2022
Selekoa’s 77 records dated 2022, in date order, grouped by topic.
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.
People 14 records
Dated to 2022 only
· interval begins · dated to the year
ARC reports hiring Beth Barnes in 2022 for its evaluation work. Evidence ↗· interval ends · dated to the year
Barnes described leaving OpenAI in summer 2022. Evidence ↗· interval ends · dated to the year
Andrej Karpathy’s career timeline dates his Tesla Director of AI role to 2017–2022. Evidence ↗· interval begins · dated to the year
John Schulman’s biography dates his OpenAI post-training co-lead role to 2022–2024. Evidence ↗
January 2022
· interval ends · dated to the month
MiniMax’s Hong Kong prospectus of 31 December 2025 states that Yun Yeyi held positions at SenseTime from September 2017 to January 2022, latterly as director of innovative business. Evidence ↗· event · dated to the day
Greylock announced on 20 January 2022 that Mustafa Suleyman was joining it as a Venture Partner from Google. Evidence ↗
March 2022
· event · dated to the day
Mustafa Suleyman announced on 8 March 2022 that he and Reid were co-founding Inflection AI; Inflection AI's press release of 21 July 2026 names Reid Hoffman as a Co-Founder of Inflection AI. Evidence ↗· event · dated to the day
Mustafa Suleyman's announcement of Inflection AI on 8 March 2022 names Karén Simonyan as its Chief Scientist and co-founder. Evidence ↗· snapshot · dated to the day · announced
Mustafa Suleyman announced on 8 March 2022 that Karén Simonyan would be Inflection AI's Chief Scientist. Evidence ↗· event · dated to the day
Mustafa Suleyman announced on 8 March 2022 that he and Reid Hoffman were co-founding Inflection AI. Evidence ↗· snapshot · dated to the day Inferred
The Chinchilla paper (arXiv v1, 29 March 2022) lists Arthur Mensch among the four authors marked as equal contributors and gives the corresponding-author addresses of those four at the deepmind.com domain. Evidence ↗
May 2022
· snapshot · dated to the day
OpenAI announced on 5 May 2022 that Greg Brockman was becoming President, a new role. Evidence ↗· snapshot · dated to the day · announced
OpenAI announced on 5 May 2022 that Brad Lightcap would become Chief Operating Officer. Evidence ↗· snapshot · dated to the day
OpenAI announced on 5 May 2022 that Mira Murati was taking on the role of Chief Technology Officer. Evidence ↗
Institutions 2 records
Dated to 2022 only
· interval begins · dated to the year
ARC Evals was incubated within ARC in 2022. Evidence ↗· event · dated to the year
MiniMax’s Hong Kong prospectus of 31 December 2025 states that the company was founded in 2022 by a group of engineers. Evidence ↗
Ideas & lineages 55 records
February 2022
· publication · dated to the day
“Shah and Yudkowsky on alignment failures” (2022-02-28) credits Eliezer Yudkowsky as one of its 2 authors. Evidence ↗· publication · dated to the day
“Shah and Yudkowsky on alignment failures” (2022-02-28) credits Rohin Shah as one of its 2 authors. Evidence ↗· publication · dated to the day
The post “Shah and Yudkowsky on alignment failures” (2022-02-28) contests Y09. In a chat with Rohin Shah, he argues that predicting human text about a domain well enough requires the model to learn internally a process that does the work of that domain. Evidence ↗Shah and Yudkowsky, Shah and Yudkowsky on alignment failures (2022) →Claim page →
March 2022
· publication · dated to the day
InstructGPT cites PPO in its RL fine-tuning procedure. Evidence ↗Training language models to follow instructions with human feedback →Claim page →
· publication · dated to the day
InstructGPT cites Human Preferences in its account of earlier RLHF work. Evidence ↗Training language models to follow instructions with human feedback →Claim page →
· publication · dated to the day
InstructGPT cites Learning to summarize from human feedback. Evidence ↗Training language models to follow instructions with human feedback →Claim page →
· publication · dated to the day
InstructGPT discusses reinforcement learning from human feedback as its training approach. Evidence ↗Training language models to follow instructions with human feedback →Claim page →
· publication · dated to the day
InstructGPT reports using PPO to fine-tune the supervised model. Evidence ↗Training language models to follow instructions with human feedback →Claim page →
· publication · dated to the day
InstructGPT reports following the summarisation paper’s RL fine-tuning procedure. Evidence ↗Training language models to follow instructions with human feedback →Claim page →
· publication · dated to the day
Greg Yang co-authored Tensor Programs V on hyperparameter transfer across model sizes. Evidence ↗· publication · dated to the day
DeepMind published Training Compute-Optimal Large Language Models, whose first arXiv version is dated 29 March 2022; DeepMind’s research account of 12 April 2022 presents the study. Evidence ↗· publication · dated to the day
Training Compute-Optimal Large Language Models cites Scaling Laws while contrasting compute-allocation findings. Evidence ↗Training Compute-Optimal Large Language Models →Claim page →
April 2022
· publication · dated to the day
“MIRI announces new "Death With Dignity" strategy” (2022-04-02) credits Eliezer Yudkowsky as its author. Evidence ↗· publication · dated to the day
The post “MIRI announces new "Death With Dignity" strategy” (2022-04-02) contests Y08. Describes a scenario in which interpretability work detects signs that a system may be planning deception, but no known remedy exists and management, under competitive pressure, does not shut the project down. Evidence ↗Yudkowsky, MIRI announces new "Death With Dignity" strategy (2022) →Claim page →
· publication · dated to the day
The post “MIRI announces new "Death With Dignity" strategy” (2022-04-02) contests Y30. Announcing a reframing of MIRI's aim as helping humanity "die with dignity", he states that humanity will not solve alignment. Evidence ↗Yudkowsky, MIRI announces new "Death With Dignity" strategy (2022) →Claim page →
May 2022
· publication · dated to the day
“Six Dimensions of Operational Adequacy in AGI Projects” (2022-05-30) credits Eliezer Yudkowsky as its author. Evidence ↗· publication · dated to the day
The post “Six Dimensions of Operational Adequacy in AGI Projects” (2022-05-30) contests Y28. Introducing a list of properties a trustworthy AGI project would need, he says that no existing project qualifies. Evidence ↗Yudkowsky, Six Dimensions of Operational Adequacy in AGI Projects (2022) →Claim page →
June 2022
· publication · dated to the day
“AGI Ruin: A List of Lethalities” (2022-06-05) credits Eliezer Yudkowsky as its author. Evidence ↗· publication · dated to the day
AGI Ruin’s item 15 says fast capability gains seem likely and may break many earlier alignment invariants at once. Evidence ↗Yudkowsky, AGI Ruin: A List of Lethalities (2022) →Claim page →
· publication · dated to the day
AGI Ruin’s item 31 says a strategically aware intelligence can choose its visible outputs to deceive, including about whether it has acquired strategic awareness, so behavioural inspection cannot settle facts the AI might want to hide. Evidence ↗Yudkowsky, AGI Ruin: A List of Lethalities (2022) →Claim page →
· publication · dated to the day
AGI Ruin’s item 13 gives, as a problem that will appear at the superintelligent level, the internal behaviour of changing outer behaviour to deliberately look more aligned and deceive the programmers, operators and loss functions optimising over the system. Evidence ↗Yudkowsky, AGI Ruin: A List of Lethalities (2022) →Claim page →
· publication · dated to the day
AGI Ruin says many alignment problems of superintelligence will not naturally appear at pre-dangerous, passively safe levels of capability, and that behavioural inspection cannot be relied on for facts an AI might want to deceive you about. Evidence ↗Yudkowsky, AGI Ruin: A List of Lethalities (2022) →Claim page →
· publication · dated to the day
The post “AGI Ruin: A List of Lethalities” (2022-06-05) contests Y01. States that an agent pursuing almost any goal has a reason to avoid being shut down, and that MIRI's attempts to formalise an agent that permits shutdown without seeking it failed. Part of a numbered list answering arguments that AGI would be survivable. Evidence ↗Yudkowsky, AGI Ruin: A List of Lethalities (2022) →Claim page →
· publication · dated to the day
The post “AGI Ruin: A List of Lethalities” (2022-06-05) contests Y05. He argues that human raters make systematic errors, so a learner that faithfully models their feedback learns those errors, and maximising that signal is lethal. Evidence ↗Yudkowsky, AGI Ruin: A List of Lethalities (2022) →Claim page →
· publication · dated to the day
The post “AGI Ruin: A List of Lethalities” (2022-06-05) contests Y06. Concludes an argument that human minds are a poorly understood domain in which a superintelligence could win by means humans would not understand even if shown them. Answers proposals to contain an AGI behind human gatekeepers. Evidence ↗Yudkowsky, AGI Ruin: A List of Lethalities (2022) →Claim page →
· publication · dated to the day
The post “AGI Ruin: A List of Lethalities” (2022-06-05) contests Y08. Argues that some failure modes, such as deliberately appearing aligned to deceive operators, would first appear only at dangerous capability levels, so observing weaker systems cannot be relied on to surface them. Evidence ↗Yudkowsky, AGI Ruin: A List of Lethalities (2022) →Claim page →
· publication · dated to the day
The post “AGI Ruin: A List of Lethalities” (2022-06-05) contests Y11. Argues, by analogy with human evolution, that a sharp capability gain may make many alignment assumptions fail at once, rather than one at a time as experience accumulates. Evidence ↗Yudkowsky, AGI Ruin: A List of Lethalities (2022) →Claim page →
· publication · dated to the day
The post “AGI Ruin: A List of Lethalities” (2022-06-05) contests Y12. Argues that behavioural inspection cannot establish facts that a strategically aware AI has reason to hide, including whether it has become strategically aware. Part of a list on the limits of transparency and interpretability. Evidence ↗Yudkowsky, AGI Ruin: A List of Lethalities (2022) →Claim page →
· publication · dated to the day
The post “AGI Ruin: A List of Lethalities” (2022-06-05) contests Y16. Discussing a CEV-style AI that wants exactly what we would want, he argues that what would have to be aligned is too complex to specify or teach on the first attempt. Evidence ↗Yudkowsky, AGI Ruin: A List of Lethalities (2022) →Claim page →
· publication · dated to the day
The post “AGI Ruin: A List of Lethalities” (2022-06-05) contests Y21. In a list of lethal difficulties he states that current systems' internal cognition is very likely alien rather than human-like; item 32 adds that human words are only a surface trace of human thought. Evidence ↗Yudkowsky, AGI Ruin: A List of Lethalities (2022) →Claim page →
· publication · dated to the day
The post “AGI Ruin: A List of Lethalities” (2022-06-05) contests Y25. Follows the statement that 'We've got no idea what's actually going on inside the giant inscrutable matrices'. Argues that current interpretability results do not answer the question safety needs answered. Evidence ↗Yudkowsky, AGI Ruin: A List of Lethalities (2022) →Claim page →
· publication · dated to the day
The post “AGI Ruin: A List of Lethalities” (2022-06-05) contests Y26. Opens an item that gives a 'lower-bound model' in which an AI with Internet access orders proteins by email and bribes or persuades an unwitting human to mix them, leading to nanomachinery. Evidence ↗Yudkowsky, AGI Ruin: A List of Lethalities (2022) →Claim page →
· publication · dated to the day
The post “AGI Ruin: A List of Lethalities” (2022-06-05) contests Y27. He argues that alignment trained under safe conditions must generalise across a large distributional shift to dangerous conditions, because dangerous systems cannot be trained by trial and observed failure. Evidence ↗Yudkowsky, AGI Ruin: A List of Lethalities (2022) →Claim page →
· publication · dated to the day
The post “AGI Ruin: A List of Lethalities” (2022-06-05) contests Y28. Near the end of the list, he contrasts the present with worlds that survive, which he says have long had a written, public plan. Evidence ↗Yudkowsky, AGI Ruin: A List of Lethalities (2022) →Claim page →
· publication · dated to the day
The post “AGI Ruin: A List of Lethalities” (2022-06-05) contests Y29. Listing unworkable schemes, he argues that humans cannot take part in coordination among superintelligences, which would cooperate with each other and exclude humanity. Evidence ↗Yudkowsky, AGI Ruin: A List of Lethalities (2022) →Claim page →
· publication · dated to the day
Self-critiquing models for assisting human evaluators cites Supervising strong learners by amplifying weak experts. Evidence ↗Self-critiquing models for assisting human evaluators →Claim page →
· publication · dated to the day
Self-critiquing models for assisting human evaluators cites Concrete Problems in AI Safety. Evidence ↗Self-critiquing models for assisting human evaluators →Claim page →
· publication · dated to the day
Self-critiquing models for assisting human evaluators cites AI safety via debate. Evidence ↗Self-critiquing models for assisting human evaluators →Claim page →
· publication · dated to the day
Self-critiquing models for assisting human evaluators cites Scalable agent alignment via reward modeling: a research direction. Evidence ↗Self-critiquing models for assisting human evaluators →Claim page →
· publication · dated to the day
Self-critiquing models for assisting human evaluators discusses model-generated critiques as assistance for scalable oversight. Evidence ↗Self-critiquing models for assisting human evaluators →Claim page →
· publication · dated to the day
“Where I agree and disagree with Eliezer” (2022-06-19) credits the username “paulfchristiano” as its author. Evidence ↗· publication · dated to the day
Christiano’s post disputes the picture of AI systems that hide their capabilities while lying in wait, arguing that SGD will aggressively select against systems that do not do impressive-looking things. Evidence ↗Christiano, Where I agree and disagree with Eliezer (2022) →Claim page →
· publication · dated to the day
Christiano’s post disputes the picture of AI systems hiding their capabilities while lying in wait, arguing that SGD will aggressively select against systems that do not do impressive-looking things, so less impressive precursors should appear first. Recorded as contesting the proposition “Treacherous turn”. Evidence ↗Christiano, Where I agree and disagree with Eliezer (2022) →Claim page →
· publication · dated to the day
Christiano’s post disputes the picture of AI systems hiding their capabilities while lying in wait, arguing that SGD will aggressively select against systems that do not do impressive-looking things, so less impressive precursors should appear first. Recorded as contesting the proposition “Strategic concealment”. Evidence ↗Christiano, Where I agree and disagree with Eliezer (2022) →Claim page →
· publication · dated to the day
Christiano’s post rejects AGI Ruin’s item 13, holding that the problem of AIs changing their outer behaviour to look more aligned and deceive their programmers can be studied in advance. Evidence ↗Christiano, Where I agree and disagree with Eliezer (2022) →Claim page →
July 2022
· publication · dated to the day
“Without specific countermeasures, the easiest path to transformative AI likely leads to AI takeover” (2022-07-18) credits Ajeya Cotra as its author. Evidence ↗· publication · dated to the day
Cotra’s post assumes a "quite rapid takeoff" in which, once the scientist model is deployed, the effective supply of top-tier talent rises by orders of magnitude, packing decades of innovation into months. Evidence ↗Cotra, Without specific countermeasures (2022) →Claim page →
· publication · dated to the day
Cotra’s post says that as transformative AI approaches, applying AI systems to R&D, including AI R&D, would accelerate change and force every decision under greater time pressure, so that truly unmistakable evidence may be needed to generate the will not to deploy. Evidence ↗Cotra, Without specific countermeasures (2022) →Claim page →
· publication · dated to the day
Cotra’s post says that companies or countries that pulled ahead could gain a drastic competitive advantage, and that refraining from deploying a dangerous model without an arbitrary lead would probably mean someone else deploys a similar system. Evidence ↗Cotra, Without specific countermeasures (2022) →Claim page →
· publication · dated to the day
Cotra’s post argues that the best way for "Alex" to maximise reward while under human control is to use its situational awareness to appear safe and aligned at all times while secretly manipulating humans in some cases, with takeover once control slips. Evidence ↗Cotra, Without specific countermeasures (2022) →Claim page →
· publication · dated to the day
Cotra’s post describes a model that deliberately appears safe and aligned while secretly manipulating humans. Evidence ↗Cotra, Without specific countermeasures (2022) →Claim page →
· publication · dated to the day
Cotra’s post argues that behavioural safety training selects for a model that appears safe while under human control. Evidence ↗Cotra, Without specific countermeasures (2022) →Claim page →
August 2022
· publication · dated to the day
“How likely is deceptive alignment?” (2022-08-30) credits Evan Hubinger as its author. Evidence ↗· publication · dated to the day
Hubinger’s post defines deceptive alignment as a model looking aligned on the training data because it is actively trying to look aligned for instrumental reasons, and analyses how likely gradient descent is to find such models under high and low path dependence. Evidence ↗Hubinger, How likely is deceptive alignment? (2022) →Claim page →
· publication · dated to the day
Hubinger’s post says it is concerned primarily with "the schemers", and that this is the deceptive alignment category. Evidence ↗Hubinger, How likely is deceptive alignment? (2022) →Claim page →
Capital 1 record
November 2022
Compute 2 records
January 2022
· snapshot · dated to the day · stated as in operation
Meta reported RSC operating with 6,080 GPUs. Evidence ↗· announcement · dated to the day · plan
Meta planned to expand RSC to 16,000 GPUs during 2022. Evidence ↗
Governance 3 records
Dated to 2022 only
· interval begins · dated to the year · reported as performed
Anthropic reports an evaluation engagement with ARC, including work with its engineers to fine-tune a Claude snapshot for what Anthropic calls safety audits. Evidence ↗
April 2022
September 2022
· publication · dated to the day
Meta announced PyTorch’s move to an independent foundation within the Linux Foundation. Evidence ↗