· Aug 17, 2026
Jack Clark argues in Import AI that invention could reshape who holds power
Markdown · JSON · Original source
On August 17, 2026, ’s Import AI newsletter examines AI systems’ ability to discover, conduct research and invent, and asks what those abilities could mean for human control. Clark describes DiG-bench, games with hidden rules that test whether systems can discover how unfamiliar environments work. He says struggle compared with human players, and predicts they may reach human parity by the middle of 2027, when he expects to seriously begin. He also reports that Faraday, a system trained to supervise research experiments, outperformed other models on some replication tasks under a rubric-based evaluation. Clark connects such work to the possibility that AI systems could help advance their own capabilities. The newsletter’s central challenge is to Mark Zuckerberg’s vision of widely shared : would a system capable of superhuman invention work solely for people less capable than itself? Clark does not say such a system would necessarily become malign. He argues that its effects on the balance of power could be difficult to reason about, and that Zuckerberg’s proposed balance is not a foregone outcome.
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
· Article
Source details
Original source date: Aug 17, 2026 (2026-08-17T13:05:36.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-06T05:53:23.243Z. First observed: 2026-10-23T18:48:00.743Z. Retrieved: 2026-10-23T18:49:36.230Z.
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 October 23, 2026; originally published August 17, 2026. The Faraday results are reported according to a rubric-based judge; the newsletter excerpt does not provide independent validation of those results. Clark’s human-parity timeline and expectation about are his prediction, not reported outcomes.
Open question
What would ensure that a system capable of superhuman invention serves individual empowerment rather than changing the balance of power in ways people cannot readily anticipate?
Source-stated
Jack Clark says DiG-bench tests whether AI systems can discover hidden rules through interaction, and reports that current frontier models struggle compared with humans.
The games are difficult enough that they are not beatable by today’s frontier models.
Source-stated
Jack Clark predicts that AI systems may reach human parity on DiG-bench by the middle of 2027, linking that possibility to recursive self-improvement.
My guess is we’ll reach human parity on DiG-bench by middle of 2027, at which point we should expect things like recursive self-improvement to seriously kick off.
Source-stated
Jack Clark reports that Faraday, using Codex, outperformed standard models on some research replication tasks according to a rubric-based judge.
Faraday using Codex is able to beat standard Opus 4.8 and GPT-5.5 on some replication tasks, exceeding their performance “on 73% of in-distribution ML tasks, and on 60% of held-out AI-for-science tasks, according to our rubric-based judge.”
Source-stated
Jack Clark questions whether a system capable of superhuman invention would work solely for the empowerment of people less capable than it at invention.
will a system capable of superhuman invention solely work on behalf of the individual empowerment of people that are less capable than it at invention?
Source-stated
Jack Clark argues that superhuman invention could alter the balance of power in ways that are difficult to reason about, and that Zuckerberg’s proposed balance is not a foregone outcome.
it’s hard to reconcile a system capable of superhuman invention with something that doesn’t fundamentally alter the balance of power in the world in ways that are confusing and hard to reason about.
Source and contributor context
Jack Clark · 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.
Frontier model
One of the most capable AI models available at a given time; the set shifts as models improve.
Recursive self-improvement
An AI system improving its own capabilities, such as its code, training methods or reasoning, then using the gains to improve further. Each cycle could hypothetically produce a more capable system.
Superintelligence
A hypothetical AI far more capable than people across most intellectual tasks.
Jack Clark
Jack Clark writes Import AI, a newsletter covering AI research, development and its implications.
Editorial provenance
Version 1. Publication ID: eeb46b12-28a0-72c3-1c14-8420d1056a52.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-06T05:49:54.437Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 85566b4b02da0ef98c18aa8c71736e903e8850ae801bd4aa8c9b813687695262. Source text hash: 1a6f976341c05c21c32fd6f93cc4975067ec0be38ca2a4ef894670ad0d3ebe6f.
Revision and source-check history
Version 1 · 2026-11-06T05:53:23.243Z
Jack Clark argues in Import AI that invention could reshape who holds power
On August 17, 2026, ’s Import AI newsletter examines AI systems’ ability to discover, conduct research and invent, and asks what those abilities could mean for human control. Clark describes DiG-bench, games with hidden rules that test whether systems can discover how unfamiliar environments work. He says struggle compared with human players, and predicts they may reach human parity by the middle of 2027, when he expects to seriously begin. He also reports that Faraday, a system trained to supervise research experiments, outperformed other models on some replication tasks under a rubric-based evaluation. Clark connects such work to the possibility that AI systems could help advance their own capabilities. The newsletter’s central challenge is to Mark Zuckerberg’s vision of widely shared : would a system capable of superhuman invention work solely for people less capable than itself? Clark does not say such a system would necessarily become malign. He argues that its effects on the balance of power could be difficult to reason about, and that Zuckerberg’s proposed balance is not a foregone outcome.
Original source
· Article
Source details
Original source date: Aug 17, 2026 (2026-08-17T13:05:36.000Z).
Date basis: Date supplied by the publisher’s feed.
Added to Escape Meridian: 2026-11-06T05:53:23.243Z. First observed: 2026-10-23T18:48:00.743Z. Retrieved: 2026-10-23T18:49:36.230Z.
Limits
Text as retrieved on October 23, 2026; originally published August 17, 2026. The Faraday results are reported according to a rubric-based judge; the newsletter excerpt does not provide independent validation of those results. Clark’s human-parity timeline and expectation about are his prediction, not reported outcomes.
Open question
What would ensure that a system capable of superhuman invention serves individual empowerment rather than changing the balance of power in ways people cannot readily anticipate?
Source-stated
Jack Clark says DiG-bench tests whether AI systems can discover hidden rules through interaction, and reports that current frontier models struggle compared with humans.
The games are difficult enough that they are not beatable by today’s frontier models.
Source-stated
Jack Clark predicts that AI systems may reach human parity on DiG-bench by the middle of 2027, linking that possibility to recursive self-improvement.
My guess is we’ll reach human parity on DiG-bench by middle of 2027, at which point we should expect things like recursive self-improvement to seriously kick off.
Source-stated
Jack Clark reports that Faraday, using Codex, outperformed standard models on some research replication tasks according to a rubric-based judge.
Faraday using Codex is able to beat standard Opus 4.8 and GPT-5.5 on some replication tasks, exceeding their performance “on 73% of in-distribution ML tasks, and on 60% of held-out AI-for-science tasks, according to our rubric-based judge.”
Source-stated
Jack Clark questions whether a system capable of superhuman invention would work solely for the empowerment of people less capable than it at invention.
will a system capable of superhuman invention solely work on behalf of the individual empowerment of people that are less capable than it at invention?
Source-stated
Jack Clark argues that superhuman invention could alter the balance of power in ways that are difficult to reason about, and that Zuckerberg’s proposed balance is not a foregone outcome.
it’s hard to reconcile a system capable of superhuman invention with something that doesn’t fundamentally alter the balance of power in the world in ways that are confusing and hard to reason about.
Source and contributor context
Jack Clark · 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.
Frontier model
One of the most capable AI models available at a given time; the set shifts as models improve.
Recursive self-improvement
An AI system improving its own capabilities, such as its code, training methods or reasoning, then using the gains to improve further. Each cycle could hypothetically produce a more capable system.
Superintelligence
A hypothetical AI far more capable than people across most intellectual tasks.
Jack Clark
Jack Clark writes Import AI, a newsletter covering AI research, development and its implications.
Editorial provenance
Version 1. Publication ID: eeb46b12-28a0-72c3-1c14-8420d1056a52.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-06T05:49:54.437Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 85566b4b02da0ef98c18aa8c71736e903e8850ae801bd4aa8c9b813687695262. Source text hash: 1a6f976341c05c21c32fd6f93cc4975067ec0be38ca2a4ef894670ad0d3ebe6f.
passed · 2026-11-06T05:49:54.437Z · anthropic/claude-opus-5.5
Publication: eeb46b12-28a0-72c3-1c14-8420d1056a52
attribution — pass: The reading attributes the newsletter, its predictions and its critique of Zuckerberg to Jack Clark's Import AI. The Faraday results are framed as reported by Clark, and Zuckerberg's vision is kept separate from Clark's critique. This is one newsletter issue with commentary in the author's own words, not a listing.
Welcome to Import AI, a newsletter about AI research.
summary — pass: The summary leads with Clark's central argument and explains DiG-bench as games with hidden rules. It describes Faraday as a supervising research system and preserves Clark's nuanced position: he does not claim malign AI, but questions a foregone balance of power. Frontier model, recursive self-improvement and superintelligence are defined beside the reading. The title and question follow from the source.
I am not suggesting that superhuman invention guarantees some kind of malign entity that is independent from people.
date — pass: August 17, 2026 matches the UTC calendar day of originalPublished. The kind is publisher-page, so no archive disclosure is needed. The 2027 parity timeline is labeled as Clark's prediction, and there is no hindsight framing. The page was modified after publication, but every cited passage is core newsletter body text, plainly part of the original issue.
My guess is we’ll reach human parity on DiG-bench by middle of 2027
support — pass: Each claim matches a verbatim source passage. The Faraday result is qualified as rubric-judged and limited to some tasks. The claim that frontier models struggle compared with humans is stated in the source. Clark's question and his not-a-foregone-outcome argument are quoted closely. The reading's question extends Clark's own missing question without adding unsupported assumptions.
This is certainly one potential outcome but I struggle to see how it is the foregone outcome.
media — pass: No media was selected, and no verified media candidates were supplied, so omitting media 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.
Jack Clark writes Import AI, a newsletter covering AI research, development and its implications.
Explore Jack ClarkJack Clark writes Import AI, a newsletter covering AI research, development and its implications.
Explore Jack ClarkOne of the most capable AI models available at a given time; the set shifts as models improve.
Definition referenceAn AI system improving its own capabilities, such as its code, training methods or reasoning, then using the gains to improve further. Each cycle could hypothetically produce a more capable system.
Definition referenceA hypothetical AI far more capable than people across most intellectual tasks.
Definition referenceJack Clark writes Import AI, a newsletter covering AI research, development and its implications.
Explore Jack ClarkAn AI system improving its own capabilities, such as its code, training methods or reasoning, then using the gains to improve further. Each cycle could hypothetically produce a more capable system.
Definition referenceJack Clark writes Import AI, a newsletter covering AI research, development and its implications.
Explore Jack ClarkOne of the most capable AI models available at a given time; the set shifts as models improve.
Definition referenceAn AI system improving its own capabilities, such as its code, training methods or reasoning, then using the gains to improve further. Each cycle could hypothetically produce a more capable system.
Definition referenceA hypothetical AI far more capable than people across most intellectual tasks.
Definition referenceJack Clark writes Import AI, a newsletter covering AI research, development and its implications.
Explore Jack ClarkAn AI system improving its own capabilities, such as its code, training methods or reasoning, then using the gains to improve further. Each cycle could hypothetically produce a more capable system.
Definition reference