2023: ideas & lineages
Selekoa’s 114 ideas & lineages records dated 2023, in date order, of 356 in all topics.
Selective: what the records hold, not a history of the year.
How this page is made
- Each record keeps the kind of date its source gives: an event, a publication, an announcement, an effective date, a revision, a signature, a snapshot, or the start or end of an interval. A snapshot is one dated report of a state.
- A record dated only to a month sits under that month, before the records dated to a day. A record dated only to the year sits under the year, before the months. No day or month is added.
- An announcement records what a party stated. It does not show that anything was done, built or paid for.
- Dates a party states for a year are listed apart, after the records: targets, deadlines and terms, and dates stated in passing, such as a reported start of operation. So is a record’s own date that falls after its sources were last checked, such as a law’s date of effect. A stated date given only in words, with no year, is not listed.
- Records with an unknown date are not listed. A record with an interval appears in the year it starts and in the year it ends, or once, with both dates, when both fall in one year.
- Each provision of a text is a record of its own. Provisions sharing a text and a date are shown here as one entry, with their number, and listed on the text’s page; a contested or unresolved one is listed here on its own. The counts on this page include each provision record.
- The page is generated from the records, with no selection or framing by an editor. Coverage is uneven across years and topics: a year or topic with few records may be one the records barely cover.
Ideas & lineages 114 records
January 2023
· publication · dated to the day
“Alexander and Yudkowsky on AGI goals” (2023-01-24) credits Eliezer Yudkowsky as one of its 2 authors. Evidence ↗· publication · dated to the day
The post “Alexander and Yudkowsky on AGI goals” (2023-01-24) contests Y05. Answering a question about whether AIs will internalise prohibitions the way most children do, he argues that children's moral learning relies on innate human mechanisms that AI training does not reproduce. Evidence ↗Alexander and Yudkowsky, Alexander and Yudkowsky on AGI goals (2023) →Claim page →
February 2023
· publication · dated to the day
The LLaMA paper discusses Hoffmann and colleagues’ compute-optimal training analysis while distinguishing inference budgets. Evidence ↗· publication · dated to the day
Meta published the LLaMA paper, whose first arXiv version is dated 27 February 2023; Meta’s announcement of 24 February 2023 introduced the work and links the paper. Evidence ↗· publication · dated to the day
The LLaMA paper reports using the Transformer architecture of Vaswani and colleagues. Evidence ↗· publication · dated to the day
Hugo Touvron co-authored the first LLaMA paper. Evidence ↗
March 2023
· announcement · dated to the day
Baidu announced ERNIE Bot on 16 March 2023, stating that it would be accessible to invited users from that day and that its API would be available on application to enterprise clients through Baidu AI Cloud. Evidence ↗· publication · dated to the day
“GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models” (2023-03-17) credits Daniel Rock as one of its 4 authors. Evidence ↗· publication · dated to the day
“GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models” (2023-03-17) credits Pamela Mishkin as one of its 4 authors. Evidence ↗· publication · dated to the day
“GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models” (2023-03-17) credits Sam Manning as one of its 4 authors. Evidence ↗· publication · dated to the day
“GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models” (2023-03-17) credits Tyna Eloundou as one of its 4 authors. Evidence ↗· publication · dated to the day
Eloundou et al., “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models” (2023-03-17) bears on M02. The authors deliberately do not classify exposure as automation versus augmentation. Evidence ↗· publication · dated to the day
Eloundou et al., “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models” (2023-03-17) bears on M03. The authors argue LLMs meet general-purpose-technology criteria of pervasiveness, improvement and complementary innovation. The authors note that co-invention requirements typical of general-purpose technologies may slow LLM diffusion. Evidence ↗· publication · dated to the day
Eloundou et al., “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models” (2023-03-17) bears on M06b. The authors call for policy preparedness and point to prior work on education, training and safety-net reform, but decline to recommend specific policies. Evidence ↗· publication · dated to the day
Eloundou et al., “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models” (2023-03-17) articulates M07. The authors estimate the share of US workers whose occupational tasks are exposed to LLMs, from O*NET task descriptions rated by humans and GPT-4. The authors describe their method: rating task exposure using a rubric applied by human annotators and GPT-4. Evidence ↗· publication · dated to the day
Eloundou et al., “GPTs are GPTs: An Early Look at the Labor Market Impact Potential of Large Language Models” (2023-03-17) bears on M07. The authors caution that exposure measures technical feasibility, not realised productivity or automation. Evidence ↗· publication · dated to the day
“Pausing AI Developments Isn't Enough. We Need to Shut it All Down” (2023-03-29) credits Eliezer Yudkowsky as its author. Evidence ↗· publication · dated to the day
“Pausing AI Developments Isn't Enough. We Need to Shut it All Down” (2023-03-29) contests Y01. Argues that the dangerous thresholds lie beyond human-level intelligence and cannot be calculated in advance, so a lab could pass them unnoticed. Written in response to the March 2023 open letter calling for a six-month pause. Evidence ↗Yudkowsky, Pausing AI Developments Isn't Enough. We Need to Shut it All Down (2023) →Claim page →
· publication · dated to the day
“Pausing AI Developments Isn't Enough. We Need to Shut it All Down” (2023-03-29) contests Y07. In an op-ed arguing that a six-month moratorium is insufficient, he cites a leading lab's stated intention to have future AI do alignment work as evidence that there is no viable survival plan. Evidence ↗Yudkowsky, Pausing AI Developments Isn't Enough. We Need to Shut it All Down (2023) →Claim page →
· publication · dated to the day
“Pausing AI Developments Isn't Enough. We Need to Shut it All Down” (2023-03-29) contests Y14. Contrasts safety for superhuman intelligence with other challenges humanity has overcome by learning from mistakes and trying again. Written in response to the March 2023 open letter calling for a six-month pause. Evidence ↗Yudkowsky, Pausing AI Developments Isn't Enough. We Need to Shut it All Down (2023) →Claim page →
· publication · dated to the day
“Pausing AI Developments Isn't Enough. We Need to Shut it All Down” (2023-03-29) bears on Y17. Setting out what he thinks would actually need to be done, he calls for an indefinite worldwide halt to large training runs, backed by multinational agreements. Evidence ↗Yudkowsky, Pausing AI Developments Isn't Enough. We Need to Shut it All Down (2023) →Claim page →
· publication · dated to the day
“Pausing AI Developments Isn't Enough. We Need to Shut it All Down” (2023-03-29) contests Y17. Responding to an open letter calling for a six-month pause on training systems more powerful than GPT-4, he argues that the letter asks for too little. Evidence ↗Yudkowsky, Pausing AI Developments Isn't Enough. We Need to Shut it All Down (2023) →Claim page →
· publication · dated to the day
“Pausing AI Developments Isn't Enough. We Need to Shut it All Down” (2023-03-29) contests Y19. In a general-audience article he repeats the formula after saying that the likely outcome of current methods is an AI that does not care for humans. Evidence ↗Yudkowsky, Pausing AI Developments Isn't Enough. We Need to Shut it All Down (2023) →Claim page →
· publication · dated to the day
“Pausing AI Developments Isn't Enough. We Need to Shut it All Down” (2023-03-29) contests Y22. Concludes his TIME essay, which calls for an indefinite worldwide moratorium on large training runs, not a six-month pause; the same essay says anyone 'talking of arms races is a fool'. Evidence ↗Yudkowsky, Pausing AI Developments Isn't Enough. We Need to Shut it All Down (2023) →Claim page →
· publication · dated to the day
“Pausing AI Developments Isn't Enough. We Need to Shut it All Down” (2023-03-29) contests Y24. Within his proposal for a worldwide moratorium, he rejects framing AI development as a national arms race on the ground that everyone shares the outcome. Discussing a moratorium that might begin in the United States, he describes the technology as one that no country could own and that would kill people in both countries. Evidence ↗Yudkowsky, Pausing AI Developments Isn't Enough. We Need to Shut it All Down (2023) →Claim page →
· publication · dated to the day
“Pausing AI Developments Isn't Enough. We Need to Shut it All Down” (2023-03-29) contests Y26. Asks readers to picture a hostile superhuman AI as a fast alien civilisation initially confined to computers, then argues that such confinement would not last because molecular biology can be ordered online. Evidence ↗Yudkowsky, Pausing AI Developments Isn't Enough. We Need to Shut it All Down (2023) →Claim page →
· publication · dated to the day
“Pausing AI Developments Isn't Enough. We Need to Shut it All Down” (2023-03-29) contests Y27. Arguing that making superhuman AI safe could take decades, he says that failure on the first attempt leaves no one to learn from it. Evidence ↗Yudkowsky, Pausing AI Developments Isn't Enough. We Need to Shut it All Down (2023) →Claim page →
· publication · dated to the day
“Pausing AI Developments Isn't Enough. We Need to Shut it All Down” (2023-03-29) contests Y30. Opening his case against a six-month pause, he states what he expects from building superhuman AI under present conditions. Evidence ↗Yudkowsky, Pausing AI Developments Isn't Enough. We Need to Shut it All Down (2023) →Claim page →
April 2023
· publication · dated to the day
“GPTs are Predictors, not Imitators” (2023-04-08) credits Eliezer Yudkowsky as its author. Evidence ↗· publication · dated to the day
The post “GPTs are Predictors, not Imitators” (2023-04-08) contests Y09. Using the example of text listing products of primes followed by the primes, he argues that predicting text can demand more intelligence than generating it, so the task does not cap out at the level of the human writers. Concludes that predicting all Internet text, including the errors and private thoughts behind it, is a harder task than being a human, while conceding that GPT-4 is not yet as capable as a human in many ways. Evidence ↗Yudkowsky, GPTs are Predictors, not Imitators (2023) →Claim page →
· publication · dated to the day
The post “GPTs are Predictors, not Imitators” (2023-04-08) contests Y21. He argues that predicting human text requires modelling the processes behind it, a different and harder task than being human, so a predictor need not think like the humans it predicts. Evidence ↗Yudkowsky, GPTs are Predictors, not Imitators (2023) →Claim page →
June 2023
· publication · dated to the month
“The Great Inflection? A Debate About AI and Explosive Growth” (2023-06) credits Matt Clancy as one of its 2 authors. Evidence ↗· publication · dated to the month
“The Great Inflection? A Debate About AI and Explosive Growth” (2023-06) credits Tamay Besiroglu as one of its 2 authors. Evidence ↗· publication · dated to the month
Clancy and Besiroglu, “The Great Inflection? A Debate About AI and Explosive Growth” (2023-06) articulates G01. Speaker: Tamay Besiroglu. Besiroglu defines explosive growth as at least ten times the past century's annual growth rate, sustained for at least a decade, and gives a conditional probability. Evidence ↗Clancy and Besiroglu, The Great Inflection (2023) →Claim page →
· publication · dated to the month
Clancy and Besiroglu, “The Great Inflection? A Debate About AI and Explosive Growth” (2023-06) contests G01. Speaker: Matt Clancy. Clancy questions whether growth models can be extrapolated to growth far outside historical experience. Earlier in the debate he puts the chance of explosive growth at 10–20%. Clancy's answer to Besiroglu's 65%, given on the same condition of AI deployed throughout the economy. He adds that most of the probability is "clustered around the low end of explosive growth", and the same turn doubts that growth models can be extrapolated so far. Evidence ↗Clancy and Besiroglu, The Great Inflection (2023) →Claim page →
· publication · dated to the month
Clancy and Besiroglu, “The Great Inflection? A Debate About AI and Explosive Growth” (2023-06) articulates G02. Speaker: Tamay Besiroglu. Besiroglu argues that AI restores the population-ideas feedback that drove historical acceleration. Evidence ↗Clancy and Besiroglu, The Great Inflection (2023) →Claim page →
· publication · dated to the month
Clancy and Besiroglu, “The Great Inflection? A Debate About AI and Explosive Growth” (2023-06) articulates G03. Speaker: Tamay Besiroglu. Points to faster growth after the Neolithic and Industrial Revolutions, and to double-digit catch-up growth, as precedent. Evidence ↗Clancy and Besiroglu, The Great Inflection (2023) →Claim page →
· publication · dated to the month
Clancy and Besiroglu, “The Great Inflection? A Debate About AI and Explosive Growth” (2023-06) bears on G03. Speaker: Matt Clancy. Clancy questions whether past accelerations are informative about an acceleration driven by intelligence. Evidence ↗Clancy and Besiroglu, The Great Inflection (2023) →Claim page →
· publication · dated to the month
Clancy and Besiroglu (2023-06) articulates G04. Speaker: Matt Clancy. Clancy gives a verbal version of the Aghion, Jones and Jones task model applied to innovation: with every task required and automation advancing about 1% a year, growth stays steady; only faster or full automation gives explosive growth. Clancy says AI is "pretty likely to boost growth" and then lists bottlenecks that slow AGI's takeover of tasks: physical-world tasks, raw materials, training data, the time natural processes take, preferences for human performers, regulation and zero-sum competition. Clancy's closing turn: different sectors are held back by materials, time, data or regulation; sectors that automate easily become cheap and "no longer count for much of GDP". Evidence ↗Clancy and Besiroglu, The Great Inflection (2023) →Claim page →
· publication · dated to the month
Clancy and Besiroglu, “The Great Inflection? A Debate About AI and Explosive Growth” (2023-06) contests G04. Speaker: Tamay Besiroglu. Besiroglu argues that with complementary tasks, scaling digital workers on automated tasks still raises output by one or two orders of magnitude unless a large share of tasks stays unautomated. Evidence ↗Clancy and Besiroglu, The Great Inflection (2023) →Claim page →
· publication · dated to the month
Clancy and Besiroglu, “The Great Inflection? A Debate About AI and Explosive Growth” (2023-06) contests G05. Speaker: Tamay Besiroglu. Besiroglu sets three conditions a resource bottleneck must meet (vital, hard to substitute, impossible to expand by an order of magnitude); he also argues that experiments can be parallelised and that regulation faces strong incentives to be bypassed. Evidence ↗Clancy and Besiroglu, The Great Inflection (2023) →Claim page →
· publication · dated to the month
Clancy and Besiroglu, “The Great Inflection? A Debate About AI and Explosive Growth” (2023-06) articulates G06. Speaker: Matt Clancy. Clancy answers Besiroglu's point about fast growth in AI training compute. Evidence ↗Clancy and Besiroglu, The Great Inflection (2023) →Claim page →
· publication · dated to the month
Clancy and Besiroglu, “The Great Inflection? A Debate About AI and Explosive Growth” (2023-06) articulates G09. Speakers: Tamay Besiroglu; Matt Clancy. Besiroglu's estimate conditional on AI automating the vast majority of tasks. In his closing turn he gives 25–75% "conditional on AGI". He separately puts the development of such AI this century at about 80%; he does not combine the figures. Clancy's estimate on the same condition, with "most of that probability clustered around the low end of explosive growth". Evidence ↗Clancy and Besiroglu, The Great Inflection (2023) →Claim page →
· publication · dated to the day
The preprint “Can large language models democratize access to dual-use biotechnology?” (arXiv v1, 6 June 2023) states that its results suggest that large language models will make pandemic-class agents widely accessible as soon as they are credibly identified, even to people with little or no laboratory training. Evidence ↗Can large language models democratize access to dual-use biotechnology? →Claim page →
· publication · dated to the day
The preprint "Can large language models democratize access to dual-use biotechnology?" (arXiv v1, 6 June 2023) states that the agents it discusses would become widely accessible as soon as they are credibly identified, even to people with little or no laboratory training, against the premise that laboratory execution is the binding barrier. Evidence ↗Can large language models democratize access to dual-use biotechnology? →Claim page →
· publication · dated to the day
The preprint of 6 June 2023 states that promising nonproliferation measures include pre-release evaluations of language models by third parties, curation of training datasets, and verifiable screening of all DNA generated by synthesis providers. Evidence ↗Can large language models democratize access to dual-use biotechnology? →Claim page →
· publication · dated to the day
Meta published an account of I-JEPA’s representation-prediction approach. Evidence ↗· publication · dated to the day
The I-JEPA account discusses self-supervised learning from unlabeled observations. Evidence ↗· publication · dated to the day
Sandbrink's preprint on AI and biological misuse (arXiv v1, 24 June 2023) states that large language models such as GPT-4 are already able to provide dual-use information that could have enabled historical biological weapons efforts to succeed. Evidence ↗Artificial intelligence and biological misuse: Differentiating risks of language models and biological design tools →Claim page →
· publication · dated to the day
Sandbrink's preprint (arXiv v1, 24 June 2023) differentiates large language models from biological design tools, stating that language models in particular lower barriers to biological misuse while design tools expand the capabilities of sophisticated actors. Evidence ↗Artificial intelligence and biological misuse: Differentiating risks of language models and biological design tools →Claim page →
· publication · dated to the day
Sandbrink's preprint (arXiv v1, 24 June 2023) states that enhanced screening of gene synthesis is among the measures key to reducing the risks it describes. Evidence ↗Artificial intelligence and biological misuse: Differentiating risks of language models and biological design tools →Claim page →
· publication · dated to the day
Sandbrink's preprint (arXiv v1, 24 June 2023) states that biological design tools will expand the capabilities of sophisticated actors, and proposes interventions to deter biological misuse by such actors. Evidence ↗Artificial intelligence and biological misuse: Differentiating risks of language models and biological design tools →Claim page →
July 2023
· event · dated to the day
OpenAI published Introducing Superalignment, setting out its new team’s alignment research programme. Evidence ↗· publication · dated to the day
Introducing Superalignment proposes AI assistance with evaluating other AI systems as scalable oversight. Evidence ↗· publication · dated to the day
Introducing Superalignment credits Ilya Sutskever as an author. Evidence ↗· publication · dated to the day
Introducing Superalignment credits Jan Leike as an author. Evidence ↗· publication · dated to the day
Introducing Superalignment argues that current alignment techniques including RLHF will not scale to superintelligence because they rely on human supervision that the authors expect to become unreliable for much smarter systems. Evidence ↗· publication · dated to the day
ARC Evals published an account of its autonomous-task pilot. Evidence ↗· publication · dated to the day
ARC Evals’ pilot announcement points to the GPT-4 system card as reporting earlier experiments. Evidence ↗ARC Evals autonomous-task pilot announcement (2023) →Claim page →
September 2023
· publication · dated to the day
Anthropic RSP v1.0 says its commitments are designed in the spirit of the Responsible Scaling Policy framework being developed by Paul Christiano and ARC Evals. Evidence ↗· event · dated to the day
The RSP announcement explicitly credited ARC Evals’ framework as inspiring Anthropic’s approach. Evidence ↗· publication · dated to the day
Jingren Zhou co-authored the Qwen Technical Report. Evidence ↗· publication · dated to the day
Junyang Lin co-authored the Qwen Technical Report. Evidence ↗· publication · dated to the day
The Qwen report says it uses PPO in reinforcement learning from human feedback. Evidence ↗· publication · dated to the day
Qwen Team published its 2023 technical report on pretraining and alignment. Evidence ↗· publication · dated to the day
"Open-Sourcing Highly Capable Foundation Models" (arXiv v1, submitted 29 September 2023) states that for some highly capable foundation models likely to be developed in the near future, open-sourcing may pose sufficiently extreme risks to outweigh the benefits, and names access to biological weapons knowledge among the dangerous capabilities at issue. Evidence ↗Open-Sourcing Highly Capable Foundation Models: An evaluation of risks, benefits, and alternative methods for pursuing open-source objectives →Claim page →
October 2023
· publication · dated to the month
The October 2023 NBER working paper “Economic Growth under Transformative AI (October 2023 version)” (2023-10) credits Anton Korinek as one of its 2 authors. Evidence ↗· publication · dated to the month
The April 2026 revision of the NBER working paper “Economic Growth under Transformative AI” (2023-10) credits Anton Korinek as one of its 2 authors. Evidence ↗· publication · dated to the month
The October 2023 NBER working paper “Economic Growth under Transformative AI (October 2023 version)” (2023-10) credits Philip Trammell as one of its 2 authors. Evidence ↗· publication · dated to the month
The April 2026 revision of the NBER working paper “Economic Growth under Transformative AI” (2023-10) credits Philip Trammell as one of its 2 authors. Evidence ↗· publication · dated to the month
Trammell and Korinek, “Economic Growth under Transformative AI (October 2023 version)” (2023-10) articulates E01. Discussing AI safety, the authors say models in which capital recursively self-improves without human input are the most worrying for existential risk. Evidence ↗Trammell and Korinek, Economic Growth under Transformative AI (2023) →Claim page →
· publication · dated to the month
Trammell and Korinek, “Economic Growth under Transformative AI (October 2023 version)” (2023-10) bears on E01. The abstract distinguishes AI letting capital self-replicate from AI letting it self-improve via knowledge production. Evidence ↗Trammell and Korinek, Economic Growth under Transformative AI (2023) →Claim page →
· publication · dated to the month
Trammell and Korinek, “Economic Growth under Transformative AI” (2023-10) articulates E01. Abstract sentence contrasting automation of production (self-replication) with automation of R&D (self-improvement). Citing the high capital share in AI R&D (Besiroglu et al., 2024), the authors describe a feedback in which automating AI R&D speeds AI capability growth, then cite calibrated models of this feedback (Erdil et al. 2025; Lifland et al. 2025) as having 'speculative elements'. In the risk subsection, the authors contrast a scenario of acceleration from R&D automation alone with scenarios of automated physical production, each with distinct safety concerns. Evidence ↗Trammell and Korinek, Economic Growth under Transformative AI (2023, revised 2026) →Claim page →
· publication · dated to the month
Trammell and Korinek, “Economic Growth under Transformative AI” (2023-10) bears on E01. Opening paragraph: the authors say future models may reach AGI and, unlike earlier technologies, could automate R&D itself. The clause names AI creating better AI systems. The authors characterise the 'singularity' hypothesis and add that absent current economic effects cannot refute it because the theory predicts little advance warning. Second of the review's 'robust findings': automating R&D in isolation has ambiguous growth effects. Evidence ↗Trammell and Korinek, Economic Growth under Transformative AI (2023, revised 2026) →Claim page →
· publication · dated to the month
Trammell and Korinek, “Economic Growth under Transformative AI (October 2023 version)” (2023-10) articulates E02. Explaining their focus on average wages and the labour share, the authors state that the likeliest transformative long-run scenario is AI outsmarting all humans. Evidence ↗Trammell and Korinek, Economic Growth under Transformative AI (2023) →Claim page →
· publication · dated to the month
Trammell and Korinek, “Economic Growth under Transformative AI” (2023-10) bears on E02. Definition of AGI used in the review's introduction. Evidence ↗Trammell and Korinek, Economic Growth under Transformative AI (2023, revised 2026) →Claim page →
· publication · dated to the month
Trammell and Korinek, “Economic Growth under Transformative AI (October 2023 version)” (2023-10) articulates E03. The authors frame AI safety as a public good to which public-goods and climate-risk economics could apply. Evidence ↗Trammell and Korinek, Economic Growth under Transformative AI (2023) →Claim page →
· publication · dated to the month
Trammell and Korinek, “Economic Growth under Transformative AI” (2023-10) articulates E03. Concluding subsection placing AI existential risk within the paper's framing of economic transformation; it goes on to say the scenarios carry different risk profiles. Evidence ↗Trammell and Korinek, Economic Growth under Transformative AI (2023, revised 2026) →Claim page →
· publication · dated to the month
Trammell and Korinek, “Economic Growth under Transformative AI” (2023-10) bears on E04. Closing sentence of the dangerous-AI subsection. Evidence ↗Trammell and Korinek, Economic Growth under Transformative AI (2023, revised 2026) →Claim page →
· publication · dated to the month
Trammell and Korinek, “Economic Growth under Transformative AI (October 2023 version)” (2023-10) articulates E05. Closing their review, Trammell and Korinek say that existential risk from AI, central in futurist discussion, is almost absent from economics, because existing growth models offer no obvious mechanism for it. The authors attribute the profession's inattention to growth singularities to a modelling norm tied to the Kaldor facts, not to the models themselves. Footnote dating Sandberg's 2013 overview of singularity models to before mainstream economists' engagement with transformative AI. Evidence ↗Trammell and Korinek, Economic Growth under Transformative AI (2023) →Claim page →
· publication · dated to the month
Trammell and Korinek, “Economic Growth under Transformative AI (October 2023 version)” (2023-10) bears on E05. The review opens by dating economists' awareness of AI's transformative potential to Simon's prediction. Evidence ↗Trammell and Korinek, Economic Growth under Transformative AI (2023) →Claim page →
· publication · dated to the month
Trammell and Korinek, “Economic Growth under Transformative AI” (2023-10) articulates E05. Cites Jones (2024, 2025) and Shulman and Thornley (2025), and Trammell and Aschenbrenner (2026) on precaution possibly increasing long-term risk. The authors introduce their synthesis of 'the most influential papers on the topic' of transformative AI and growth. Footnote: most models Sandberg summarised did not spell out how AI would fit standard economic models; 'Our review fills this gap.'. Evidence ↗Trammell and Korinek, Economic Growth under Transformative AI (2023, revised 2026) →Claim page →
· publication · dated to the day
“Explosive growth from AI automation: A review of the arguments” (2023-10-01) credits Ege Erdil as one of its 2 authors. Evidence ↗· publication · dated to the day
“Explosive growth from AI automation: A review of the arguments” (2023-10-01) credits Tamay Besiroglu as one of its 2 authors. Evidence ↗· publication · dated to the day
Erdil and Besiroglu, “Explosive growth from AI automation: A review of the arguments” (2023-10-01) articulates G01. Opens the paper's summary of its assessment: substitution of AI for labour on most or all tasks makes explosive growth the default prediction of growth models. Evidence ↗Erdil and Besiroglu, Explosive growth from AI automation (2023) →Claim page →
· publication · dated to the day
Erdil and Besiroglu, “Explosive growth from AI automation: A review of the arguments” (2023-10-01) contests G04. Based on a CES task model calibrated with a range of substitution parameters: fast automation of most tasks can give a large transitory boost even with bottlenecks. Section 3.1 treats the objection that "the training or deployment of AI systems will be sufficiently impeded by regulation"; the authors argue that incentives for AI adoption resemble those for nuclear proliferation, so coordination to slow it is likely to be imperfect. Evidence ↗Erdil and Besiroglu, Explosive growth from AI automation (2023) →Claim page →
· publication · dated to the day
Erdil and Besiroglu, “Explosive growth from AI automation: A review of the arguments” (2023-10-01) contests G05. Accepts eventual limits and contests their timing, with an estimate of returns to scale. Evidence ↗Erdil and Besiroglu, Explosive growth from AI automation (2023) →Claim page →
· publication · dated to the day
Erdil and Besiroglu, “Explosive growth from AI automation: A review of the arguments” (2023-10-01) articulates G09. The authors' unconditional estimate. The same paragraph says high confidence "seems unwarranted". Evidence ↗Erdil and Besiroglu, Explosive growth from AI automation (2023) →Claim page →
November 2023
· publication · dated to the month
“The A.I. Dilemma: Growth versus Existential Risk” (2023-11) credits Charles I. Jones as its author. Evidence ↗· publication · dated to the month
“Can Socially-Minded Governance Control the AGI Beast?” (2023-11) credits Joshua S. Gans as its author. Evidence ↗· publication · dated to the month
Jones, “The A.I. Dilemma: Growth versus Existential Risk” (2023-11) bears on E01. Cites Aghion, Jones and Jones (2019), Trammell and Korinek (2020), Davidson (2021), Nordhaus (2021), Erdil and Besiroglu (2023). Evidence ↗· publication · dated to the month
Jones, “The A.I. Dilemma: Growth versus Existential Risk” (2023-11) articulates E02. Preceded by 'many experts worry that these advances entail existential risk:'. Jones reports warnings from 'a substantial contingent of the A.I. research community, including leading researchers at OpenAI and Google'. Evidence ↗· publication · dated to the month
Jones, “The A.I. Dilemma: Growth versus Existential Risk” (2023-11) bears on E02. Jones's own statement of possibility, followed by growth models in which machines produce ideas. Evidence ↗· publication · dated to the month
Gans, “Can Socially-Minded Governance Control the AGI Beast?” (2023-11) bears on E03. Refers to the 'potential existential risks' that motivated OpenAI's funders; the model then assumes the risks are real. Evidence ↗Gans, Can Socially-Minded Governance Control the AGI Beast (2023) →Claim page →
· publication · dated to the month
Jones, “The A.I. Dilemma: Growth versus Existential Risk” (2023-11) articulates E03. The abstract sets growth gains against existential risk, which it attributes to 'many experts', and frames the paper as a welfare analysis of whether to continue or stop. First of four insights: 'With log utility, the models are remarkably unconcerned with existential risk'. Insight 2 says CRRA utility with γ of 2 or more is 'quite conservative'. Key Point 3: with γ ≤ 1 any risk short of sure annihilation is acceptable for an immediate singularity; with γ = 2 or 3 the cutoffs are much smaller. The 'key exception' to the conservative result: mortality gains are 'in the same units' as existential risk. Evidence ↗· publication · dated to the month
Gans, “Can Socially-Minded Governance Control the AGI Beast?” (2023-11) articulates E04. Abstract: a model under idealised conditions (AGI risks real, safe AGI possible, socially-minded funders exist) motivated by OpenAI's governance. Case in which the for-profit entity is viable in direct competition; 'they' is the socially-minded entity. Conclusion; Gans adds that whether national or global intervention is needed is left to future research, while the results point to the latter. Evidence ↗Gans, Can Socially-Minded Governance Control the AGI Beast (2023) →Claim page →
· publication · dated to the month
Jones, “The A.I. Dilemma: Growth versus Existential Risk” (2023-11) bears on E05. Distinguishes the paper from Jones (2016) on growth versus lives lost; earlier the section lists Joy, Bostrom, Rees, Posner, Yudkowsky et al. and Ord as having highlighted AI existential risk. Evidence ↗· publication · dated to the day
xAI’s initial agenda proposes integrating formal verification. Evidence ↗· publication · dated to the day
xAI published an initial research programme alongside its Grok announcement. Evidence ↗· publication · dated to the day
xAI’s initial agenda discusses scalable oversight assisted by tools and human feedback. Evidence ↗· publication · dated to the day
The preprint "Towards Responsible Governance of Biological Design Tools" (arXiv v1, 27 November 2023) states that biological design tools pose distinct regulatory challenges and that regulatory proposals tailored to large language models may be less effective for them. Evidence ↗Towards Responsible Governance of Biological Design Tools →Claim page →
· publication · dated to the day
The preprint of 27 November 2023 states that the design capabilities of biological design tools could be used to evade DNA screening techniques, against the premise that screening at synthesis is a dependable control point. Evidence ↗Towards Responsible Governance of Biological Design Tools →Claim page →
· publication · dated to the day
The preprint of 27 November 2023 identifies, among the challenges of regulating biological design tools, the potential for safeguards to be circumvented by open-sourcing model weights. Evidence ↗Towards Responsible Governance of Biological Design Tools →Claim page →
· publication · dated to the day
Meta published a retrospective locating FAIR’s launch in late 2013 and describing a decade of open research. Evidence ↗
December 2023
· publication · dated to the month
CSET's explainer "AI and Biorisk" (December 2023) states that the barrier to scientific information is not high even for beginners, and that feasibility depends on both information and hands-on production success, without stating which of the two limits biological misuse. Evidence ↗· publication · dated to the month
CSET's explainer of December 2023 treats biological design tools as a separate class from chatbots, stating that their use in its current state requires both scientific and programming expertise. Evidence ↗· publication · dated to the month
CSET's explainer of December 2023 states that the biorisk landscape is not uniform and that specific scenarios and actors should be assessed individually, without stating that assessment must cover actors who already hold relevant skills. Evidence ↗· event · dated to the day
Google DeepMind published its Gemini research introduction through CEO Demis Hassabis, on behalf of the Gemini team. Evidence ↗· publication · dated to the day
OpenAI’s Weak-to-strong generalization announcement presents the linked paper as the Superalignment team’s first paper. Evidence ↗· publication · dated to the day
Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision cites Supervising strong learners by amplifying weak experts. Evidence ↗Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision →Claim page →
· publication · dated to the day
Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision cites Self-critiquing models for assisting human evaluators. Evidence ↗Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision →Claim page →
· publication · dated to the day
Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision cites AI safety via debate. Evidence ↗Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision →Claim page →
· publication · dated to the day
Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision cites Scalable agent alignment via reward modeling: a research direction. Evidence ↗Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision →Claim page →
· publication · dated to the day
Weak-to-Strong Generalization distinguishes generalizing beyond weak supervision from improving supervision through scalable oversight. Evidence ↗Weak-to-Strong Generalization: Eliciting Strong Capabilities With Weak Supervision →Claim page →