2025: ideas & lineages
Selekoa’s 67 ideas & lineages records dated 2025, in date order, of 772 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 67 records
January 2025
· publication · dated to the day
MiniMax posted the MiniMax-01 technical report, “MiniMax-01: Scaling Foundation Models with Lightning Attention”, to arXiv on 14 January 2025 (version 1). Evidence ↗· publication · dated to the day
MiniMax’s MiniMax-01 GitHub repository README, at a commit of 15 January 2025, introduced two models, MiniMax-Text-01 and MiniMax-VL-01. Evidence ↗· publication · dated to the day
A Chinese-language page on ByteDance Seed's website, dated 22 January 2025 and titled 'Doubao 1.5pro - Doubao Team', describes the Doubao-1.5-pro model as using a mixture-of-experts architecture. Evidence ↗· publication · dated to the day
The Kimi-k1.5 repository README, as of 22 January 2025, offered testing of Kimi k1.5 through the Kimi OpenPlatform API on application, and published no model weights. Evidence ↗· publication · dated to the day
Yuxin Wu is listed among the authors of the Kimi k1.5 technical report (arXiv version 1). Evidence ↗· publication · dated to the day
Zhilin Yang is listed among the authors of the Kimi k1.5 technical report (arXiv version 1). Evidence ↗· publication · dated to the day
Xinyu Zhou is listed among the authors of the Kimi k1.5 technical report (arXiv version 1). Evidence ↗· publication · dated to the day
Moonshot AI’s Kimi Team posted the Kimi k1.5 technical report to arXiv on 22 January 2025 (version 1). Evidence ↗· publication · dated to the day
DeepSeek published the R1 research report. Evidence ↗· publication · dated to the day
R1’s distillation section cites Qwen2.5’s announcement while naming Qwen base models fine-tuned on R1-generated samples. Evidence ↗· publication · dated to the day
The R1 report distinguishes R1-Zero’s RL without preliminary SFT from R1’s cold-start training. Evidence ↗
March 2025
· publication · dated to the month
“How Much Should We Spend to Reduce A.I.'s Existential Risk?” (2025-03) credits Charles I. Jones as its author. Evidence ↗· publication · dated to the month
Jones, “How Much Should We Spend to Reduce A.I.'s Existential Risk?” (2025-03) bears on E02. The preceding sentence reports experts on catastrophic outcomes 'either from malicious use or even from a superintelligent A.I. itself'. Evidence ↗Jones, How Much Should We Spend to Reduce A.I.'s Existential Risk (2025) →Claim page →
· publication · dated to the month
Jones, “How Much Should We Spend to Reduce A.I.'s Existential Risk?” (2025-03) articulates E03. Sentence continues 'in the Covid-19 pandemic'; Jones adds that mortality risk to people alive 'more than justifies spending large amounts'. Abstract follows a back-of-envelope calculation from a $10 million value of life and a 1% mortality risk. Preceded by the note that 12–33% of runs lead to zero mitigation investment (very small or hard-to-mitigate risk). After saying that valuing future generations would raise the amount. Evidence ↗Jones, How Much Should We Spend to Reduce A.I.'s Existential Risk (2025) →Claim page →
· publication · dated to the month
Jones, “How Much Should We Spend to Reduce A.I.'s Existential Risk?” (2025-03) articulates E04. Listed among 'many related questions that merit further study'. Evidence ↗Jones, How Much Should We Spend to Reduce A.I.'s Existential Risk (2025) →Claim page →
· publication · dated to the month
Jones, “How Much Should We Spend to Reduce A.I.'s Existential Risk?” (2025-03) articulates E05. Jones then says the Covid-19 experience supplies a key insight, letting existential risk be modelled like pandemic mortality. Evidence ↗Jones, How Much Should We Spend to Reduce A.I.'s Existential Risk (2025) →Claim page →
· publication · dated to the month
Jones, “How Much Should We Spend to Reduce A.I.'s Existential Risk?” (2025-03) bears on E05. Related-literature paragraph that also cites Aschenbrenner (2020), Aschenbrenner and Trammell (2024) and Martin and Pindyck. Evidence ↗Jones, How Much Should We Spend to Reduce A.I.'s Existential Risk (2025) →Claim page →
· announcement · dated to the day
Baidu announced ERNIE 4.5 and the reasoning model ERNIE X1 on 16 March 2025, stating that both were accessible to individual users free of charge through ERNIE Bot. Evidence ↗
April 2025
· publication · dated to the day
Axiom Space's release (April 2025) states that limitations of terrestrial data centres, including energy, cooling, real estate, licensing and permitting, are becoming increasingly apparent, and presents its orbital nodes as mitigating them. Evidence ↗· publication · dated to the day
Moonshot AI published the Kimi-VL-A3B-Instruct model on its moonshotai Hugging Face account, with a model card whose metadata gives the licence as “mit”, at a commit of 9 April 2025. Evidence ↗· publication · dated to the day
ByteDance Seed posted the Seed-Thinking-v1.5 reasoning-model report to arXiv on 10 April 2025 (version 1); later arXiv versions are titled Seed1.5-Thinking. Evidence ↗· publication · dated to the day
Moonshot AI’s Kimi Team posted the Kimi-VL technical report to arXiv on 10 April 2025 (version 1). Evidence ↗
May 2025
June 2025
· publication · dated to the day
ByteDance Seed posted the Seed-Coder technical report to arXiv on 4 June 2025 (version 1). Evidence ↗· publication · dated to the day
The RAND working paper "Contemporary AI foundation models increase biological weapons risk" (WR-A3853-1, arXiv v1 of 12 June 2025) states that contemporary AI models can meaningfully contribute to biological weapons development, against published safety assessments that assert minimal risk. Evidence ↗Contemporary AI foundation models increase biological weapons risk →Claim page →
· publication · dated to the day
The RAND working paper of 12 June 2025 states that safety assessments mistakenly assume that biological weapons development requires tacit knowledge, that is, expertise gained through experience that cannot be conveyed in words. Evidence ↗Contemporary AI foundation models increase biological weapons risk →Claim page →
· publication · dated to the day
The RAND working paper of 12 June 2025 states that flawed safety assessments resting on tacit knowledge and inadequate benchmarks may create a false sense of security, leading to weak safeguards and regulatory gaps. Evidence ↗Contemporary AI foundation models increase biological weapons risk →Claim page →
· publication · dated to the day
The RAND working paper of 12 June 2025 states that safety assessments rely on benchmarks that overlook the full range of potential threat actors and fail to account for how AI could enhance the capabilities of skilled individuals as well as enabling non-experts. Evidence ↗Contemporary AI foundation models increase biological weapons risk →Claim page →
· publication · dated to the day
MiniMax posted the MiniMax-M1 technical report to arXiv on 16 June 2025 (version 1); the abstract states that MiniMax publicly releases MiniMax-M1 in versions with 40K and 80K thinking budgets. Evidence ↗· revision · dated to the day
The PaddlePaddle/ERNIE repository README at its commit of 30 June 2025 introduces ERNIE 4.5 as a family of ten model variants and states that the models, including pre-trained weights and inference code, have been released on Hugging Face and AI Studio. Evidence ↗
July 2025
· publication · dated to the month
The Frontier AI Risk Management Framework (v1.0) states that it builds upon the International AI Safety Report (January 2025) and TC260's AI Safety Governance Framework v1.0, which it cites in a footnote. Evidence ↗· publication · dated to the day
The README of Moonshot AI’s Kimi-K2 GitHub repository, at the repository’s initial commit dated 11 July 2025, introduced two model variants, Kimi-K2-Base and Kimi-K2-Instruct, and pointed to their checkpoints on Hugging Face. Evidence ↗· publication · dated to the day
ByteDance Seed posted the Seed-X multilingual translation model report to arXiv on 18 July 2025 (version 1). Evidence ↗· publication · dated to the day
Yuxin Wu is listed among the authors of the Kimi K2 technical report (arXiv version 1). Evidence ↗· publication · dated to the day
Zhilin Yang is listed among the authors of the Kimi K2 technical report (arXiv version 1). Evidence ↗· publication · dated to the day
Xinyu Zhou is listed among the authors of the Kimi K2 technical report (arXiv version 1). Evidence ↗· publication · dated to the day
Moonshot AI’s Kimi Team posted the Kimi K2 technical report, “Kimi K2: Open Agentic Intelligence”, to arXiv on 28 July 2025 (version 1). Evidence ↗
August 2025
· publication · dated to the day
“Re: recent Anthropic safety research” (2025-08-06) credits Eliezer Yudkowsky as its author. Evidence ↗· publication · dated to the day
The post “Re: recent Anthropic safety research” (2025-08-06) contests Y21. Commenting on reported behaviour of a language model, he distinguishes a learned human-like persona from the system's underlying preferences, which he expects to be alien. Evidence ↗Yudkowsky, Re: recent Anthropic safety research (2025) →Claim page →
· announcement · dated to the day
ByteDance Seed's Seed-OSS model card states that on 20 August 2025 it released Seed-OSS-36B-Base, in versions with and without synthetic data, and Seed-OSS-36B-Instruct. Evidence ↗
September 2025
· publication · dated to the month
“A Research Agenda for the Economics of Transformative AI” (2025-09) credits Ajay K. Agrawal as one of its 3 authors. Evidence ↗· publication · dated to the month
“A Research Agenda for the Economics of Transformative AI” (2025-09) credits Anton Korinek as one of its 3 authors. Evidence ↗· publication · dated to the month
“A Research Agenda for the Economics of Transformative AI” (2025-09) credits Erik Brynjolfsson as one of its 3 authors. Evidence ↗· publication · dated to the month
Andrews and Farboodi, “Do Markets Believe in Transformative AI?” (2025-09) bears on E05. Event study of US Treasury and TIPS yields around 30 frontier-lab model releases (Nov 2022–Dec 2025). Evidence ↗Andrews and Farboodi, Do Markets Believe in Transformative AI (2025) →Claim page →
· publication · dated to the month
Brynjolfsson, Korinek and Agrawal, “A Research Agenda for the Economics of Transformative AI” (2025-09) bears on E05. Opening sentence of the research-agenda abstract ('As we approach Transformative Artificial Intelligence (TAI), …'). The agenda's list; 'safety risks' is one of nine challenges. Evidence ↗Brynjolfsson, Korinek and Agrawal, A Research Agenda for the Economics of Transformative AI (2025) →Claim page →
· publication · dated to the day
Anthropic's Frontier Red Team post "LLMs and biorisk" (5 September 2025) states that its 2024 pilot study, while not providing strong evidence either way about AI uplift, suggested that tacit knowledge may not be as significant a bottleneck as many have suggested. Evidence ↗· publication · dated to the day
Anthropic's Frontier Red Team post of 5 September 2025 states that text-based uplift trials of the kind it reports are imperfect proxies for real-world scenarios, which involve further factors it names. Evidence ↗
October 2025
· publication · dated to the month
“Genius on Demand: The Value of Transformative Artificial Intelligence” (2025-10) credits Ajay K. Agrawal as one of its 3 authors. Evidence ↗· publication · dated to the month
“Genius on Demand: The Value of Transformative Artificial Intelligence” (2025-10) credits Avi Goldfarb as one of its 3 authors. Evidence ↗· publication · dated to the month
“Genius on Demand: The Value of Transformative Artificial Intelligence” (2025-10) credits Joshua S. Gans as one of its 3 authors. Evidence ↗· publication · dated to the month
“We Won't be Missed: Work and Growth in the AGI World” (2025-10) credits Pascual Restrepo as its author. Evidence ↗· publication · dated to the month
Agrawal, Gans and Goldfarb, “Genius on Demand: The Value of Transformative Artificial Intelligence” (2025-10) bears on E02. Model of 'genius on demand' AI; human geniuses specialise where their comparative advantage over AI is greatest. Evidence ↗Agrawal, Gans and Goldfarb, Genius on Demand (2025) →Claim page →
· publication · dated to the month
Restrepo, “We Won't be Missed: Work and Growth in the AGI World” (2025-10) bears on E02. Restrepo's chapter for the NBER volume The Economics of Transformative AI; the abstract derives labour-share convergence to zero. Evidence ↗· publication · dated to the day
NVIDIA's post (October 2025) quotes Starcloud's chief executive Philip Johnston as saying that, apart from launch, orbital data centres would bring tenfold carbon-dioxide savings over their life compared with terrestrial power. Evidence ↗· announcement · dated to the day
MiniMax announced on 27 October 2025 that it was making MiniMax M2 open source and launching it, and stated that the complete model weights had been released on Hugging Face. Evidence ↗· publication · dated to the day
Aili, Choi, Ong and Wen’s Perspective on carbon-neutral data centres in space (Nature Electronics, 2025) states that retired orbital hardware ends up as space junk that could burn up in the atmosphere, releasing emissions and other harmful chemicals, a disadvantage compared with terrestrial data centres. Evidence ↗Aili, Choi, Ong and Wen, Carbon-neutral data centres in space (2025) →Claim page →
November 2025
· publication · dated to the day
Google's Suncatcher post (November 2025) states that in the right orbit a solar panel can be up to eight times more productive than on Earth and produce power nearly continuously. Evidence ↗Google Research, Project Suncatcher post (2025) →Claim page →
· publication · dated to the day
Google's Suncatcher post (November 2025) names high-bandwidth ground communications among the significant engineering challenges that remain. Evidence ↗Google Research, Project Suncatcher post (2025) →Claim page →
· publication · dated to the day
Google's Suncatcher post (November 2025) names thermal management among the significant engineering challenges that remain for space-based ML compute. Evidence ↗Google Research, Project Suncatcher post (2025) →Claim page →
· publication · dated to the day
Moonshot AI published the Kimi-K2-Thinking model on its moonshotai Hugging Face account; the model card, at a commit of 6 November 2025, describes it as a thinking model built starting from Kimi K2. Evidence ↗· publication · dated to the day
ESPI Report 98 (November 2025) states that in-orbit servicing remains so costly that deorbiting a failed satellite and launching a replacement is generally considered more cost-effective than sending crews for repairs, and that the maintenance cycle of a space-based data centre is expected to depend heavily on redundancy, which itself increases costs. Evidence ↗· publication · dated to the day
ESPI Report 98 (November 2025) models gigawatt-scale space-based data centres against a terrestrial benchmark and finds that, under its modelling assumptions, a constellation architecture has double the benchmark’s capital cost and more than triple its operating cost, and a monolithic architecture an operating cost 38% higher. Evidence ↗· publication · dated to the day
The Suncatcher paper (November 2025) argues that if launch to LEO reaches $200 per kilogram, launch cost amortised over spacecraft life could be roughly comparable to terrestrial data-centre energy costs per kW. Evidence ↗Agüera y Arcas et al., Suncatcher paper (2025) →Claim page →
· publication · dated to the day
The Suncatcher paper (November 2025) reports proton-beam tests in which Trillium TPU HBM showed irregularities after 2 krad(Si), about three times the authors' five-year requirement of 750 rad(Si). Evidence ↗Agüera y Arcas et al., Suncatcher paper (2025) →Claim page →
· publication · dated to the day
The Suncatcher paper (November 2025) states that replacing failed TPUs by technicians, simple on Earth, is impracticable in space, and that redundant provisioning is the simplest solution. Evidence ↗Agüera y Arcas et al., Suncatcher paper (2025) →Claim page →
· publication · dated to the day
Google researchers posted "Towards a future space-based, highly scalable AI infrastructure system design" to arXiv as version 1 on 22 November 2025. Evidence ↗
December 2025
· publication · dated to the day
1 provision record of MiniMax Group Inc. Global Offering prospectus (31 December 2025)