MIT Technology Review · Jan 20, 2025
MIT Technology Review reports a shift toward AI systems that build and maintain software
Markdown · JSON · Original source
MIT Technology Review’s article was published on January 20, 2025. It describes a second wave of AI coding tools aimed not just at producing code that runs, but at producing code that does what a developer intended. The article says companies including Cosine, Poolside, Zencoder, and Merly are pursuing different methods, from recording developers’ steps to training models through code execution or working with an intermediate representation. Their approaches differ, and Merly founder Justin Gottschlich argues that are the wrong tool for programming. The article presents a possible change in software work: developers may spend more directing and reviewing machine-written code, while tools may take on tasks such as testing, debugging, and fixing reported bugs. Cosine cofounder Yang Li says human review could become the bottleneck, and predicts that some companies will use the technology to hire fewer programmers. That is a forecast, not an established outcome. The article also connects these products to their builders’ ambitions for , while noting that such systems remain a pipe dream. Whether the different approaches can reliably produce useful software, and how widely they may change hiring and developers’ work, remains unknown.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
MIT Technology Review · Article
Source details
Original source date: Jan 20, 2025 (2025-01-20T11:52:59Z).
Date basis: Date from the publisher’s own records, through its API.
Added to Escape Meridian: 2026-11-16T00:53:26.137Z. First observed: 2026-11-01T01:02:01.183Z. Retrieved: 2026-11-01T01:03:19.848Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Text as retrieved on November 1, 2026; originally published January 20, 2025. The article reports company and individual claims about products, capabilities, and plans; it does not establish how broadly the described effects will occur. The article’s account of Cosine, Poolside, Zencoder, and Merly does not provide a common evaluation of their approaches or real-world outcomes.
Open question
Can these systems reliably produce software that meets developers’ intentions, and how much will their adoption change programmers’ work and hiring?
Source-stated
MIT Technology Review describes a new generation of coding assistants aimed at making programs function as intended, rather than merely producing code that runs.
It’s that second level of correctness that the new wave of generative coding assistants are aiming for
Source-stated
The article describes competing methods, including Cosine’s collection of developer work steps and Poolside’s use of reinforcement learning from code execution.
Poolside, based in San Francisco, is also creating a synthetic data set that captures the process of coding, but it leans more on a technique called RLCE—reinforcement learning from code execution.
Source-stated
Merly CEO and founder Justin Gottschlich argues that large language models are the wrong tool for programming and describes Merly’s different approach.
To Justin Gottschlich, CEO and founder of Merly, large language models are the wrong tool for the job—period.
Source-stated
Cosine cofounder Yang Li says that as software grows, humans’ ability to review machine-generated code may become a bottleneck.
“The bottleneck will become how fast humans can review the machine-generated code,” says Li.
Source-stated
Li predicts that some companies will use the technology to reduce programmer hiring and that smaller teams may take on work once done by hundreds.
At the other end, smaller teams of 10 to 20 people will do a job that once required hundreds of coders.
Source-stated
The article says these systems remain a pipe dream as tools and roles in software engineering change.
Down here on Earth, such systems remain a pipe dream. At the same time, software engineering is changing faster than many at the cutting edge expected.
Source and contributor context
MIT Technology Review · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Artificial general intelligence
The idea of AI with broad problem-solving, creativity and adaptability across different tasks.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: d6f3e150-d38d-45c5-8dc5-b5b828ecbd9f.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T00:48:51.998Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model.
Observation: d584a6008a99a199b71eb57dc829d7548f34f4cc89d619ebb73b029bd9bb8819. Source text hash: 251809828dedb390bf20bb658e20fa510ef2b50bc3aa9de77177b0719d0ecdac.
Revision and source-check history
Version 1 · 2026-11-16T00:53:26.137Z
MIT Technology Review reports a shift toward AI systems that build and maintain software
MIT Technology Review’s article was published on January 20, 2025. It describes a second wave of AI coding tools aimed not just at producing code that runs, but at producing code that does what a developer intended. The article says companies including Cosine, Poolside, Zencoder, and Merly are pursuing different methods, from recording developers’ steps to training models through code execution or working with an intermediate representation. Their approaches differ, and Merly founder Justin Gottschlich argues that are the wrong tool for programming. The article presents a possible change in software work: developers may spend more directing and reviewing machine-written code, while tools may take on tasks such as testing, debugging, and fixing reported bugs. Cosine cofounder Yang Li says human review could become the bottleneck, and predicts that some companies will use the technology to hire fewer programmers. That is a forecast, not an established outcome. The article also connects these products to their builders’ ambitions for , while noting that such systems remain a pipe dream. Whether the different approaches can reliably produce useful software, and how widely they may change hiring and developers’ work, remains unknown.
Original source
MIT Technology Review · Article
Source details
Original source date: Jan 20, 2025 (2025-01-20T11:52:59Z).
Date basis: Date from the publisher’s own records, through its API.
Added to Escape Meridian: 2026-11-16T00:53:26.137Z. First observed: 2026-11-01T01:02:01.183Z. Retrieved: 2026-11-01T01:03:19.848Z.
Limits
Text as retrieved on November 1, 2026; originally published January 20, 2025. The article reports company and individual claims about products, capabilities, and plans; it does not establish how broadly the described effects will occur. The article’s account of Cosine, Poolside, Zencoder, and Merly does not provide a common evaluation of their approaches or real-world outcomes.
Open question
Can these systems reliably produce software that meets developers’ intentions, and how much will their adoption change programmers’ work and hiring?
Source-stated
MIT Technology Review describes a new generation of coding assistants aimed at making programs function as intended, rather than merely producing code that runs.
It’s that second level of correctness that the new wave of generative coding assistants are aiming for
Source-stated
The article describes competing methods, including Cosine’s collection of developer work steps and Poolside’s use of reinforcement learning from code execution.
Poolside, based in San Francisco, is also creating a synthetic data set that captures the process of coding, but it leans more on a technique called RLCE—reinforcement learning from code execution.
Source-stated
Merly CEO and founder Justin Gottschlich argues that large language models are the wrong tool for programming and describes Merly’s different approach.
To Justin Gottschlich, CEO and founder of Merly, large language models are the wrong tool for the job—period.
Source-stated
Cosine cofounder Yang Li says that as software grows, humans’ ability to review machine-generated code may become a bottleneck.
“The bottleneck will become how fast humans can review the machine-generated code,” says Li.
Source-stated
Li predicts that some companies will use the technology to reduce programmer hiring and that smaller teams may take on work once done by hundreds.
At the other end, smaller teams of 10 to 20 people will do a job that once required hundreds of coders.
Source-stated
The article says these systems remain a pipe dream as tools and roles in software engineering change.
Down here on Earth, such systems remain a pipe dream. At the same time, software engineering is changing faster than many at the cutting edge expected.
Source and contributor context
MIT Technology Review · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Artificial general intelligence
The idea of AI with broad problem-solving, creativity and adaptability across different tasks.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: d6f3e150-d38d-45c5-8dc5-b5b828ecbd9f.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T00:48:51.998Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model.
Observation: d584a6008a99a199b71eb57dc829d7548f34f4cc89d619ebb73b029bd9bb8819. Source text hash: 251809828dedb390bf20bb658e20fa510ef2b50bc3aa9de77177b0719d0ecdac.
uncertain · 2026-11-16T00:48:51.998Z · anthropic/claude-opus-5.5
Publication: d6f3e150-d38d-45c5-8dc5-b5b828ecbd9f
attribution — pass: The reading identifies one MIT Technology Review article and attributes its views to the named people: Gottschlich at Merly, Li and Pullen at Cosine, and Kant at Poolside. The 'Deep Dive' list at the end is page furniture, so the evidence is a single piece and not a listing.
To Justin Gottschlich, CEO and founder of Merly, large language models are the wrong tool for the job—period.
summary — pass: The reading leads with the article's thesis: a second wave of tools aiming at functional correctness. It names the competing companies and methods and keeps Gottschlich's dissent. It ties AGI ambitions to the pipe-dream caveat. LLM and AGI are defined on the page. 'Intermediate representation' is not explained, but it is peripheral. The question follows from the evidence.
It’s that second level of correctness that the new wave of generative coding assistants are aiming for
date — pass: January 20, 2025 matches the UTC day of originalPublished. The 'Deep Dive' sidebar has later items, such as 'Ten years after AlphaGo', but the reading does not use them. Li's prediction is labeled a forecast, not an outcome. This is a publisher page, so no archive disclosure is needed. The later edit does not affect the body passages the reading relies on.
support — uncertain: Claim 5 credits Li with predicting that companies will reduce hiring, but that sentence is the journalist's own ('Many companies will...'). Li predicts tiers of engineers and small teams. Repair: attribute the hiring cut to the article, or describe Li's tiers and team-size forecast. Claim 6 says 'these systems remain a pipe dream', which can read as the coding tools. The source applies that phrase to AGI-style agents. Repair: say 'AGI-style systems'. Claims 4 and 5 also soften Li's 'will become' to 'may'. Restore his conviction.
But there’s another angle too. Many companies will use this technology to cut down on the number of programmers they hire. Li thinks we will soon see tiers of software engineers.
media — pass: No media is selected, and no verified candidate media is available. Leaving the media field empty is appropriate.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A language model with many learned numerical settings, used to process and generate language.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEThe idea of AI with broad problem-solving, creativity and adaptability across different tasks.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEThe idea of AI with broad problem-solving, creativity and adaptability across different tasks.
Definition reference