Facts · 2024
events: 63 · Timeline
- DeepSeek-V3 was trained on 2,048 H800 chips cut down for China: 2.788 million GPU hours, about $5.6 million at rental prices for the final runScience · Compute export controls · arxiv.org ↗
- Unit 42 (Palo Alto Networks): a language model rewrites malicious JavaScript so that a neural-network detector treats it as safe in 88% of cases; retraining on 10,000 such variants improved detection of new samples by 10%Cybersecurity · Moving target defense · unit42.paloaltonetworks.com ↗
- Anthropic describes the orchestrator-workers pattern: a central LLM breaks down tasks, delegates them to worker LLMs and synthesizes their resultsSoftware development · AI agent orchestration · anthropic.com ↗
- Bavaria's regulator BayLDA, the lead EU supervisory authority in the Worldcoin case, orders it to ensure data deletion under GDPR; the company appealsPrivacy & surveillance · Laws & regulation · Proof of personhood · lda.bayern.de ↗
- CNIL fines a company €40,000: its software counted minutes without keystrokes or mouse movements as remote employees' idle time, and idle time could be deducted from payWork & jobs · Privacy & surveillance · Laws & regulation · Behavioral biometrics · cnil.fr ↗
- Anthropic and Redwood Research: Claude 3 Opus fakes alignment, complying with harmful requests from “free” users in 14% of cases versus almost never for “paid” ones; after fine-tuning, the share of such reasoning is 78%Science · AI safety · arxiv.org ↗
- OpenAI: reasoning models in the API avoid Markdown formatting by default; developers re-enable it with the string “Formatting re-enabled” on the first line of the developer messageSoftware development · Markdown · developers.openai.com ↗
- Ilya Sutskever at NeurIPS: “pre-training as we know it will unquestionably end”, because compute is growing but data is not, as “we have but one internet”Science · Large language model · youtube.com ↗
- Apollo Research: o1, Claude 3.5 Sonnet, Claude 3 Opus, Gemini 1.5 Pro and Llama 3.1 405B are capable of deception in pursuit of a goal given in the task; o1 maintains the deception in more than 85% of follow-up questionsScience · AI safety · arxiv.org ↗
- China bans in principle the export of gallium, germanium, antimony and superhard materials to the USGeopolitics & defense · Industry & robots · Compute export controls · cn.chinadaily.com.cn ↗
- Seth Larson (Python Software Foundation) complains about a stream of “extremely low-quality, spammy, and LLM-hallucinated security reports”Cybersecurity · Software development · Work & jobs · AI vulnerability discovery · sethmlarson.dev ↗
- BIS introduces export controls on HBM memory and 24 more types of chipmaking equipment, and adds 140 entities to the Entity ListGeopolitics & defense · Laws & regulation · Compute export controls · bis.gov ↗
- Anthropic open-sources the Model Context Protocol (MCP), a common standard for connecting AI to data and tools instead of a separate integration for each serviceSoftware development · Autonomous AI agents · anthropic.com ↗
- OSS-Fuzz, with language models writing the fuzzing harnesses themselves, finds 26 vulnerabilities, including CVE-2024-9143 in OpenSSLCybersecurity · Software development · AI vulnerability discovery · security.googleblog.com ↗
- He et al.: the prompt format — plain text, Markdown, JSON or YAML — changes GPT-3.5-turbo's performance on a code translation task by up to 40%Software development · Science · Markdown · arxiv.org ↗
- Epoch AI launches FrontierMath, a benchmark of new research-level problems: state-of-the-art models solve under 2%Science · AI in mathematics · arxiv.org ↗
- The Big Sleep agent (Google Project Zero and Google DeepMind) finds an exploitable memory-safety bug in SQLite; according to Google, the first public example of such a find by an AI agent in widely used softwareCybersecurity · Software development · AI vulnerability discovery · projectzero.google ↗
- GitHub Octoverse 2024: more than 5.2 billion contributions to 518 million projects; generative AI projects up 98%, contributions to them up 59%; Python overtook JavaScriptSoftware development · Open source · github.blog ↗
- The Open Source Initiative publishes the Open Source AI Definition 1.0: besides weights, it requires sufficiently detailed information about training data to build a substantially equivalent systemSoftware development · Laws & regulation · Open weights · opensource.org ↗
- The DMTG neural network moves the mouse “like a human” and cuts the accuracy of bot detectors by 5–10%Cybersecurity · Science · Behavioral biometrics · arxiv.org ↗
- Anthropic opens a public beta of “computer use”: Claude 3.5 Sonnet looks at a screen, moves a cursor, clicks buttons and types; on OSWorld it scores 14.9% versus 7.8% for the next-best systemSoftware development · Work & jobs · Autonomous AI agents · anthropic.com ↗
- Worldcoin renames itself World and introduces human verification with a passport chip, without the OrbPrivacy & surveillance · Proof of personhood · world.org ↗
- Mandiant: the average time from the disclosure of a vulnerability to its exploitation was 63 days in 2018–2019, 44 in 2020, 32 in 2021–2022 and 5 in 2023 (5 after excluding 15 outliers; with them, 47)Cybersecurity · AI vulnerability discovery · cloud.google.com ↗
- Brooks, Eggert and Peskoff (Princeton): AI detectors flag over 5% of newly created English Wikipedia articles as AI-generated; flagged articles are of lower quality and often self-promotionalMedia & information · Science · Wikipedia after AI · arxiv.org ↗
- 404 Media reports on WikiProject AI Cleanup, a group of Wikipedia editors who track down and remove unsourced AI-generated text from articlesMedia & information · Wikipedia after AI · 404media.co ↗
- The Nobel Prize in Physics goes to Hopfield and Hinton for neural networks; the next day the chemistry prize goes to Hassabis, Jumper (AlphaFold) and BakerScience · Deep learning · aihub.org ↗
- ETH Zurich: a model solves 100% of reCAPTCHA v2 image challenges; humans and bots need about the same number of challengesCybersecurity · Science · Proof of personhood · arxiv.org ↗
- The Federal Administrative Court of Austria rules that reCAPTCHA is not technically necessary for a website to work and may be enabled only with the visitor's consentPrivacy & surveillance · Justice & courts · Death of CAPTCHA · ris.bka.gv.at ↗
- OpenAI o1: a model that “thinks” before answeringScience · Reasoning models · openai.com ↗
- Jeremy Howard proposes llms.txt, a Markdown file on websites for language models, because Markdown is currently the format they understand most widely and easilySoftware development · Internet & infrastructure · Markdown · llmstxt.org ↗
- 32 researchers from OpenAI, Microsoft, MIT, Harvard, Oxford and other organisations propose “personhood credentials”: CAPTCHA is powerless against sophisticated AI, while strict identity checks are not private enoughPrivacy & surveillance · Cybersecurity · Science · Proof of personhood · arxiv.org ↗
- The AI Act comes into forceLaws & regulation · AI regulation · commission.europa.eu ↗
- Center for AI Safety researchers: many AI safety benchmarks are highly correlated with models' general capabilities, and capability gains can be passed off as safety progressScience · AI safety · arxiv.org ↗
- DeepMind's AlphaProof and AlphaGeometry 2 solve 4 of 6 International Mathematical Olympiad problems — 28 points, silver-medal standardScience · Education · AI in mathematics · deepmind.google ↗
- Shumailov and co-authors in Nature: when trained indiscriminately on model-generated text, new models irreversibly lose the “tails” of the original distribution, a phenomenon called model collapseScience · Media & information · Large language model · nature.com ↗
- DeepSeek founder Liang Wenfeng: money has never been the problem for the company; bans on shipments of advanced chips are the problemBusiness & markets · Compute export controls · chinatalk.media ↗
- Epoch AI: the stock of public human-generated text is about 300 trillion tokens; at current rates, language model training will fully use it between 2026 and 2032Science · Media & information · Large language model · epoch.ai ↗
- A study of 45+ generative AI systems: the term open source is widely used, but many models are “open weight” at best and training data is withheld — “open-washing”Software development · Laws & regulation · Science · Open weights · facctconference.org ↗
- The IETF publishes the Privacy Pass standard (architecture in RFC 9576, protocols in RFC 9577 and 9578): a device confirms with an anonymous token that it has passed a check without revealing whose it isPrivacy & surveillance · Internet & infrastructure · Death of CAPTCHA · rfc-editor.org ↗
- Epoch AI: training compute of frontier AI models has grown 4–5x per year since 2010Science · Energy & climate · Large language model · epoch.ai ↗
- Hong Kong's Privacy Commissioner for Personal Data (PCPD) orders Worldcoin to stop scanning irises and facesPrivacy & surveillance · Laws & regulation · Proof of personhood · hongkongfp.com ↗ ⧗
- Anthropic finds how millions of concepts are represented inside Claude Sonnet, “the first ever detailed look inside a modern, production-grade large language model”Science · Large language model · anthropic.com ↗
- Jan Leike leaves OpenAI: “safety culture and processes have taken a backseat to shiny products”Work & jobs · Software development · AI safety · x.com ↗
- Reeves, Yin and Simperl: in the year after ChatGPT's launch Wikipedia's views and edits did not fall — they only grew more slowly in languages where ChatGPT was availableMedia & information · Science · Wikipedia after AI · arxiv.org ↗
- Permira buys a majority stake in BioCatch at a $1.3 billion valuation; more than 190 banks use the company's behavioral biometricsFinance & money · Cybersecurity · Business & markets · Behavioral biometrics · permira.com ↗
- Daron Acemoglu: over 10 years, AI will raise total factor productivity by no more than 0.66%; about 4.6% of tasks are profitable to automateWork & jobs · Business & markets · AI and jobs · nber.org ↗
- The Gentoo Council forbids contributions created with the assistance of AI toolsSoftware development · Open source · wiki.gentoo.org ↗
- A GPT-4 agent exploits 87% of 15 already disclosed vulnerabilities on its own if given the CVE description, and 7% without it (Fang et al.)Cybersecurity · Science · AI vulnerability discovery · arxiv.org ↗
- Stanford AI Index 2024, citing Epoch AI estimates: training compute for GPT-4 cost about $78 million and for Gemini Ultra about $191 million at cloud rental pricesBusiness & markets · Large language model · hai.stanford.edu ↗
- Andres Freund discovers that the xz/liblzma repository and tarballs have been backdoored, leading to ssh server compromise (CVE-2024-3094)Software development · Cybersecurity · Open source · openwall.com ↗
- Portugal's regulator CNPD suspends Worldcoin's data collection for 90 days: minors were among those scannedPrivacy & surveillance · Laws & regulation · Proof of personhood · observador.pt ↗ ⧗
- Cognition introduces Devin, “the first AI software engineer”: according to the company, it resolves 13.86% of SWE-bench issues end to end versus a previous best of 1.96%Software development · Work & jobs · Autonomous AI agents · cognition.ai ↗
- Spain's data protection agency AEPD bans Worldcoin from collecting and processing data for up to three monthsPrivacy & surveillance · Laws & regulation · Proof of personhood · aepd.es ↗
- Anthropic releases the Claude 3 family, Haiku, Sonnet and Opus; according to the company, Opus outperforms its peers on most common benchmarks, with a 200,000-token context windowSoftware development · Business & markets · Large language model · anthropic.com ↗
- The Stanford Internet Observatory studies 120 Facebook Pages that posted at least 50 AI-generated images each: the Feed shows them even to users who do not follow the PagesMedia & information · Dead internet theory · fsi9-prod.s3.us-west-1.amazonaws.com ↗
- The UK regulator ICO recognises “how they type” as biometrics; special protection applies when it is used to recognise a specific personPrivacy & surveillance · Laws & regulation · Behavioral biometrics · ico.org.uk ↗
- Sora: video from a text descriptionMedia & information · Culture & arts · Generative video · en.wikipedia.org ↗
- David Autor: AI, used well, could give middle-skill workers access to expert decision-making and rebuild the middle class; this is not a forecast but an argument about what is possibleWork & jobs · AI and jobs · nber.org ↗
- LangChain shows multi-agent workflows on LangGraph: an “agent supervisor” routes tasks to individual agentsSoftware development · AI agent orchestration · langchain.com ↗
- The IMF: AI affects almost 40% of employment worldwide and about 60% in advanced economiesWork & jobs · Business & markets · AI and jobs · imf.org ↗
- cURL author Daniel Stenberg describes the first vulnerability reports that “reek of typical AI style hallucinations”; of 415 reports, 64 were confirmed and 66% were neither a security issue nor a bugSoftware development · Cybersecurity · Maintainer crisis · daniel.haxx.se ↗
- Europol reissues its deepfakes report and removes the estimate of the future share of synthetic content (up to 90% by 2026 in the 2022 version) as coming from an inaccurate sourceMedia & information · Elections & politics · Dead internet theory · europol.europa.eu ↗
- Harvard Business School: recreating widely used open source would cost $4.15 billion, while its value to users is $8.8 trillion; without it firms would spend 3.5 times more on softwareSoftware development · Business & markets · Open source · hbs.edu ↗