Meta's Cicero finished in the top ten percent of human Diplomacy players
On webDiplomacy.net, Meta's Cicero scored more than double the human average and finished in the top ten percent of players who played more than one game.
Facts
Atomic, verifiable facts - every one tied to a primary source.
268 entries, all primary-sourced
NeurIPS 2023 made authors responsible for any LLM-written text
NeurIPS 2023 let authors use LLMs to help write papers but held them responsible for all text being correct and original.
Clarkesworld shut submissions in a single morning over AI spam
Clarkesworld closed submissions on Feb 20, 2023 after banned-behavior submissions spiked past fifty before noon, mostly machine-written.
ChatGLM-6B could run on a 6 GB consumer GPU
With INT4 quantization, Tsinghua and Zhipu AI's ChatGLM-6B ran in about 6 GB of GPU memory, fitting on an ordinary consumer graphics card.
OpenAI never disclosed GPT-4's size or architecture
The 2023 GPT-4 Technical Report declined to reveal the model's architecture, size, hardware, or training compute, citing competition and safety.
GPT-4 scored about 297 of 400 on the Uniform Bar Exam
A 2023 study estimated GPT-4 scored roughly 297 of 400 on the Uniform Bar Exam, above the passing threshold in every UBE jurisdiction.
The Pause letter asked for a six-month halt above GPT-4
The 2023 Future of Life Institute open letter asked AI labs to pause for at least six months the training of systems more powerful than GPT-4.
BloombergGPT was trained on about 708 billion tokens, roughly half of them financial
BloombergGPT trained on ~708 billion tokens: 363 billion from financial sources plus 345 billion general.
NASA's DAGGER AI gives a 30-minute solar storm warning
NASA's deep-learning model DAGGER can predict geomagnetic storm impacts worldwide about 30 minutes before they hit.
Turnitin's AI detector reached 10,700 institutions at launch
At its April 2023 launch, Turnitin's AI writing detector was switched on for over 10,700 institutions and 2.1 million educators.
Med-PaLM 2 scored 86.5% on USMLE-style medical questions
Google's Med-PaLM 2 reached up to 86.5% on the MedQA benchmark of USMLE-style questions, improving on Med-PaLM by over 19 percentage points.
Meta's MMS halved Whisper's error rate while covering 11x more languages
On shared languages, Meta's MMS speech models hit half OpenAI Whisper's word error rate while covering 11 times as many languages.
Voyager remembered Minecraft skills by saving them as executable code
Voyager's agent stored each learned Minecraft skill as a reusable program, collecting 3.3x more unique items than prior methods.
The Statement on AI Risk is 22 words long
The Center for AI Safety's 2023 Statement on AI Risk is a single sentence of exactly 22 words.
AlphaDev's sorting code went into the C++ standard library
Sorting routines discovered by DeepMind's AlphaDev were added to the LLVM libc++ standard library, run trillions of times a day.
Insilico took an AI drug from target to Phase I in under 30 months
Insilico Medicine says its generative-AI drug INS018_055 went from novel target discovery to a Phase I trial in under 30 months, about half the usual time.
Pangu-Weather makes a 24-hour global forecast in about 1.4 seconds
Huawei's Pangu-Weather produces a global 24-hour forecast in roughly 1.4 seconds on one GPU, about 10,000 times faster than numerical weather models.
On WebArena, a GPT-4 agent finished 14% of web tasks versus 78% for humans
WebArena reported a GPT-4 agent completing 14.41% of realistic web tasks against 78.24% for humans, a stark capability gap.
China enacted the world's first national law specifically governing generative AI
China's Interim Measures for Generative AI Services, effective 15 August 2023, were the first national regulation aimed specifically at generative AI services.
CS50 bans ChatGPT but allows its own AI duck
Harvard's CS50 course policy forbids ChatGPT yet permits cs50.ai, its own AI tutor built to guide students instead of answering.
Amazon capped Kindle self-publishing at three new titles a day
To curb a flood of AI-generated e-books, Amazon limited most KDP accounts to publishing three new titles per day in September 2023.
The Nvidia H20 was the export-compliant chip Nvidia built for China after the 2023 controls
After US export controls tightened in October 2023, Nvidia's China data-center chip became the H20, later needing a license and a 4.5 billion dollar charge.
28 countries plus the EU signed the Bletchley Declaration
The Bletchley Declaration was signed at the November 2023 AI Safety Summit by 28 countries together with the European Union.
GraphCast beat the leading forecast system on 90% of targets
DeepMind's GraphCast beat the leading conventional weather system on 90 percent of 1,380 targets, producing 10-day forecasts in under a minute.
Sam Altman was fired and back within days
OpenAI's board removed Sam Altman as CEO on November 17, 2023, and OpenAI announced his return on November 22, 2023 - a span of less than a week.
GNoME predicted 2.2 million new crystal structures
DeepMind's GNoME tool predicted 2.2 million new crystal structures, of which about 380,000 were flagged as stable candidates for synthesis.
Mixtral 8x7B uses only about 13B of its 47B parameters per token
Mixtral 8x7B holds about 47 billion parameters but activates only around 13 billion for each token it processes.
FunSearch made the biggest cap-set gain in 20 years
DeepMind's FunSearch found new cap-set constructions, the largest improvement to that problem in twenty years, using an LLM.
Rite Aid ran facial recognition on shoppers for about eight years
An FTC complaint says Rite Aid used facial recognition in stores from about October 2012 to July 2020, often flagging shoppers by mistake.
Backdoored 'sleeper agent' models survived safety training
Anthropic's 2024 Sleeper Agents study found that standard safety training failed to remove a deceptive backdoor from a language model.
Klarna said its AI assistant did the work of 700 agents in one month
Klarna reported in 2024 that its OpenAI-powered assistant handled 2.3 million chats in a month, equal to 700 full-time agents.
US Census data showed firm AI use rose from 3.7% to 5.4% in 2023-2024
Census Bureau survey data showed the share of US businesses using AI rose from 3.7% in September 2023 to 5.4% by February 2024.
Spain ordered Worldcoin to stop scanning irises in 2024
In March 2024 Spain's data regulator ordered Worldcoin to stop collecting iris data, citing minors, consent, and sensitive biometrics.
NVIDIA's Blackwell GPU packs 208 billion transistors
NVIDIA's Blackwell B200, announced in March 2024, holds 208 billion transistors and adds 4-bit FP4 inference.
Coaches preferred TacticAI's corners 90% of the time
Liverpool FC experts preferred DeepMind's TacticAI corner-kick suggestions 90% of the time over real setups.
TacticAI used pitch symmetry to multiply its data
TacticAI exploited the football pitch's symmetry, reflecting each corner four ways to learn from scarce data.
An algorithm found 27,500 asteroids in old telescope images
In 2024 the Asteroid Institute's THOR algorithm found 27,500 new asteroid candidates hidden in archival telescope images.
Visa's generative-AI fraud score rates a transaction in 20 milliseconds
Visa's VAAI Score uses generative AI to flag card enumeration attacks in real time, scoring each transaction in about 20 milliseconds.
Microsoft's emissions rose 29 percent on data center construction
Microsoft's 2024 sustainability report disclosed its total emissions were up 29.1 percent from its 2020 baseline, driven by data center construction.
Nvidia's data-center revenue rose 427 percent year over year in early 2024
For the quarter ended April 28, 2024, Nvidia reported Data Center revenue of $22.6 billion, up 427 percent year over year, within record total of $26.0 billion.
Google's emissions rose 13 percent in 2023, driven by data centers
Google's 2024 environmental report disclosed its total greenhouse-gas emissions rose 13 percent in 2023, driven mainly by data center energy use.
AlphaProof scored silver-medal level at the 2024 IMO
DeepMind's AlphaProof and AlphaGeometry 2 solved 4 of 6 problems at the 2024 IMO, scoring 28 of 42 points - the level of a silver medalist.
The EU AI Act entered into force on August 1, 2024
Regulation (EU) 2024/1689, the EU AI Act, was published in the Official Journal on 12 July 2024 and entered into force on 1 August 2024.
The Council of Europe opened the first legally binding international AI treaty in 2024
On 5 September 2024 the Council of Europe opened its Framework Convention on Artificial Intelligence, the first-ever international legally binding treaty on AI.
John Deere's See and Spray averaged 59% herbicide savings in 2024
John Deere said See and Spray saved about 8 million gallons of herbicide mix across 1 million acres in 2024, averaging 59% savings.
David Baker won half the 2024 Chemistry Nobel for protein design
The 2024 Chemistry Nobel gave half to David Baker for computational protein design, half to Hassabis and Jumper for prediction.
A study collected 768 papers with undeclared AI writing
A 2024 study gathered 768 published papers showing telltale signs of undeclared AI use, even in high-impact, high-fee journals.
Amazon invested $8 billion in Anthropic
Amazon's total investment in Anthropic reached $8 billion after a 2023 commitment of up to $4 billion and a second $4 billion in November 2024.