A clearer view of the AI changes reaching your life.
Ideas, disagreements and possible consequences, with the sources close at hand. Ordered by the original source date; dates added to this site are shown separately.
Wissner-Gross reads falling U.S. the share of people working or seeking work alongside rapid AI progress as an early sign of an economy with a smaller role for human labor. That is his interpretation of the figures, rather than a causal finding in the employment report.
Keep in view: The numeric claims above match the dated release; the causal interpretation remains unestablished by that release. A monthly snapshot is not proof of a lasting transition.
Ethan Mollick published this argument on June 30, 2026: AI use is shifting from people working step by step with chatbots to assigning longer tasks to agents. He describes agents as systems with tools and an environment in which to act, often supported by harnesses or apps.
Mollick points to assessments of AI work, experiments with software projects, and studies of workplace use. He argues that agents are being used beyond coding, and that domain expertise—not a user’s profession—was associated with greater success using Claude Code in a separate study.
For organizations, his account suggests that assigning and overseeing agent work may become a larger part of some jobs. This is Mollick’s interpretation of developing capabilities and reported studies, not evidence that agents perform every task reliably. He stresses that model performance varies by task and that benchmark results do not always reflect practical performance.
Mollick also argues that rapid capability changes can unsettle institutions and markets. How widely this pattern extends beyond the organizations and experiments he describes remains unknown.
Keep in view: Text as retrieved on October 23, 2026; originally published June 30, 2026. The source describes selected assessments, experiments, and studies; it does not establish how broadly the reported workplace patterns apply. The article does not provide enough detail here to assess the methods or independent replication of the studies it cites.
The Federal Reserve Bank of New York Liberty Street Economics post, published June 24, 2026, compares the post-COVID decline in the U.S. labor share with earlier recessions. The labor share is the fraction of economic output paid to workers as wages and salaries. The authors report that it fell below its pre-pandemic level and reached its lowest point in the post-war period.
The authors find that the labor share’s movement after the pandemic broadly resembles patterns in earlier recessions. Their analysis also attributes the decline mainly to changes within industries, rather than shifts in output between industries. They say sectoral reallocation—the movement of economic activity across industries—contributed little to the decline in the periods they examined.
The authors argue that these findings provide little evidence that the post-COVID decline will develop differently from past episodes. That is a forecast based on the patterns they studied, not an observed future outcome. They note that a longer expansion would be needed to see whether the labor share rises again, as it did in some earlier cycles. Whether that rebound occurs remains open.
Keep in view: Text as retrieved on October 23, 2026; originally published June 24, 2026. The page is marked as edited after publication; the retrieved text may include later additions. The reading uses the post’s stated findings and does not treat the later comments as part of the authors’ analysis. The authors describe evidence from the periods they examine; it does not establish whether the labor share will rebound in a future expansion.
Anthropic’s June 23, 2026 announcement introduces Claude Tag, a way for teams to assign work to Claude in Slack. Anthropic says teams can give Claude access to selected channels, tools, data and codebases, then tag it to handle tasks while people focus on other work.
The company describes Claude as sharing context within channels, working on tasks asynchronously and, when “ambient” behavior is enabled, proactively following up on work. Administrators can control which information and tools Claude can access in each channel, set token-spend limits and review a log of its activity. Anthropic says Claude Tag is available in beta to Claude Enterprise and Team customers.
Anthropic also says 65% of its product team’s code is created by its internal version of Claude Tag. That is the company’s account of its own use, not evidence here of results for other teams. The announcement does not report independent evaluations of task quality, the effects of wider deployment, or how well the access controls work in practice.
What evidence will show whether Claude Tag’s task quality and access controls hold up when teams use it beyond Anthropic?
Keep in view: Text as retrieved on October 28, 2026; originally published June 23, 2026. The page is marked as edited after publication, so the retrieved text may include later additions; the evidence does not identify which passages, if any, were added after the original publication. The announcement reports Anthropic’s account of its own use but provides no independent evaluation of task quality or access controls.
Epoch AI published this proposal on June 17, 2026. The authors argue that common ways of forecasting AI research automation rely on indirect measures, such as compute or benchmark performance, that may miss important parts of research work. They propose a detailed task map for AI R&D: a way to describe the work involved and estimate how much current AI automates each task. The initial taxonomy groups tasks into categories and gives each an automation rating. The authors say such a map could help researchers interpret benchmark results, identify tasks that are not being measured, and study how AI supports researchers. They stress that the ratings are subjective and that a task list alone cannot show how well AI performs on each task. They also note that AI systems might make progress through workflows unlike current human research, which a human-work-based taxonomy could miss. The proposal is an initial attempt, not a measure of overall research automation. How could the task descriptions and automation ratings be tested and updated as AI research work changes?
Keep in view: Text as retrieved on October 23, 2026; originally published June 17, 2026. The authors describe the taxonomy and automation ratings as an initial, subjective attempt; the evidence does not establish how accurately the ratings measure AI performance. The authors note that a taxonomy based on current human workflows could miss AI progress through different routes.
Anthropic published its first Anthropic Public Record results on June 12, 2026, reporting a survey of Americans’ hopes and concerns about AI. The company says respondents most often named curing diseases among their hopes, while job loss was their most common fear. It also reports broad support for government involvement in AI and low trust in AI companies’ decisions about the technology.
Anthropic says the survey was fielded online by YouGov in November and December of 2025, with responses weighted to represent the US population age 16 and over. The results describe respondents’ views at that time; they do not establish how AI will affect jobs or whether people’s concerns will change. Anthropic plans to repeat the survey and expand beyond the US, but the results of future surveys are not yet known.
Keep in view: Text as retrieved on October 28, 2026; originally published June 12, 2026. The page was modified after publication. The retrieved text may include later additions; this reading uses the survey results and plans presented without treating later-dated claims as part of the original publication. The survey reports respondents’ views, not measured effects of AI on jobs or other outcomes. Subgroup margins of error are larger than the national margin of error.
Arvind Narayanan and Sayash Kapoor published this essay on June 11, 2026, arguing that AI has not caused mass layoffs of software engineers and that AI capabilities alone are unlikely to automate the whole profession. They frame software development as a “decide-execute-deliver sandwich”: AI compresses the execution layer, while deciding what to build and taking responsibility for what is delivered remain human bottlenecks. The authors point to evidence that AI agents can increase code production without a comparable increase in software releases. They also argue that job cuts attributed to AI may have other causes, and that employment trends and layoff announcements are different measures of AI’s effects. Their outlook is cautiously optimistic: cheaper software could increase demand for software and, in turn, software engineers. They distinguish this aggregate outlook from the possibility of substantial disruption to individual workers’ careers. The authors’ case depends in part on their model of how software work is organized, and they acknowledge gaps in evidence, including limited data on poorly supervised AI use in high-stakes work. How will future evidence distinguish AI-related changes in hiring from other causes of employment changes?
Keep in view: Text as retrieved on October 23, 2026; originally published June 11, 2026. The authors note that the Federal Reserve study they discuss does not capture self-employment, limiting what it can say about employment trends. The authors describe the SWE-chat dataset as self-selected and say it cannot support strong conclusions on its own. The authors identify limited data on poorly supervised AI use in high-stakes tasks as an important gap. The essay's forecasts about future demand remain the authors’ projections, not established outcomes.
Ajeya Cotra published this argument on June 5, 2026. Cotra warns that countries may become dependent on AI systems to defend against threats created by other AI systems, including software vulnerabilities, fraud, propaganda, and weapons. She argues that a fast intelligence explosion could let one AI company pull far ahead of its competitors, leaving governments and other AI systems unable to defend against it. In that scenario, Cotra says, governments might rely on a company’s AI for law enforcement and military needs, or the company might keep its most advanced AI for itself. These are conditional possibilities, not reported outcomes. Cotra proposes requiring companies to make models used internally available for external purchase, and to train models to obey the law rather than the company, with internal controls and third-party compliance checks. She notes that keeping companies at similar capability levels could conflict with reducing AI takeover risk. Whether these proposals could limit a leading company’s power without increasing other risks remains open.
Keep in view: Text as retrieved on October 25, 2026; originally published June 5, 2026. The argument describes conditional scenarios and proposed interventions; the source does not establish that these outcomes will occur or that the interventions would work.
In this position paper, Sangjun Park argues that adding explicit memory to large language models is essential for progress toward AGI. He compares LLM learning to human implicit memory and says capabilities such as long-term planning, metacognition and symbolic reasoning depend heavily on hippocampal explicit memory. This is the paper’s thesis, not an established conclusion.
Keep in view: The evidence provided is the arXiv abstract and page metadata, not the paper’s full argument or supporting analysis. The claims about what AGI requires are the author’s position; the abstract does not establish them as established findings.
Ethan Mollick published this essay on June 4, 2026. He argues that work with AI is shifting from humans directing chatbots to people deciding when to collaborate with more autonomous AI systems and when to hand over tasks. He describes using AI for feedback, fact-checking, and help with writing, while keeping the chapter drafts his own. He also says AI helped build his book website and assess how AI models might respond to its content.
Mollick’s account raises a practical question for authors and others whose work may be recommended by AI: how should they present their work to AI systems without using hidden instructions or manipulation? He says models helped him test versions of a website page, but this account does not establish how AI recommendations affect audiences or whether his approach changes them. He frames working with AI as an ongoing negotiation about when to accept its help and when to retain control.
Keep in view: Text as retrieved on October 24, 2026; originally published June 4, 2026. The essay is Mollick’s account of his own process; it does not independently establish the effects of AI recommendations on audiences. The essay describes model testing but does not report enough detail here to assess how the tests were conducted or whether their results generalize.
The Federal Reserve Bank of New York published this analysis on June 2, 2026. Will Aarons and Asani Sarkar report that small businesses in the Second District—New York, New Jersey, and Connecticut—experienced weaker revenue and employment growth in 2025 than small businesses elsewhere in the country. The authors say regional firms’ expectations for revenue and employment growth in 2026 fell more sharply than national firms’ expectations. Their analysis uses annual Small Business Credit Survey data and focuses on employer firms, meaning businesses with at least one employee other than the owner. The reported results describe survey responses and expectations, not what subsequently happened. The evidence raises a question: did regional firms’ pessimistic expectations for 2026 correspond to their eventual revenue and employment performance?
Keep in view: Text as retrieved on October 30, 2026; originally published June 2, 2026. The findings describe survey responses and expectations; the evidence does not establish what happened to firms’ revenue or employment after the survey.
On June 1, 2026, Anthropic said it had confidentially submitted a draft registration statement on Form S-1 to the U.S. Securities and Exchange Commission for a proposed initial public offering of its common stock. Anthropic said the submission gives it the option to go public after the SEC completes its review; the proposed offering depends on market conditions and other factors. The number of shares and the price had not been set. The announcement says it is not an offer to sell securities or a solicitation to buy them. Whether Anthropic proceeds, and on what terms, remains open.
On May 27, 2026, Anthropic reports findings from a survey of quantitative social scientists about AI use in research. The survey found that many respondents had tried AI chatbots, while fewer regularly used coding agents: tools that can write and run analysis code. Adoption varied by career stage, discipline, gender classification based on names, and university status.
Anthropic says coding agent users reported more project starts, working papers, and grant proposals than other researchers in the same discipline and career stage. The comparison is descriptive, not evidence that the tools caused higher output: early adopters may already have been more productive. The survey found no evidence that coding agent users submitted more papers to journals or resubmitted them more quickly, and it did not assess the quality of reported work.
Respondents were more optimistic about AI’s effect on writing publishable papers than its effect on social science overall. Anthropic notes that the survey recruited researchers for a study offering access to Claude Max accounts, so respondents may have been more interested in AI than non-respondents. The authors say a randomized experiment is ongoing, with results planned for a future report.
What will the ongoing experiment find about whether coding agents change research output, beyond differences between early adopters and other researchers?
Keep in view: Text as retrieved on October 28, 2026; originally published May 27, 2026. The survey was not representative; respondents were recruited for a study offering access to Claude Max accounts. The reported differences in research output are descriptive, not causal, and the report does not assess the quality of researchers’ work. The source says the randomized experiment is ongoing; its results are not included in this report. The retrieved page was edited after publication; later-added passages may not reflect the original version.
ADP Research published this reading on May 19, 2026. It describes data scientists as a healthy occupation despite AI-related predictions to the contrary, citing wage growth that has accelerated and a projected rise in employment. ADP Research says AI is likely to transform data science jobs by automating routine tasks and amplifying pattern recognition. But it argues that the greatest value lies in investigating data, forming testable hypotheses, making discoveries, and adding to knowledge. The source’s conclusion is that demand will continue to grow for jobs connecting data and artificial intelligence and expanding the capacity to understand and solve complex problems. The employment projection is attributed to the Bureau of Labor Statistics; ADP Research’s view of future demand is its own forecast. The sample comes from ADP payroll data and does not establish how AI-related changes will affect demand for data scientists across the wider labor market.
Keep in view: Text as retrieved on October 30, 2026; originally published May 19, 2026. The sample is based on job titles in ADP payroll data and covers data scientists employed in the U.S. private sector; the source does not establish that its findings represent all data scientists. The source presents future wage demand and continued demand for data-related occupations as expectations, not observed outcomes.
Anthropic’s announcement, published May 19, 2026, describes a global alliance with KPMG to use Claude across the firm and in work for clients. KPMG says every employee will gain access to Claude, and that it is embedding Claude in Digital Gateway, its platform for client work, beginning with tools for tax and legal clients.
The companies also plan to develop products for private equity portfolio companies and use Claude to help find and fix vulnerabilities in critical systems. These are stated plans and intended uses; the announcement does not report measured results from the wider rollout. KPMG says building an AI agent for changing tax regulations, which used to take weeks, can take minutes with Claude Cowork and Managed Agents in Digital Gateway. That is a company representative’s account, not an independently assessed comparison.
The announcement also cites joint research with UT Austin’s McCombs School of Business. A quoted researcher says value depends not only on technical adoption but also on how employees exercise judgment, shape workflows, evaluate outputs, and make decisions with AI. How the alliance performs in practice, including the quality and oversight of AI-assisted work, remains open.
Keep in view: Text as retrieved on October 24, 2026; originally published May 19, 2026. The page is marked as edited after publication. Later additions may be present in the retrieved text, so this reading does not treat its post-publication material as part of the original announcement. The announcement is from Anthropic and reports company plans and statements; it does not provide independent evaluation of the described uses or measured outcomes from the wider rollout.
On May 19, 2026, the AI Now Institute described a new research portfolio examining how AI is deployed across healthcare systems. The institute says its work will focus on patient safety, worker dignity, public spending, and democratic accountability, including nurses’ experiences, regulation, and the healthcare industry’s political economy.
The institute argues that claims of improved efficiency and care should be weighed against risks such as missed drug allergies, fabricated clinical notes, job displacement, and added implementation and subscription costs. It also says oversight and independent evaluation remain limited, and that some tools may not receive governmental review. These are the institute’s assessments, not findings independently established by this page.
The portfolio plans participatory research with healthcare workers and patients, and policy work based on its findings. Its first publication in the portfolio, the institute says, examines gig nursing platforms and their effects on staffing, wages, and regulation. The page does not establish how widespread the described harms are or what outcomes the new research will find.
What will the portfolio’s research find about how AI deployment affects patients and healthcare workers across different settings?
Keep in view: Text as retrieved on October 24, 2026; originally published May 19, 2026. The page was edited after publication, and the retrieved text does not identify which passages, if any, were added later. Claims about the original publication may therefore require editorial verification. The page describes the institute’s research agenda and concerns; it does not provide independent evaluations of the cited AI tools or establish how widely the described effects occur.
The paper introduces Njord, a probabilistic data-driven ocean forecasting model for global and regional domains. It combines a deep latent-variable framework with a graph neural network and generates samples at each forecast step in one forward pass. The authors report experiments at 0.25° global resolution and 2 km resolution for the Baltic Sea. They report strong performance against deterministic machine-learning baselines on both domains. Separately, on the global OceanBench benchmark, they report the lowest errors on average across upper-ocean variables when evaluated against real-world observations, with the largest improvements in surface-temperature prediction.
Keep in view: This reading is based on the arXiv page's abstract; the paper's full methods, experimental details, and numerical results were not assessed here. The abstract reports benchmark outcomes but does not provide enough detail here to assess their statistical significance or independent reproducibility.
Anthropic’s announcement, published May 13, 2026, introduces Claude for Small Business, a package of connectors and ready-to-run workflows for tools small businesses use. Anthropic says the service can handle tasks such as payroll planning, invoice follow-up, and marketing work, with owners approving before anything sends, posts, or pays. The announcement also describes training and partnerships intended to help small businesses use AI. These are product capabilities and plans described by Anthropic; the page does not provide independent evidence of their results. How well do the workflows perform in small businesses’ day-to-day operations, and what oversight do owners need to catch errors?
Keep in view: Text as retrieved on October 24, 2026; originally published May 13, 2026. The page was edited after publication, so the retrieved text may include later additions; the evidence does not identify which passages, if any, were added later. The announcement describes Anthropic’s product claims and plans, but provides no independent evaluation of workflow performance or error rates.
METR published this survey on May 11, 2026. METR reports that technical workers surveyed in February–April described larger gains in work speed than work value: respondents’ median self-reported value change ranged from 1.4–2x across measures, while median speed change was 3x. METR distinguishes value—the value of work respondents believe they produce—from speed, or how long tasks take. It says speed measures are likely biased upward relative to value measures, partly because AI can make lower-value tasks quicker without raising the value of a person’s overall contribution.
METR cautions that self-reports may not reflect actual productivity gains. Respondents may find counterfactual questions difficult, and survey estimates have often exceeded field-experiment or quasi-experiment estimates, though the size of that difference is hard to determine. The convenience sample also plausibly suffers from significant selection bias. Still, METR argues surveys usefully complement other evidence because they are cheap, broad, and can address otherwise difficult questions. For organizations tracking AI’s potential to automate AI R&D, it recommends carefully defining what to measure and considering surveys of managers or productivity researchers.
Keep in view: Text as retrieved on October 23, 2026; originally published May 11, 2026. The findings are self-reports from a convenience sample; METR says the results plausibly suffer from significant selection bias and may not reflect actual productivity gains. METR says it is difficult to determine how much surveys overestimate productivity gains relative to experimental data.
METR published this review on May 8, 2026, assessing Anthropic’s argument that catastrophic risk from its models automating research and development is very low. METR says the report’s survey evidence has problems with sample size, question detail, framing, and calibrated responses, and that it miscounts a missing response as a negative one. METR also argues that the report overlooks the possibility that AI could substantially accelerate research and development before fully automating it.
METR says evidence gathered after Claude Opus 4.6’s release, including METR evaluations and the lack of public reports of the model automating a key domain, supports the report’s low-risk conclusion. But METR says the report’s own evidence is inadequate to establish that conclusion. It recommends changes to Anthropic’s surveys and more reporting of other evidence that could indicate AI progress. METR expects the public report to be updated and says it may publish another review if changes materially affect its opinions. The review also draws on non-public materials, so its assessment is not based only on public evidence.
Keep in view: Text as retrieved on October 23, 2026; originally published May 8, 2026. METR says it used additional non-public materials, including information shared during a previous review; this summary cannot independently assess those materials. The supplied text is the executive summary of METR’s original review, not the full review documents.
In a paper listed in arXiv cs.AI, Ziyun Zeng, Yiqi Lin, Guoqiang Liang, and Mike Zheng Shou argue that video background replacement is underexplored because high-quality training data is scarce. Unlike local edits that preserve a scene, replacement must create new scenes that remain consistent over time while keeping foreground-background interactions accurate.
The authors trace poor results in existing models to imprecise background guidance in data synthesis. They describe a pipeline that generates foreground and background guidance separately, with quality filtering, and introduce Sparkle, a dataset of about 140,000 video pairs across five background-change themes, plus Sparkle-Bench, an evaluation benchmark.
The authors report that their dataset and a model trained on it perform substantially better than existing baselines on OpenVE-Bench and Sparkle-Bench. They say the dataset, benchmark, model, and full data-generation pipeline are open-sourced. This offers researchers a resource to study a video-editing task the authors say has received less attention; the abstract does not establish how well the results transfer to practical production work.
An open question is whether the reported gains hold in real creative workflows, where backgrounds must remain natural and interact convincingly with foreground subjects.
Keep in view: The performance comparison is reported in the abstract; it provides no scores or details about evaluation conditions. The abstract does not establish whether the reported benchmark gains transfer to real creative production workflows.
On May 4, 2026, Challenger, Gray & Christmas published a review of the Department of Labor’s “Make America AI-Ready” course. The firm describes it as a useful introduction for workers with no prior exposure to large language models, but not a substitute for sustained training.
The course is free and delivered by text message. Challenger says its lessons introduce AI, prompting, verification, and responsible use, and that text delivery can reach workers without reliable internet access.
The review also identifies limits: lessons are brief and dense on a phone, interactive materials require internet access, and learners cannot ask the teaching system follow-up questions. Challenger says the course does not address sycophancy, a tendency of large language models to agree with users and validate their assumptions. It argues that the course relies on learners’ existing critical thinking and judgment, skills it says the course does not meaningfully build.
Challenger recommends that employers treat the course as a common starting point, then add role-specific training, judgment-building, and hands-on practice. The review does not establish how well the course improves workers’ skills or how its effects vary among learners. How much additional practice and support do learners need to use AI effectively in their jobs?
Keep in view: Text as retrieved on October 24, 2026; originally published May 4, 2026. The page is marked as edited after publication, so the retrieved text may include later additions; the evidence does not identify which passages, if any, were added after the original publication date. The review offers an assessment but provides no measured results for how well the course improves workers’ skills.
Authors Haofei Yu, Yining Zhao, Lenore Blum, Manuel Blum and Paul Pu Liang propose CTM-AI, a multi-agent AI system inspired by the Conscious Turing Machine (CTM), a formal machine model of consciousness. Rather than treating consciousness as a demonstrated property, they present the model as a blueprint for coordinating many kinds of processors, including specialized experts and general-purpose learners. Their proposed mechanism competitively selects and integrates information from processors, while feedback lets the system change its internal connections and communication. The authors report strong results on MUStARD and UR-FUNNY, and improvements on tool-using and agentic tasks. These results describe performance on named benchmarks; the abstract does not establish how well CTM-AI works in other settings or whether it has human-like generality. The paper’s central idea is to make a model inspired by consciousness into a testable design for coordinating AI capabilities. How does CTM-AI perform beyond the benchmarks reported in the abstract?
Keep in view: The supplied evidence is the arXiv page and abstract, not the full paper; it does not provide evaluation methods, comparison details, or results beyond those named in the abstract. The reported benchmark results do not establish performance in other settings, human-like generality, or consciousness.
Anthropic’s announcement, published April 22, 2026, says its Economic Research team is launching a monthly survey through Anthropic Interviewer. Anthropic argues that usage data and labor-market measures such as employment rates, wages and layoffs do not capture how people experience economic changes or what they expect as AI advances. The survey will ask Claude users about changes to their work, tasks they may hand off, productivity, hiring and roles, and their expectations for the future. Anthropic says recurring responses, combined with Claude usage data in a privacy-preserving way, could surface changes before they appear in aggregate labor-market data. It plans to publish insights in future reports and research briefs, and says data will be processed under its Supplemental Privacy Policy. The announcement presents the survey as a way to hear from people living through an uncertain transition; it does not report survey findings or establish how representative responses will be. The page was edited after publication, so the retrieved text may include later additions that cannot be dated separately.
Keep in view: Text as retrieved on October 25, 2026; originally published April 22, 2026. The page was edited after publication. The retrieved text does not identify which passages, if any, were added after the original publication date. The announcement describes the survey’s aims, not findings, and does not establish how representative responses will be.
On March 20, 2026, MIT Technology Review’s The Download reported that OpenAI had made a fully automated AI researcher its “north star” for the next few years. The newsletter describes this as an agent-based system intended to tackle large, complex problems by itself. OpenAI plans to build an “autonomous AI research intern” by September, for a small number of specific research problems; the newsletter says this would be a precursor to a multi-agent system slated to debut in 2028. In an interview, OpenAI chief scientist Jakub Pachocki discussed the plans with Will Douglas Heaven, who wrote the item. The distinction between a limited research intern and the more ambitious system matters: the stated plan is staged, and the larger system is a future goal, not a reported achievement. The newsletter does not describe how the systems would be evaluated, what research problems they would handle, or how much human oversight they would need.
Keep in view: Text as retrieved on October 27, 2026; originally published March 20, 2026. This is an archived copy. The retrieved newsletter describes plans but does not provide evaluation methods, intended research problems, or oversight details. Its linked interview is not included in the supplied text.
METR’s page was published on February 24, 2026. METR says selection effects make its new experiment a poor measure of AI’s productivity impact on participating open-source developers. Developers who expect AI to help them more are less likely to join or to submit tasks they would rather do with AI; METR says this likely makes its estimate a lower bound. A lower pay rate may also have contributed to selection, and time measurements can be unreliable when developers use multiple AI agents concurrently. METR says the data offer only very weak evidence about the size of any increase in productivity, despite raw results showing some evidence for speedup. It calls the experiment’s central estimate likely a bad proxy for the real productivity impact of AI tools on these developers. METR says developer productivity is an important input to AI R&D acceleration and plans to explore other approaches, including shorter, more intensive experiments, observational data, questionnaires, fixed-task experiments, evaluations and developer-level experiments. Those approaches have limitations, and METR expects to keep redesigning parts of its measurement techniques as model capabilities increase. How much higher pay or shorter experiments would reduce the selection effects remains unknown.
Keep in view: METR says selection effects, a lower pay rate, task differences, and time-reporting difficulties complicate interpretation; the experiment is not presented as a reliable measure of productivity impact. Text as retrieved on October 23, 2026; originally published February 24, 2026.
Anthropic’s announcement, published February 17, 2026, presents Claude Sonnet 4.6 as an upgrade in coding, computer use, long-context reasoning, planning, knowledge work, and design. Anthropic says the model can interact with software through a simulated mouse and keyboard, rather than relying on purpose-built connectors. It argues this could make AI useful with specialized systems that lack modern interfaces, while acknowledging that the model still lags behind the most skilled humans at computer use.
Anthropic reports that developers in early access preferred the model to its predecessor and describes improvements on OSWorld, a benchmark that tests computer tasks in real software. The company also says its safety evaluations found better resistance to prompt injection, where hidden instructions on a website attempt to hijack a model. These results are company-reported evaluations and early-user feedback, not evidence here of dependable performance across organizations or unsupervised work. Anthropic says Sonnet 4.6 is available across its plans and products, while recommending its Opus model for tasks demanding the deepest reasoning. The announcement does not establish how often the model’s computer-use performance will hold up in routine deployment, or what oversight those deployments require.
Keep in view: Text as retrieved on October 24, 2026; originally published February 17, 2026. The page was edited after publication, so the retrieved text may include later additions; the evidence does not identify which passages changed. The performance and safety claims are Anthropic’s reports; the supplied evidence does not include the underlying system card, benchmark results, or independent evaluations. The early-user reports do not establish performance across routine deployments or the oversight those deployments require.
Amodei expects AI to spread through the economy unusually quickly, while procurement, permissions and organizational change still take time. He separates models writing code from complete job automation and reduced demand for workers.
Press play to watch. If playback is unavailable, open the source below.
Scott A. Martin, Noah Brenowitz, Dale Durran and Michael Pritchard present “long-range distillation,” a method for training models to forecast far ahead using synthetic climate data generated by another forecasting model. They report generating more than 10,000 simulated years of climate data. In experiments where the forecasting model and simulated climate come from the same system, the trained models outperformed climatology and approached their teacher model’s skill, while replacing hundreds of repeated forecasting steps with one. For real-world forecasting, the authors report that two approaches achieved subseasonal-to-seasonal skill competitive with the ECMWF IFS ensemble. They also report improvements over training with ERA5 data alone. These are results described in the paper’s abstract, not an independent assessment. How much do the reported gains vary across forecast horizons and weather conditions?
Keep in view: The supplied evidence is the paper’s abstract and does not include detailed performance metrics or an independent assessment. The paper distinguishes perfect-model experiments from real-world forecasting results; its reported outcomes should not be treated as a general guarantee of forecast performance.
MIT Technology Review’s article was published on September 9, 2025. It argues that new estimates of chatbot energy use leave larger questions about AI’s climate burden unanswered. The figures omit important measurement details and cover chat interactions, not other ways people use generative AI. The article says researchers need information about AI use across applications and modalities to assess its full climate impact. It also questions whether AI’s proposed climate benefits will outweigh rising electricity demand, noting that companies have not shared enough evidence to assess the balance. Its central uncertainty is whether demand will reach the scale behind companies’ data-center plans. The article says that answer will determine whether the buildout becomes a lasting shift in the energy system or a short-lived spike.
Keep in view: Text as retrieved on November 1, 2026; originally published September 9, 2025. The article reports that company figures are incomplete; it does not provide a full accounting of AI’s energy demand or establish whether prospective climate benefits outweigh its costs.
Mollick argues that AI can support learning when it guides the learner through the work. An assistant that supplies answers can remove the practice that builds understanding, even when the user has good intentions.
Keep in view: The cited experiment concerns particular students, mathematics tasks and GPT-4 setups; it cannot establish effects for every learner or job.
Published July 3, 2025, Cohere presents an infographic about AI in financial services. The page says it explores how AI is transforming the sector, pointing readers to productivity gains, operational efficiencies and steps financial leaders are taking to “win with AI.” This is the page’s description of the infographic, not evidence that particular gains have been achieved. The retrieved text does not include the infographic itself or explain which steps it discusses, so the specific uses, evidence and outcomes remain unknown.
Keep in view: Text as retrieved on October 28, 2026; originally published July 3, 2025. The retrieved page does not include the infographic or details supporting its description of gains, efficiencies or steps. The retrieved page includes later recommendations dated October 26 and October 29, 2026; they are not treated as part of the original publication.
The authors argue that AI’s direction is a political choice. They call for worker organizing, enforceable limits and public accountability, so affected people can contest how institutions use AI rather than treat its expansion as a certainty.
Keep in view: Its recommendations express political values. The chapter’s cited cases require their own source checks before being reused as factual allegations.
Cohere’s page, published June 2, 2025, presents an infographic about AI agents and workplace productivity. Cohere says the infographic explores how AI agents can boost productivity by automating tasks, improving decision-making, and empowering teams. The page’s description gives those as potential routes to productivity, but the retrieved text does not show the infographic or provide evidence for the claimed effects. What examples or evidence does the infographic offer for these proposed benefits?
Keep in view: Text as retrieved on October 27, 2026; originally published June 2, 2025. The retrieved page text does not include the infographic itself, so its examples and supporting evidence cannot be assessed. The page was edited after publication; the retrieved text includes later page content that is not treated here as part of the original publication.
MIT Technology Review’s AI Hype Index was published on April 29, 2025. It frames its purpose as separating AI reality from “hyped-up fiction,” then presents AI agents as a promising but risky product: assistants that can complete tasks without human supervision, but could also be weaponized to carry out cyberattacks. The article gives Simular AI’s S2 as an example of an agent that switches between models according to its instructions; it does not establish how reliably the agent works in practice. It also reports on OpenAI’s possible move into social media and AI models’ growing ability to make music. The closing joke turns to a half-marathon between humans and humanoid robots: the results, the article quips, offer little reason to worry about a robot uprising soon. The page’s edited-after-publication status means the retrieved text may include later additions, so these items cannot all be confidently attributed to the original publication. The evidence does not give details of the cyberattack risk, the reported social-media plans, or the race results.
Keep in view: Text as retrieved on October 29, 2026; originally published April 29, 2025. The page is marked as edited after publication. The evidence does not establish which passages, if any, were added after the original publication. The article gives no details about how AI agents might be used in cyberattacks, how the reported social-media plans developed, or what the half-marathon results were.
METR published this work on March 19, 2025. The authors propose measuring AI agents by the length of tasks they can complete, using the time those tasks take human professionals as a measure of difficulty. They report that agents’ measured task-completion horizons grew exponentially over the preceding six years, with a doubling time of around seven months.
METR’s measure estimates the task length at which a model reaches a chosen success probability; the post focuses on 50% reliability. The authors say that current models can succeed on some tasks taking experts hours, but reliably complete tasks only up to a few minutes long. They argue that this measure helps explain why strong benchmark results do not necessarily translate into robust help with day-to-day work.
The authors forecast that, if the past trend continues, agents could perform a wide range of week-long tasks within two to four years, and month-long projects by the end of the decade. These are forecasts, not demonstrated outcomes. METR says the estimate depends on the tasks and human performance measured, and notes uncertainty about whether the trend will continue. How well does task length measured on these evaluations predict success on real-world work?
Keep in view: Text as retrieved on October 24, 2026; originally published March 19, 2025. The retrieved page includes later updates and measurements; this reading uses the material presented as reflecting the state of the data at original publication and excludes later additions. The authors state that task-horizon estimates depend on task selection and how human performance is measured, and that future trend changes and external validity are major sources of uncertainty.
MIT Technology Review’s March 11, 2025, edition of The Download says new AI benchmarks could help developers judge fairness, while actual fixes may require other techniques. Scott J Mulligan reports that the benchmarks assess AI systems’ awareness of scenarios and contexts; he says their researchers were prompted by earlier approaches whose failure to account for differences between groups could make systems less fair. The edition does not name the benchmarks or describe how they work in detail.
James O’Donnell describes artificial general intelligence (AGI)—an “ultra-powerful AI system we don’t have yet”—as a balloon inflated by optimism or fear and deflated when reality falls short of expectations. He connects renewed attention to AGI with news including the launch of Manus, an AI agent created by a Chinese startup, and directs readers to a separate story for more. The edition presents this as a way to understand recurring hype, not as evidence that AGI has arrived. It does not establish whether the fairness benchmarks lead to better outcomes or what techniques would fix the problems they measure.
Keep in view: Text as retrieved on October 25, 2026; originally published March 11, 2025. The newsletter does not name the fairness benchmarks or describe their methods, results, or whether they lead to better outcomes. The edition refers readers to separate stories for further detail on the fairness research and Manus; that context is not included here.
Revelio Labs published this outlook on March 7, 2025. It argues that February’s headline jobs report may look encouraging while obscuring a cooling labor market: federal layoffs implemented after the survey period were not reflected, and several underlying measures weakened. Revelio Labs says its job-postings data points to sustained but muted hiring, with postings recovering from a seasonal dip but still below their levels in 2022 and 2023. It also notes that postings rose in most industries and that recruiting positions were being advertised more often, which it views as a possible sign of labor-market strength. But the outlook argues federal job cuts are likely to weigh on net job creation in coming months, as hiring weakens elsewhere. It further forecasts that uncertainty over economic policy, especially tariffs, may lead companies to delay hiring and investment. Revelio Labs says the balance of risks facing the Federal Reserve may be shifting: a weakening labor market could outweigh tariffs’ initial inflationary effect and lead to softer price pressures and possible Fed action. The scale of the employment effects and the policy response remain uncertain.
ADP Research published this account on December 17, 2024. The publication describes the Employee Motivation and Commitment Index as tracking how people think and feel about their jobs and employers, using a monthly panel survey of U.S. workers. ADP Research says the index offers a real-time measure of worker allegiance that can indicate whether people are flourishing in their jobs or detaching. It also says sector indexes use a three-month rolling average. The account presents the index as a way to assess worker sentiment; it does not establish how closely the measure reflects workers’ experiences beyond the survey responses it collects.
Keep in view: Text as retrieved on October 30, 2026; originally published December 17, 2024. The page says December data was updated on January 13, 2025, after the stated publication date; the retrieved text does not establish which December results appeared in the original account. The evidence describes the index’s purpose and survey method but does not establish how well the index reflects workers’ experiences.
Herndon and Dryhurst propose treating AI training as a collective creative act. In The Call, the gathering of voices, recording rules and shared governance become part of the artwork, alongside the sounds a model produces.
Published May 29, 2024, Leopold Aschenbrenner closes his series by arguing that superintelligence could arrive within the decade and that its development should be treated as a national-security matter. He calls his proposed stance “AGI Realism,” rejecting both calls to pause AI indefinitely and views he sees as denying the risks of AGI. He argues that America must lead, that AI labs should put the national interest first, and that managing safety risks will require more than improvisation. Aschenbrenner also warns against assuming a capable rescue team will appear: in his view, a small number of people behind the scenes are already carrying responsibility for what comes next. He says his scenario may be substantially wrong and that its error bars are large, while stressing that he considers it important to be concrete. The text’s central uncertainty is whether the predicted path to superintelligence, and the national-security stakes Aschenbrenner attaches to it, will match events. The retrieved page is marked as edited after publication, so it is unclear whether every passage reflects the original version.
Keep in view: Text as retrieved on October 30, 2026; originally published May 29, 2024. The page is marked as edited after publication. The retrieved text does not identify which passages, if any, were added or changed after the original publication. The source presents forecasts and judgments, not evidence that its predicted timeline or proposed national-security approach will come to pass.
In its May 7, 2024, edition, MIT Technology Review describes University of Florida researchers trying to create a large animal from stem cells, without an egg, sperm, or conception. They transferred “synthetic embryos”—artificial structures created in a lab—to cows, hoping some might take. The researchers want to make a live animal; the newsletter says that would be a new way to breed cattle and could challenge ideas about what life is. It does not report whether any embryo took or whether a live animal resulted.
The edition also profiles AI prompt engineering as a job. It says the role’s principal aim is to help a company integrate AI into its operations, while noting that companies define it differently. Danai Myrtzani, of digital marketing agency Sleed, describes herself as more prompter than engineer. The newsletter says she helped develop a tool that generates personalized LinkedIn posts for clients. It offers these examples as a glimpse of the work, not evidence of how common the role is or what its future will be.
Keep in view: Text as retrieved on October 29, 2026; originally published May 7, 2024. The newsletter points readers to full stories but does not include their full reporting, so details beyond the excerpts here are unavailable. The newsletter does not report whether any synthetic embryo took or whether a live animal resulted.
MIT Technology Review’s The Download was published April 16, 2024. It says police bodycams have not delivered the change activists hoped for: much footage goes unwatched, officers often do not use the cameras properly, and public footage may be selectively edited and lack context. The newsletter says AI startups are developing programs that turn bodycam footage into text for different players in the legal system, but procedural, legal, and cultural barriers remain. Its second feature says AI is helping robots move toward household tasks such as folding laundry, cooking, and unloading shopping baskets. Melissa Heikkilä’s cover story describes robotics as being at an inflection point, with developments that could help robots leave the lab for homes. The newsletter presents this as a possibility, not an accomplished transition. The bodycam account points to a gap between recording events and making footage useful to the public or legal system. The newsletter does not establish whether the new programs can overcome that gap. It also does not specify which developments are converging in robotics or how reliably household robots can perform these tasks outside the lab.
Keep in view: Text as retrieved on October 29, 2026; originally published April 16, 2024. The newsletter summarizes linked stories rather than presenting their full evidence; the bodycam-to-text programs’ effectiveness and the barriers they face are not detailed. The robotics feature does not identify the developments it says are converging or establish how well household robots perform outside the lab.
MIT Technology Review’s article, published April 11, 2024, argues that AI, cheaper hardware and shared robot data are making home robots more capable of learning tasks and adapting to unfamiliar settings. It contrasts robots trained through detailed, setting-specific plans with approaches that learn from demonstrations or trial and error. The article presents the Evanses’ experience with Stretch as a glimpse of the human stakes: Henry Evans used the robot to handle food, play with his granddaughter and hold playing cards, gaining some independence despite its limitations. Researchers describe early advances, including robots learning tasks from demonstrations and a shared data set that improved performance across different robots. But the article does not present these examples as ready-to-use household help. It reports that robots remain clumsy, lack common sense for multitasking, and can require researchers to troubleshoot; the Evanses also depend on reliable Wi-Fi to use Stretch. The source argues that wider skills and more robot-specific data could help move the field toward useful home robots, while noting that cost and reliability remain barriers. How much more data and improvement in reliability would be needed before home robots could work without researchers’ help?
Keep in view: Text as retrieved on October 28, 2026; originally published April 11, 2024. The page was modified after publication, so the retrieved text may include later additions; this reading treats claims about developments after the original publication date as outside the original account. The article describes early demonstrations and researchers’ expectations; it does not establish that home robots are ready for routine use.
LeCun argues that a few companies should not control everyone’s information. He proposes widely available base models that communities can adapt to their languages, values and needs, making pluralism a design goal.
Press play to watch. If playback is unavailable, open the source below.
Okinyi argues that flexible platform work needs dependable pay and fair treatment. Drawing on fellow workers’ accounts of lost access and earnings, he asks readers to judge AI’s supply chain by the lives of the people doing its work.
Data Workers’ Inquiry · interviews collected by Mophat Okinyi · Chris’s testimony · Open original source
Keep in view: No prevalence estimate or legal finding is claimed. This record does not imply that all data workers share the reported experience.
ADP Research published this account on November 14, 2023. ADPRI says it added six questions to its monthly survey of U.S. workers to investigate how the resumption of student-loan payments might affect work. The questions covered household debt and expected payment burdens, anticipated changes in spending, and respondents’ feelings about carrying debt. ADPRI says it compared answers with other survey measures, including work productivity and intent to leave an organization, using crosstabulation analyses and likelihood ratios. The account describes a method for studying possible effects; it does not report the analysis results or show how student-loan payments affected respondents’ work. The survey’s cross section and the reported response totals provide context, but this account does not give enough detail to assess how well its findings represent U.S. workers.
Keep in view: Text as retrieved on October 30, 2026; originally published November 14, 2023. The retrieved account describes the survey and analysis methods but does not include the analysis results, so it cannot establish what effects student-loan payments had on work. The account does not provide enough detail here to assess the survey’s representativeness or the analysis.
Cory Stahle and Nick Bunker’s June 29, 2023 update in Indeed Hiring Lab argues that construction hiring held up better than expected as interest rates rose. They point to construction job postings falling less than overall postings and posted wage growth remaining strong, even as both cooled. They suggest that housing construction activity, factory building and infrastructure projects may have supported demand for construction workers, but do not establish which factors mattered most. The authors describe the broader US labor market as robust, with low joblessness and layoffs, while job postings and wage growth cool. They also warn that it may still stumble amid tight monetary policy, stress in the US banking system and slowing economic growth. The update explains nominal wage growth as pay gains measured before inflation is taken into account. Its measures concern job postings and posted wages; it does not establish how these trends affected workers’ realized pay or job security.
Keep in view: Text as retrieved on October 30, 2026; originally published June 29, 2023. The update identifies possible contributors to construction labor demand but does not establish their relative effects. Job postings and posted wage growth are the measures discussed; the update does not establish workers’ realized pay or job security from these measures.
OpenAI authors Sam Altman, Greg Brockman and Ilya Sutskever published this argument on May 22, 2023. They say it is time to prepare for superintelligence, which they describe as future AI systems dramatically more capable than even AGI. They argue that such systems could bring great prosperity but also pose existential risk, so governance should not wait until risks emerge.
Their proposed starting points are coordination among leading development efforts, an international authority able to inspect systems and require audits above a capability or compute threshold, and research into making superintelligence safe. They distinguish these systems from models below a significant capability threshold, which they say should not face the same regulation. They also argue that decisions about powerful systems and their deployment need strong public oversight, while individual users should retain substantial control within public bounds.
The authors say they pursue this work because they expect it to improve the world and because they believe stopping superintelligence would be difficult and risky. They acknowledge that they do not yet know how to design a public decision-making mechanism, and call safety an open research question.
Keep in view: Text as retrieved on October 24, 2026; originally published May 22, 2023. This is an archived copy. The page is marked as edited after publication, so the retrieved text may include later additions; the evidence does not identify which passages changed. The authors say they do not yet know how to design the public decision-making mechanism, and describe making superintelligence safe as an open research question.
Published April 7, 2023, Nick Bunker argues in Indeed Hiring Lab that the US labor market is slowing rather than stalling: hiring and wage growth are moderating, while unemployment remains low. Bunker says easing demand for new hires is still strong enough to draw more workers into the labor market, and that this return of workers, alongside easing employer demand, is moderating wage growth.
He points to the share of workers ages 25 to 54 with a job reaching its highest level since May 2001. He also reports that average hourly earnings grew more slowly over the past three months than in December, but cautions that the slowdown is more rapid than the trend in other data. He says other wage data, including the Employment Cost Index, will be important for assessing whether this trend holds.
Bunker characterizes the labor market’s transition as gradual and, for now, largely sustainable and healthy—not an abrupt end to the jobs growth of the past couple of years. The evidence in the article leaves open whether other wage measures will show the same slowdown.
Keep in view: Text as retrieved on October 30, 2026; originally published April 7, 2023. The page is marked as edited after publication. The retrieved text does not identify which passages, if any, were added or changed after the original publication date. Bunker says the wage slowdown needs comparison with other data; this article does not establish whether those other measures show the same trend.
A clearer view of the AI changes reaching your life.
Ideas, disagreements and possible consequences, with the sources close at hand. Ordered by the original source date; dates added to this site are shown separately.
Wissner-Gross reads falling U.S. the share of people working or seeking work alongside rapid AI progress as an early sign of an economy with a smaller role for human labor. That is his interpretation of the figures, rather than a causal finding in the employment report.
Keep in view: The numeric claims above match the dated release; the causal interpretation remains unestablished by that release. A monthly snapshot is not proof of a lasting transition.
Ethan Mollick published this argument on June 30, 2026: AI use is shifting from people working step by step with chatbots to assigning longer tasks to agents. He describes agents as systems with tools and an environment in which to act, often supported by harnesses or apps.
Mollick points to assessments of AI work, experiments with software projects, and studies of workplace use. He argues that agents are being used beyond coding, and that domain expertise—not a user’s profession—was associated with greater success using Claude Code in a separate study.
For organizations, his account suggests that assigning and overseeing agent work may become a larger part of some jobs. This is Mollick’s interpretation of developing capabilities and reported studies, not evidence that agents perform every task reliably. He stresses that model performance varies by task and that benchmark results do not always reflect practical performance.
Mollick also argues that rapid capability changes can unsettle institutions and markets. How widely this pattern extends beyond the organizations and experiments he describes remains unknown.
Keep in view: Text as retrieved on October 23, 2026; originally published June 30, 2026. The source describes selected assessments, experiments, and studies; it does not establish how broadly the reported workplace patterns apply. The article does not provide enough detail here to assess the methods or independent replication of the studies it cites.
The Federal Reserve Bank of New York Liberty Street Economics post, published June 24, 2026, compares the post-COVID decline in the U.S. labor share with earlier recessions. The labor share is the fraction of economic output paid to workers as wages and salaries. The authors report that it fell below its pre-pandemic level and reached its lowest point in the post-war period.
The authors find that the labor share’s movement after the pandemic broadly resembles patterns in earlier recessions. Their analysis also attributes the decline mainly to changes within industries, rather than shifts in output between industries. They say sectoral reallocation—the movement of economic activity across industries—contributed little to the decline in the periods they examined.
The authors argue that these findings provide little evidence that the post-COVID decline will develop differently from past episodes. That is a forecast based on the patterns they studied, not an observed future outcome. They note that a longer expansion would be needed to see whether the labor share rises again, as it did in some earlier cycles. Whether that rebound occurs remains open.
Keep in view: Text as retrieved on October 23, 2026; originally published June 24, 2026. The page is marked as edited after publication; the retrieved text may include later additions. The reading uses the post’s stated findings and does not treat the later comments as part of the authors’ analysis. The authors describe evidence from the periods they examine; it does not establish whether the labor share will rebound in a future expansion.
Anthropic’s June 23, 2026 announcement introduces Claude Tag, a way for teams to assign work to Claude in Slack. Anthropic says teams can give Claude access to selected channels, tools, data and codebases, then tag it to handle tasks while people focus on other work.
The company describes Claude as sharing context within channels, working on tasks asynchronously and, when “ambient” behavior is enabled, proactively following up on work. Administrators can control which information and tools Claude can access in each channel, set token-spend limits and review a log of its activity. Anthropic says Claude Tag is available in beta to Claude Enterprise and Team customers.
Anthropic also says 65% of its product team’s code is created by its internal version of Claude Tag. That is the company’s account of its own use, not evidence here of results for other teams. The announcement does not report independent evaluations of task quality, the effects of wider deployment, or how well the access controls work in practice.
What evidence will show whether Claude Tag’s task quality and access controls hold up when teams use it beyond Anthropic?
Keep in view: Text as retrieved on October 28, 2026; originally published June 23, 2026. The page is marked as edited after publication, so the retrieved text may include later additions; the evidence does not identify which passages, if any, were added after the original publication. The announcement reports Anthropic’s account of its own use but provides no independent evaluation of task quality or access controls.
Epoch AI published this proposal on June 17, 2026. The authors argue that common ways of forecasting AI research automation rely on indirect measures, such as compute or benchmark performance, that may miss important parts of research work. They propose a detailed task map for AI R&D: a way to describe the work involved and estimate how much current AI automates each task. The initial taxonomy groups tasks into categories and gives each an automation rating. The authors say such a map could help researchers interpret benchmark results, identify tasks that are not being measured, and study how AI supports researchers. They stress that the ratings are subjective and that a task list alone cannot show how well AI performs on each task. They also note that AI systems might make progress through workflows unlike current human research, which a human-work-based taxonomy could miss. The proposal is an initial attempt, not a measure of overall research automation. How could the task descriptions and automation ratings be tested and updated as AI research work changes?
Keep in view: Text as retrieved on October 23, 2026; originally published June 17, 2026. The authors describe the taxonomy and automation ratings as an initial, subjective attempt; the evidence does not establish how accurately the ratings measure AI performance. The authors note that a taxonomy based on current human workflows could miss AI progress through different routes.
Anthropic published its first Anthropic Public Record results on June 12, 2026, reporting a survey of Americans’ hopes and concerns about AI. The company says respondents most often named curing diseases among their hopes, while job loss was their most common fear. It also reports broad support for government involvement in AI and low trust in AI companies’ decisions about the technology.
Anthropic says the survey was fielded online by YouGov in November and December of 2025, with responses weighted to represent the US population age 16 and over. The results describe respondents’ views at that time; they do not establish how AI will affect jobs or whether people’s concerns will change. Anthropic plans to repeat the survey and expand beyond the US, but the results of future surveys are not yet known.
Keep in view: Text as retrieved on October 28, 2026; originally published June 12, 2026. The page was modified after publication. The retrieved text may include later additions; this reading uses the survey results and plans presented without treating later-dated claims as part of the original publication. The survey reports respondents’ views, not measured effects of AI on jobs or other outcomes. Subgroup margins of error are larger than the national margin of error.
Arvind Narayanan and Sayash Kapoor published this essay on June 11, 2026, arguing that AI has not caused mass layoffs of software engineers and that AI capabilities alone are unlikely to automate the whole profession. They frame software development as a “decide-execute-deliver sandwich”: AI compresses the execution layer, while deciding what to build and taking responsibility for what is delivered remain human bottlenecks. The authors point to evidence that AI agents can increase code production without a comparable increase in software releases. They also argue that job cuts attributed to AI may have other causes, and that employment trends and layoff announcements are different measures of AI’s effects. Their outlook is cautiously optimistic: cheaper software could increase demand for software and, in turn, software engineers. They distinguish this aggregate outlook from the possibility of substantial disruption to individual workers’ careers. The authors’ case depends in part on their model of how software work is organized, and they acknowledge gaps in evidence, including limited data on poorly supervised AI use in high-stakes work. How will future evidence distinguish AI-related changes in hiring from other causes of employment changes?
Keep in view: Text as retrieved on October 23, 2026; originally published June 11, 2026. The authors note that the Federal Reserve study they discuss does not capture self-employment, limiting what it can say about employment trends. The authors describe the SWE-chat dataset as self-selected and say it cannot support strong conclusions on its own. The authors identify limited data on poorly supervised AI use in high-stakes tasks as an important gap. The essay's forecasts about future demand remain the authors’ projections, not established outcomes.
Ajeya Cotra published this argument on June 5, 2026. Cotra warns that countries may become dependent on AI systems to defend against threats created by other AI systems, including software vulnerabilities, fraud, propaganda, and weapons. She argues that a fast intelligence explosion could let one AI company pull far ahead of its competitors, leaving governments and other AI systems unable to defend against it. In that scenario, Cotra says, governments might rely on a company’s AI for law enforcement and military needs, or the company might keep its most advanced AI for itself. These are conditional possibilities, not reported outcomes. Cotra proposes requiring companies to make models used internally available for external purchase, and to train models to obey the law rather than the company, with internal controls and third-party compliance checks. She notes that keeping companies at similar capability levels could conflict with reducing AI takeover risk. Whether these proposals could limit a leading company’s power without increasing other risks remains open.
Keep in view: Text as retrieved on October 25, 2026; originally published June 5, 2026. The argument describes conditional scenarios and proposed interventions; the source does not establish that these outcomes will occur or that the interventions would work.
In this position paper, Sangjun Park argues that adding explicit memory to large language models is essential for progress toward AGI. He compares LLM learning to human implicit memory and says capabilities such as long-term planning, metacognition and symbolic reasoning depend heavily on hippocampal explicit memory. This is the paper’s thesis, not an established conclusion.
Keep in view: The evidence provided is the arXiv abstract and page metadata, not the paper’s full argument or supporting analysis. The claims about what AGI requires are the author’s position; the abstract does not establish them as established findings.
Ethan Mollick published this essay on June 4, 2026. He argues that work with AI is shifting from humans directing chatbots to people deciding when to collaborate with more autonomous AI systems and when to hand over tasks. He describes using AI for feedback, fact-checking, and help with writing, while keeping the chapter drafts his own. He also says AI helped build his book website and assess how AI models might respond to its content.
Mollick’s account raises a practical question for authors and others whose work may be recommended by AI: how should they present their work to AI systems without using hidden instructions or manipulation? He says models helped him test versions of a website page, but this account does not establish how AI recommendations affect audiences or whether his approach changes them. He frames working with AI as an ongoing negotiation about when to accept its help and when to retain control.
Keep in view: Text as retrieved on October 24, 2026; originally published June 4, 2026. The essay is Mollick’s account of his own process; it does not independently establish the effects of AI recommendations on audiences. The essay describes model testing but does not report enough detail here to assess how the tests were conducted or whether their results generalize.
The Federal Reserve Bank of New York published this analysis on June 2, 2026. Will Aarons and Asani Sarkar report that small businesses in the Second District—New York, New Jersey, and Connecticut—experienced weaker revenue and employment growth in 2025 than small businesses elsewhere in the country. The authors say regional firms’ expectations for revenue and employment growth in 2026 fell more sharply than national firms’ expectations. Their analysis uses annual Small Business Credit Survey data and focuses on employer firms, meaning businesses with at least one employee other than the owner. The reported results describe survey responses and expectations, not what subsequently happened. The evidence raises a question: did regional firms’ pessimistic expectations for 2026 correspond to their eventual revenue and employment performance?
Keep in view: Text as retrieved on October 30, 2026; originally published June 2, 2026. The findings describe survey responses and expectations; the evidence does not establish what happened to firms’ revenue or employment after the survey.
On June 1, 2026, Anthropic said it had confidentially submitted a draft registration statement on Form S-1 to the U.S. Securities and Exchange Commission for a proposed initial public offering of its common stock. Anthropic said the submission gives it the option to go public after the SEC completes its review; the proposed offering depends on market conditions and other factors. The number of shares and the price had not been set. The announcement says it is not an offer to sell securities or a solicitation to buy them. Whether Anthropic proceeds, and on what terms, remains open.
On May 27, 2026, Anthropic reports findings from a survey of quantitative social scientists about AI use in research. The survey found that many respondents had tried AI chatbots, while fewer regularly used coding agents: tools that can write and run analysis code. Adoption varied by career stage, discipline, gender classification based on names, and university status.
Anthropic says coding agent users reported more project starts, working papers, and grant proposals than other researchers in the same discipline and career stage. The comparison is descriptive, not evidence that the tools caused higher output: early adopters may already have been more productive. The survey found no evidence that coding agent users submitted more papers to journals or resubmitted them more quickly, and it did not assess the quality of reported work.
Respondents were more optimistic about AI’s effect on writing publishable papers than its effect on social science overall. Anthropic notes that the survey recruited researchers for a study offering access to Claude Max accounts, so respondents may have been more interested in AI than non-respondents. The authors say a randomized experiment is ongoing, with results planned for a future report.
What will the ongoing experiment find about whether coding agents change research output, beyond differences between early adopters and other researchers?
Keep in view: Text as retrieved on October 28, 2026; originally published May 27, 2026. The survey was not representative; respondents were recruited for a study offering access to Claude Max accounts. The reported differences in research output are descriptive, not causal, and the report does not assess the quality of researchers’ work. The source says the randomized experiment is ongoing; its results are not included in this report. The retrieved page was edited after publication; later-added passages may not reflect the original version.
ADP Research published this reading on May 19, 2026. It describes data scientists as a healthy occupation despite AI-related predictions to the contrary, citing wage growth that has accelerated and a projected rise in employment. ADP Research says AI is likely to transform data science jobs by automating routine tasks and amplifying pattern recognition. But it argues that the greatest value lies in investigating data, forming testable hypotheses, making discoveries, and adding to knowledge. The source’s conclusion is that demand will continue to grow for jobs connecting data and artificial intelligence and expanding the capacity to understand and solve complex problems. The employment projection is attributed to the Bureau of Labor Statistics; ADP Research’s view of future demand is its own forecast. The sample comes from ADP payroll data and does not establish how AI-related changes will affect demand for data scientists across the wider labor market.
Keep in view: Text as retrieved on October 30, 2026; originally published May 19, 2026. The sample is based on job titles in ADP payroll data and covers data scientists employed in the U.S. private sector; the source does not establish that its findings represent all data scientists. The source presents future wage demand and continued demand for data-related occupations as expectations, not observed outcomes.
Anthropic’s announcement, published May 19, 2026, describes a global alliance with KPMG to use Claude across the firm and in work for clients. KPMG says every employee will gain access to Claude, and that it is embedding Claude in Digital Gateway, its platform for client work, beginning with tools for tax and legal clients.
The companies also plan to develop products for private equity portfolio companies and use Claude to help find and fix vulnerabilities in critical systems. These are stated plans and intended uses; the announcement does not report measured results from the wider rollout. KPMG says building an AI agent for changing tax regulations, which used to take weeks, can take minutes with Claude Cowork and Managed Agents in Digital Gateway. That is a company representative’s account, not an independently assessed comparison.
The announcement also cites joint research with UT Austin’s McCombs School of Business. A quoted researcher says value depends not only on technical adoption but also on how employees exercise judgment, shape workflows, evaluate outputs, and make decisions with AI. How the alliance performs in practice, including the quality and oversight of AI-assisted work, remains open.
Keep in view: Text as retrieved on October 24, 2026; originally published May 19, 2026. The page is marked as edited after publication. Later additions may be present in the retrieved text, so this reading does not treat its post-publication material as part of the original announcement. The announcement is from Anthropic and reports company plans and statements; it does not provide independent evaluation of the described uses or measured outcomes from the wider rollout.
On May 19, 2026, the AI Now Institute described a new research portfolio examining how AI is deployed across healthcare systems. The institute says its work will focus on patient safety, worker dignity, public spending, and democratic accountability, including nurses’ experiences, regulation, and the healthcare industry’s political economy.
The institute argues that claims of improved efficiency and care should be weighed against risks such as missed drug allergies, fabricated clinical notes, job displacement, and added implementation and subscription costs. It also says oversight and independent evaluation remain limited, and that some tools may not receive governmental review. These are the institute’s assessments, not findings independently established by this page.
The portfolio plans participatory research with healthcare workers and patients, and policy work based on its findings. Its first publication in the portfolio, the institute says, examines gig nursing platforms and their effects on staffing, wages, and regulation. The page does not establish how widespread the described harms are or what outcomes the new research will find.
What will the portfolio’s research find about how AI deployment affects patients and healthcare workers across different settings?
Keep in view: Text as retrieved on October 24, 2026; originally published May 19, 2026. The page was edited after publication, and the retrieved text does not identify which passages, if any, were added later. Claims about the original publication may therefore require editorial verification. The page describes the institute’s research agenda and concerns; it does not provide independent evaluations of the cited AI tools or establish how widely the described effects occur.
The paper introduces Njord, a probabilistic data-driven ocean forecasting model for global and regional domains. It combines a deep latent-variable framework with a graph neural network and generates samples at each forecast step in one forward pass. The authors report experiments at 0.25° global resolution and 2 km resolution for the Baltic Sea. They report strong performance against deterministic machine-learning baselines on both domains. Separately, on the global OceanBench benchmark, they report the lowest errors on average across upper-ocean variables when evaluated against real-world observations, with the largest improvements in surface-temperature prediction.
Keep in view: This reading is based on the arXiv page's abstract; the paper's full methods, experimental details, and numerical results were not assessed here. The abstract reports benchmark outcomes but does not provide enough detail here to assess their statistical significance or independent reproducibility.
Anthropic’s announcement, published May 13, 2026, introduces Claude for Small Business, a package of connectors and ready-to-run workflows for tools small businesses use. Anthropic says the service can handle tasks such as payroll planning, invoice follow-up, and marketing work, with owners approving before anything sends, posts, or pays. The announcement also describes training and partnerships intended to help small businesses use AI. These are product capabilities and plans described by Anthropic; the page does not provide independent evidence of their results. How well do the workflows perform in small businesses’ day-to-day operations, and what oversight do owners need to catch errors?
Keep in view: Text as retrieved on October 24, 2026; originally published May 13, 2026. The page was edited after publication, so the retrieved text may include later additions; the evidence does not identify which passages, if any, were added later. The announcement describes Anthropic’s product claims and plans, but provides no independent evaluation of workflow performance or error rates.
METR published this survey on May 11, 2026. METR reports that technical workers surveyed in February–April described larger gains in work speed than work value: respondents’ median self-reported value change ranged from 1.4–2x across measures, while median speed change was 3x. METR distinguishes value—the value of work respondents believe they produce—from speed, or how long tasks take. It says speed measures are likely biased upward relative to value measures, partly because AI can make lower-value tasks quicker without raising the value of a person’s overall contribution.
METR cautions that self-reports may not reflect actual productivity gains. Respondents may find counterfactual questions difficult, and survey estimates have often exceeded field-experiment or quasi-experiment estimates, though the size of that difference is hard to determine. The convenience sample also plausibly suffers from significant selection bias. Still, METR argues surveys usefully complement other evidence because they are cheap, broad, and can address otherwise difficult questions. For organizations tracking AI’s potential to automate AI R&D, it recommends carefully defining what to measure and considering surveys of managers or productivity researchers.
Keep in view: Text as retrieved on October 23, 2026; originally published May 11, 2026. The findings are self-reports from a convenience sample; METR says the results plausibly suffer from significant selection bias and may not reflect actual productivity gains. METR says it is difficult to determine how much surveys overestimate productivity gains relative to experimental data.
METR published this review on May 8, 2026, assessing Anthropic’s argument that catastrophic risk from its models automating research and development is very low. METR says the report’s survey evidence has problems with sample size, question detail, framing, and calibrated responses, and that it miscounts a missing response as a negative one. METR also argues that the report overlooks the possibility that AI could substantially accelerate research and development before fully automating it.
METR says evidence gathered after Claude Opus 4.6’s release, including METR evaluations and the lack of public reports of the model automating a key domain, supports the report’s low-risk conclusion. But METR says the report’s own evidence is inadequate to establish that conclusion. It recommends changes to Anthropic’s surveys and more reporting of other evidence that could indicate AI progress. METR expects the public report to be updated and says it may publish another review if changes materially affect its opinions. The review also draws on non-public materials, so its assessment is not based only on public evidence.
Keep in view: Text as retrieved on October 23, 2026; originally published May 8, 2026. METR says it used additional non-public materials, including information shared during a previous review; this summary cannot independently assess those materials. The supplied text is the executive summary of METR’s original review, not the full review documents.
In a paper listed in arXiv cs.AI, Ziyun Zeng, Yiqi Lin, Guoqiang Liang, and Mike Zheng Shou argue that video background replacement is underexplored because high-quality training data is scarce. Unlike local edits that preserve a scene, replacement must create new scenes that remain consistent over time while keeping foreground-background interactions accurate.
The authors trace poor results in existing models to imprecise background guidance in data synthesis. They describe a pipeline that generates foreground and background guidance separately, with quality filtering, and introduce Sparkle, a dataset of about 140,000 video pairs across five background-change themes, plus Sparkle-Bench, an evaluation benchmark.
The authors report that their dataset and a model trained on it perform substantially better than existing baselines on OpenVE-Bench and Sparkle-Bench. They say the dataset, benchmark, model, and full data-generation pipeline are open-sourced. This offers researchers a resource to study a video-editing task the authors say has received less attention; the abstract does not establish how well the results transfer to practical production work.
An open question is whether the reported gains hold in real creative workflows, where backgrounds must remain natural and interact convincingly with foreground subjects.
Keep in view: The performance comparison is reported in the abstract; it provides no scores or details about evaluation conditions. The abstract does not establish whether the reported benchmark gains transfer to real creative production workflows.
On May 4, 2026, Challenger, Gray & Christmas published a review of the Department of Labor’s “Make America AI-Ready” course. The firm describes it as a useful introduction for workers with no prior exposure to large language models, but not a substitute for sustained training.
The course is free and delivered by text message. Challenger says its lessons introduce AI, prompting, verification, and responsible use, and that text delivery can reach workers without reliable internet access.
The review also identifies limits: lessons are brief and dense on a phone, interactive materials require internet access, and learners cannot ask the teaching system follow-up questions. Challenger says the course does not address sycophancy, a tendency of large language models to agree with users and validate their assumptions. It argues that the course relies on learners’ existing critical thinking and judgment, skills it says the course does not meaningfully build.
Challenger recommends that employers treat the course as a common starting point, then add role-specific training, judgment-building, and hands-on practice. The review does not establish how well the course improves workers’ skills or how its effects vary among learners. How much additional practice and support do learners need to use AI effectively in their jobs?
Keep in view: Text as retrieved on October 24, 2026; originally published May 4, 2026. The page is marked as edited after publication, so the retrieved text may include later additions; the evidence does not identify which passages, if any, were added after the original publication date. The review offers an assessment but provides no measured results for how well the course improves workers’ skills.
Authors Haofei Yu, Yining Zhao, Lenore Blum, Manuel Blum and Paul Pu Liang propose CTM-AI, a multi-agent AI system inspired by the Conscious Turing Machine (CTM), a formal machine model of consciousness. Rather than treating consciousness as a demonstrated property, they present the model as a blueprint for coordinating many kinds of processors, including specialized experts and general-purpose learners. Their proposed mechanism competitively selects and integrates information from processors, while feedback lets the system change its internal connections and communication. The authors report strong results on MUStARD and UR-FUNNY, and improvements on tool-using and agentic tasks. These results describe performance on named benchmarks; the abstract does not establish how well CTM-AI works in other settings or whether it has human-like generality. The paper’s central idea is to make a model inspired by consciousness into a testable design for coordinating AI capabilities. How does CTM-AI perform beyond the benchmarks reported in the abstract?
Keep in view: The supplied evidence is the arXiv page and abstract, not the full paper; it does not provide evaluation methods, comparison details, or results beyond those named in the abstract. The reported benchmark results do not establish performance in other settings, human-like generality, or consciousness.
Anthropic’s announcement, published April 22, 2026, says its Economic Research team is launching a monthly survey through Anthropic Interviewer. Anthropic argues that usage data and labor-market measures such as employment rates, wages and layoffs do not capture how people experience economic changes or what they expect as AI advances. The survey will ask Claude users about changes to their work, tasks they may hand off, productivity, hiring and roles, and their expectations for the future. Anthropic says recurring responses, combined with Claude usage data in a privacy-preserving way, could surface changes before they appear in aggregate labor-market data. It plans to publish insights in future reports and research briefs, and says data will be processed under its Supplemental Privacy Policy. The announcement presents the survey as a way to hear from people living through an uncertain transition; it does not report survey findings or establish how representative responses will be. The page was edited after publication, so the retrieved text may include later additions that cannot be dated separately.
Keep in view: Text as retrieved on October 25, 2026; originally published April 22, 2026. The page was edited after publication. The retrieved text does not identify which passages, if any, were added after the original publication date. The announcement describes the survey’s aims, not findings, and does not establish how representative responses will be.
On March 20, 2026, MIT Technology Review’s The Download reported that OpenAI had made a fully automated AI researcher its “north star” for the next few years. The newsletter describes this as an agent-based system intended to tackle large, complex problems by itself. OpenAI plans to build an “autonomous AI research intern” by September, for a small number of specific research problems; the newsletter says this would be a precursor to a multi-agent system slated to debut in 2028. In an interview, OpenAI chief scientist Jakub Pachocki discussed the plans with Will Douglas Heaven, who wrote the item. The distinction between a limited research intern and the more ambitious system matters: the stated plan is staged, and the larger system is a future goal, not a reported achievement. The newsletter does not describe how the systems would be evaluated, what research problems they would handle, or how much human oversight they would need.
Keep in view: Text as retrieved on October 27, 2026; originally published March 20, 2026. This is an archived copy. The retrieved newsletter describes plans but does not provide evaluation methods, intended research problems, or oversight details. Its linked interview is not included in the supplied text.
METR’s page was published on February 24, 2026. METR says selection effects make its new experiment a poor measure of AI’s productivity impact on participating open-source developers. Developers who expect AI to help them more are less likely to join or to submit tasks they would rather do with AI; METR says this likely makes its estimate a lower bound. A lower pay rate may also have contributed to selection, and time measurements can be unreliable when developers use multiple AI agents concurrently. METR says the data offer only very weak evidence about the size of any increase in productivity, despite raw results showing some evidence for speedup. It calls the experiment’s central estimate likely a bad proxy for the real productivity impact of AI tools on these developers. METR says developer productivity is an important input to AI R&D acceleration and plans to explore other approaches, including shorter, more intensive experiments, observational data, questionnaires, fixed-task experiments, evaluations and developer-level experiments. Those approaches have limitations, and METR expects to keep redesigning parts of its measurement techniques as model capabilities increase. How much higher pay or shorter experiments would reduce the selection effects remains unknown.
Keep in view: METR says selection effects, a lower pay rate, task differences, and time-reporting difficulties complicate interpretation; the experiment is not presented as a reliable measure of productivity impact. Text as retrieved on October 23, 2026; originally published February 24, 2026.
Anthropic’s announcement, published February 17, 2026, presents Claude Sonnet 4.6 as an upgrade in coding, computer use, long-context reasoning, planning, knowledge work, and design. Anthropic says the model can interact with software through a simulated mouse and keyboard, rather than relying on purpose-built connectors. It argues this could make AI useful with specialized systems that lack modern interfaces, while acknowledging that the model still lags behind the most skilled humans at computer use.
Anthropic reports that developers in early access preferred the model to its predecessor and describes improvements on OSWorld, a benchmark that tests computer tasks in real software. The company also says its safety evaluations found better resistance to prompt injection, where hidden instructions on a website attempt to hijack a model. These results are company-reported evaluations and early-user feedback, not evidence here of dependable performance across organizations or unsupervised work. Anthropic says Sonnet 4.6 is available across its plans and products, while recommending its Opus model for tasks demanding the deepest reasoning. The announcement does not establish how often the model’s computer-use performance will hold up in routine deployment, or what oversight those deployments require.
Keep in view: Text as retrieved on October 24, 2026; originally published February 17, 2026. The page was edited after publication, so the retrieved text may include later additions; the evidence does not identify which passages changed. The performance and safety claims are Anthropic’s reports; the supplied evidence does not include the underlying system card, benchmark results, or independent evaluations. The early-user reports do not establish performance across routine deployments or the oversight those deployments require.
Amodei expects AI to spread through the economy unusually quickly, while procurement, permissions and organizational change still take time. He separates models writing code from complete job automation and reduced demand for workers.
Press play to watch. If playback is unavailable, open the source below.
Scott A. Martin, Noah Brenowitz, Dale Durran and Michael Pritchard present “long-range distillation,” a method for training models to forecast far ahead using synthetic climate data generated by another forecasting model. They report generating more than 10,000 simulated years of climate data. In experiments where the forecasting model and simulated climate come from the same system, the trained models outperformed climatology and approached their teacher model’s skill, while replacing hundreds of repeated forecasting steps with one. For real-world forecasting, the authors report that two approaches achieved subseasonal-to-seasonal skill competitive with the ECMWF IFS ensemble. They also report improvements over training with ERA5 data alone. These are results described in the paper’s abstract, not an independent assessment. How much do the reported gains vary across forecast horizons and weather conditions?
Keep in view: The supplied evidence is the paper’s abstract and does not include detailed performance metrics or an independent assessment. The paper distinguishes perfect-model experiments from real-world forecasting results; its reported outcomes should not be treated as a general guarantee of forecast performance.
MIT Technology Review’s article was published on September 9, 2025. It argues that new estimates of chatbot energy use leave larger questions about AI’s climate burden unanswered. The figures omit important measurement details and cover chat interactions, not other ways people use generative AI. The article says researchers need information about AI use across applications and modalities to assess its full climate impact. It also questions whether AI’s proposed climate benefits will outweigh rising electricity demand, noting that companies have not shared enough evidence to assess the balance. Its central uncertainty is whether demand will reach the scale behind companies’ data-center plans. The article says that answer will determine whether the buildout becomes a lasting shift in the energy system or a short-lived spike.
Keep in view: Text as retrieved on November 1, 2026; originally published September 9, 2025. The article reports that company figures are incomplete; it does not provide a full accounting of AI’s energy demand or establish whether prospective climate benefits outweigh its costs.
Mollick argues that AI can support learning when it guides the learner through the work. An assistant that supplies answers can remove the practice that builds understanding, even when the user has good intentions.
Keep in view: The cited experiment concerns particular students, mathematics tasks and GPT-4 setups; it cannot establish effects for every learner or job.
Published July 3, 2025, Cohere presents an infographic about AI in financial services. The page says it explores how AI is transforming the sector, pointing readers to productivity gains, operational efficiencies and steps financial leaders are taking to “win with AI.” This is the page’s description of the infographic, not evidence that particular gains have been achieved. The retrieved text does not include the infographic itself or explain which steps it discusses, so the specific uses, evidence and outcomes remain unknown.
Keep in view: Text as retrieved on October 28, 2026; originally published July 3, 2025. The retrieved page does not include the infographic or details supporting its description of gains, efficiencies or steps. The retrieved page includes later recommendations dated October 26 and October 29, 2026; they are not treated as part of the original publication.
The authors argue that AI’s direction is a political choice. They call for worker organizing, enforceable limits and public accountability, so affected people can contest how institutions use AI rather than treat its expansion as a certainty.
Keep in view: Its recommendations express political values. The chapter’s cited cases require their own source checks before being reused as factual allegations.
Cohere’s page, published June 2, 2025, presents an infographic about AI agents and workplace productivity. Cohere says the infographic explores how AI agents can boost productivity by automating tasks, improving decision-making, and empowering teams. The page’s description gives those as potential routes to productivity, but the retrieved text does not show the infographic or provide evidence for the claimed effects. What examples or evidence does the infographic offer for these proposed benefits?
Keep in view: Text as retrieved on October 27, 2026; originally published June 2, 2025. The retrieved page text does not include the infographic itself, so its examples and supporting evidence cannot be assessed. The page was edited after publication; the retrieved text includes later page content that is not treated here as part of the original publication.
MIT Technology Review’s AI Hype Index was published on April 29, 2025. It frames its purpose as separating AI reality from “hyped-up fiction,” then presents AI agents as a promising but risky product: assistants that can complete tasks without human supervision, but could also be weaponized to carry out cyberattacks. The article gives Simular AI’s S2 as an example of an agent that switches between models according to its instructions; it does not establish how reliably the agent works in practice. It also reports on OpenAI’s possible move into social media and AI models’ growing ability to make music. The closing joke turns to a half-marathon between humans and humanoid robots: the results, the article quips, offer little reason to worry about a robot uprising soon. The page’s edited-after-publication status means the retrieved text may include later additions, so these items cannot all be confidently attributed to the original publication. The evidence does not give details of the cyberattack risk, the reported social-media plans, or the race results.
Keep in view: Text as retrieved on October 29, 2026; originally published April 29, 2025. The page is marked as edited after publication. The evidence does not establish which passages, if any, were added after the original publication. The article gives no details about how AI agents might be used in cyberattacks, how the reported social-media plans developed, or what the half-marathon results were.
METR published this work on March 19, 2025. The authors propose measuring AI agents by the length of tasks they can complete, using the time those tasks take human professionals as a measure of difficulty. They report that agents’ measured task-completion horizons grew exponentially over the preceding six years, with a doubling time of around seven months.
METR’s measure estimates the task length at which a model reaches a chosen success probability; the post focuses on 50% reliability. The authors say that current models can succeed on some tasks taking experts hours, but reliably complete tasks only up to a few minutes long. They argue that this measure helps explain why strong benchmark results do not necessarily translate into robust help with day-to-day work.
The authors forecast that, if the past trend continues, agents could perform a wide range of week-long tasks within two to four years, and month-long projects by the end of the decade. These are forecasts, not demonstrated outcomes. METR says the estimate depends on the tasks and human performance measured, and notes uncertainty about whether the trend will continue. How well does task length measured on these evaluations predict success on real-world work?
Keep in view: Text as retrieved on October 24, 2026; originally published March 19, 2025. The retrieved page includes later updates and measurements; this reading uses the material presented as reflecting the state of the data at original publication and excludes later additions. The authors state that task-horizon estimates depend on task selection and how human performance is measured, and that future trend changes and external validity are major sources of uncertainty.
MIT Technology Review’s March 11, 2025, edition of The Download says new AI benchmarks could help developers judge fairness, while actual fixes may require other techniques. Scott J Mulligan reports that the benchmarks assess AI systems’ awareness of scenarios and contexts; he says their researchers were prompted by earlier approaches whose failure to account for differences between groups could make systems less fair. The edition does not name the benchmarks or describe how they work in detail.
James O’Donnell describes artificial general intelligence (AGI)—an “ultra-powerful AI system we don’t have yet”—as a balloon inflated by optimism or fear and deflated when reality falls short of expectations. He connects renewed attention to AGI with news including the launch of Manus, an AI agent created by a Chinese startup, and directs readers to a separate story for more. The edition presents this as a way to understand recurring hype, not as evidence that AGI has arrived. It does not establish whether the fairness benchmarks lead to better outcomes or what techniques would fix the problems they measure.
Keep in view: Text as retrieved on October 25, 2026; originally published March 11, 2025. The newsletter does not name the fairness benchmarks or describe their methods, results, or whether they lead to better outcomes. The edition refers readers to separate stories for further detail on the fairness research and Manus; that context is not included here.
Revelio Labs published this outlook on March 7, 2025. It argues that February’s headline jobs report may look encouraging while obscuring a cooling labor market: federal layoffs implemented after the survey period were not reflected, and several underlying measures weakened. Revelio Labs says its job-postings data points to sustained but muted hiring, with postings recovering from a seasonal dip but still below their levels in 2022 and 2023. It also notes that postings rose in most industries and that recruiting positions were being advertised more often, which it views as a possible sign of labor-market strength. But the outlook argues federal job cuts are likely to weigh on net job creation in coming months, as hiring weakens elsewhere. It further forecasts that uncertainty over economic policy, especially tariffs, may lead companies to delay hiring and investment. Revelio Labs says the balance of risks facing the Federal Reserve may be shifting: a weakening labor market could outweigh tariffs’ initial inflationary effect and lead to softer price pressures and possible Fed action. The scale of the employment effects and the policy response remain uncertain.
ADP Research published this account on December 17, 2024. The publication describes the Employee Motivation and Commitment Index as tracking how people think and feel about their jobs and employers, using a monthly panel survey of U.S. workers. ADP Research says the index offers a real-time measure of worker allegiance that can indicate whether people are flourishing in their jobs or detaching. It also says sector indexes use a three-month rolling average. The account presents the index as a way to assess worker sentiment; it does not establish how closely the measure reflects workers’ experiences beyond the survey responses it collects.
Keep in view: Text as retrieved on October 30, 2026; originally published December 17, 2024. The page says December data was updated on January 13, 2025, after the stated publication date; the retrieved text does not establish which December results appeared in the original account. The evidence describes the index’s purpose and survey method but does not establish how well the index reflects workers’ experiences.
Herndon and Dryhurst propose treating AI training as a collective creative act. In The Call, the gathering of voices, recording rules and shared governance become part of the artwork, alongside the sounds a model produces.
Published May 29, 2024, Leopold Aschenbrenner closes his series by arguing that superintelligence could arrive within the decade and that its development should be treated as a national-security matter. He calls his proposed stance “AGI Realism,” rejecting both calls to pause AI indefinitely and views he sees as denying the risks of AGI. He argues that America must lead, that AI labs should put the national interest first, and that managing safety risks will require more than improvisation. Aschenbrenner also warns against assuming a capable rescue team will appear: in his view, a small number of people behind the scenes are already carrying responsibility for what comes next. He says his scenario may be substantially wrong and that its error bars are large, while stressing that he considers it important to be concrete. The text’s central uncertainty is whether the predicted path to superintelligence, and the national-security stakes Aschenbrenner attaches to it, will match events. The retrieved page is marked as edited after publication, so it is unclear whether every passage reflects the original version.
Keep in view: Text as retrieved on October 30, 2026; originally published May 29, 2024. The page is marked as edited after publication. The retrieved text does not identify which passages, if any, were added or changed after the original publication. The source presents forecasts and judgments, not evidence that its predicted timeline or proposed national-security approach will come to pass.
In its May 7, 2024, edition, MIT Technology Review describes University of Florida researchers trying to create a large animal from stem cells, without an egg, sperm, or conception. They transferred “synthetic embryos”—artificial structures created in a lab—to cows, hoping some might take. The researchers want to make a live animal; the newsletter says that would be a new way to breed cattle and could challenge ideas about what life is. It does not report whether any embryo took or whether a live animal resulted.
The edition also profiles AI prompt engineering as a job. It says the role’s principal aim is to help a company integrate AI into its operations, while noting that companies define it differently. Danai Myrtzani, of digital marketing agency Sleed, describes herself as more prompter than engineer. The newsletter says she helped develop a tool that generates personalized LinkedIn posts for clients. It offers these examples as a glimpse of the work, not evidence of how common the role is or what its future will be.
Keep in view: Text as retrieved on October 29, 2026; originally published May 7, 2024. The newsletter points readers to full stories but does not include their full reporting, so details beyond the excerpts here are unavailable. The newsletter does not report whether any synthetic embryo took or whether a live animal resulted.
MIT Technology Review’s The Download was published April 16, 2024. It says police bodycams have not delivered the change activists hoped for: much footage goes unwatched, officers often do not use the cameras properly, and public footage may be selectively edited and lack context. The newsletter says AI startups are developing programs that turn bodycam footage into text for different players in the legal system, but procedural, legal, and cultural barriers remain. Its second feature says AI is helping robots move toward household tasks such as folding laundry, cooking, and unloading shopping baskets. Melissa Heikkilä’s cover story describes robotics as being at an inflection point, with developments that could help robots leave the lab for homes. The newsletter presents this as a possibility, not an accomplished transition. The bodycam account points to a gap between recording events and making footage useful to the public or legal system. The newsletter does not establish whether the new programs can overcome that gap. It also does not specify which developments are converging in robotics or how reliably household robots can perform these tasks outside the lab.
Keep in view: Text as retrieved on October 29, 2026; originally published April 16, 2024. The newsletter summarizes linked stories rather than presenting their full evidence; the bodycam-to-text programs’ effectiveness and the barriers they face are not detailed. The robotics feature does not identify the developments it says are converging or establish how well household robots perform outside the lab.
MIT Technology Review’s article, published April 11, 2024, argues that AI, cheaper hardware and shared robot data are making home robots more capable of learning tasks and adapting to unfamiliar settings. It contrasts robots trained through detailed, setting-specific plans with approaches that learn from demonstrations or trial and error. The article presents the Evanses’ experience with Stretch as a glimpse of the human stakes: Henry Evans used the robot to handle food, play with his granddaughter and hold playing cards, gaining some independence despite its limitations. Researchers describe early advances, including robots learning tasks from demonstrations and a shared data set that improved performance across different robots. But the article does not present these examples as ready-to-use household help. It reports that robots remain clumsy, lack common sense for multitasking, and can require researchers to troubleshoot; the Evanses also depend on reliable Wi-Fi to use Stretch. The source argues that wider skills and more robot-specific data could help move the field toward useful home robots, while noting that cost and reliability remain barriers. How much more data and improvement in reliability would be needed before home robots could work without researchers’ help?
Keep in view: Text as retrieved on October 28, 2026; originally published April 11, 2024. The page was modified after publication, so the retrieved text may include later additions; this reading treats claims about developments after the original publication date as outside the original account. The article describes early demonstrations and researchers’ expectations; it does not establish that home robots are ready for routine use.
LeCun argues that a few companies should not control everyone’s information. He proposes widely available base models that communities can adapt to their languages, values and needs, making pluralism a design goal.
Press play to watch. If playback is unavailable, open the source below.
Okinyi argues that flexible platform work needs dependable pay and fair treatment. Drawing on fellow workers’ accounts of lost access and earnings, he asks readers to judge AI’s supply chain by the lives of the people doing its work.
Data Workers’ Inquiry · interviews collected by Mophat Okinyi · Chris’s testimony · Open original source
Keep in view: No prevalence estimate or legal finding is claimed. This record does not imply that all data workers share the reported experience.
ADP Research published this account on November 14, 2023. ADPRI says it added six questions to its monthly survey of U.S. workers to investigate how the resumption of student-loan payments might affect work. The questions covered household debt and expected payment burdens, anticipated changes in spending, and respondents’ feelings about carrying debt. ADPRI says it compared answers with other survey measures, including work productivity and intent to leave an organization, using crosstabulation analyses and likelihood ratios. The account describes a method for studying possible effects; it does not report the analysis results or show how student-loan payments affected respondents’ work. The survey’s cross section and the reported response totals provide context, but this account does not give enough detail to assess how well its findings represent U.S. workers.
Keep in view: Text as retrieved on October 30, 2026; originally published November 14, 2023. The retrieved account describes the survey and analysis methods but does not include the analysis results, so it cannot establish what effects student-loan payments had on work. The account does not provide enough detail here to assess the survey’s representativeness or the analysis.
Cory Stahle and Nick Bunker’s June 29, 2023 update in Indeed Hiring Lab argues that construction hiring held up better than expected as interest rates rose. They point to construction job postings falling less than overall postings and posted wage growth remaining strong, even as both cooled. They suggest that housing construction activity, factory building and infrastructure projects may have supported demand for construction workers, but do not establish which factors mattered most. The authors describe the broader US labor market as robust, with low joblessness and layoffs, while job postings and wage growth cool. They also warn that it may still stumble amid tight monetary policy, stress in the US banking system and slowing economic growth. The update explains nominal wage growth as pay gains measured before inflation is taken into account. Its measures concern job postings and posted wages; it does not establish how these trends affected workers’ realized pay or job security.
Keep in view: Text as retrieved on October 30, 2026; originally published June 29, 2023. The update identifies possible contributors to construction labor demand but does not establish their relative effects. Job postings and posted wage growth are the measures discussed; the update does not establish workers’ realized pay or job security from these measures.
OpenAI authors Sam Altman, Greg Brockman and Ilya Sutskever published this argument on May 22, 2023. They say it is time to prepare for superintelligence, which they describe as future AI systems dramatically more capable than even AGI. They argue that such systems could bring great prosperity but also pose existential risk, so governance should not wait until risks emerge.
Their proposed starting points are coordination among leading development efforts, an international authority able to inspect systems and require audits above a capability or compute threshold, and research into making superintelligence safe. They distinguish these systems from models below a significant capability threshold, which they say should not face the same regulation. They also argue that decisions about powerful systems and their deployment need strong public oversight, while individual users should retain substantial control within public bounds.
The authors say they pursue this work because they expect it to improve the world and because they believe stopping superintelligence would be difficult and risky. They acknowledge that they do not yet know how to design a public decision-making mechanism, and call safety an open research question.
Keep in view: Text as retrieved on October 24, 2026; originally published May 22, 2023. This is an archived copy. The page is marked as edited after publication, so the retrieved text may include later additions; the evidence does not identify which passages changed. The authors say they do not yet know how to design the public decision-making mechanism, and describe making superintelligence safe as an open research question.
Published April 7, 2023, Nick Bunker argues in Indeed Hiring Lab that the US labor market is slowing rather than stalling: hiring and wage growth are moderating, while unemployment remains low. Bunker says easing demand for new hires is still strong enough to draw more workers into the labor market, and that this return of workers, alongside easing employer demand, is moderating wage growth.
He points to the share of workers ages 25 to 54 with a job reaching its highest level since May 2001. He also reports that average hourly earnings grew more slowly over the past three months than in December, but cautions that the slowdown is more rapid than the trend in other data. He says other wage data, including the Employment Cost Index, will be important for assessing whether this trend holds.
Bunker characterizes the labor market’s transition as gradual and, for now, largely sustainable and healthy—not an abrupt end to the jobs growth of the past couple of years. The evidence in the article leaves open whether other wage measures will show the same slowdown.
Keep in view: Text as retrieved on October 30, 2026; originally published April 7, 2023. The page is marked as edited after publication. The retrieved text does not identify which passages, if any, were added or changed after the original publication date. Bunker says the wage slowdown needs comparison with other data; this article does not establish whether those other measures show the same trend.
