Cognition claimed Devin resolved 13.86% of SWE-bench issues end-to-end
Cognition's March 2024 Devin launch post stated that Devin resolved 13.86% of SWE-bench issues end-to-end, up from a prior state of the art of 1.96%.
Facts
Atomic, verifiable facts - every one tied to a primary source.
278 entries, all primary-sourced
Common Crawl has collected petabytes of web data since 2008
Common Crawl describes its openly available web corpus as petabytes of data that it has regularly collected since 2008.
DARPA's Strategic Computing program cost about a billion dollars
DARPA's Strategic Computing program spent about a billion dollars from 1983 to 1993 chasing military machine intelligence.
ETS uses an NLP engine to help score standardized-test essays
ETS's e-rater engine uses natural language processing to grade essays alongside human raters on the GRE and TOEFL.
Every AIME answer is an integer from 0 to 999
The AIME is a 15-question, 3-hour exam in which every answer is an integer between 0 and 999, making it trivial to score automatically.
F.E.A.R.'s acclaimed combat AI ran on a three-state machine
F.E.A.R.'s acclaimed combat AI ran on a finite state machine with only three states.
Ford took a $2.7 billion impairment shutting down Argo AI
Ford booked a $2.7 billion impairment when it and Volkswagen wound down their self-driving venture Argo AI in 2022.
Forza's Drivatars learn real players' driving styles
Forza's Drivatars learn real players' driving styles in the cloud and race as AI stand-ins for them.
Frank Rosenblatt drowned on his 43rd birthday
Perceptron inventor Frank Rosenblatt drowned in 1971 in a Chesapeake Bay boating accident on his 43rd birthday, his ideas decades ahead of the hardware.
Frege's 1879 Begriffsschrift introduced the quantifiers every logician still uses
Gottlob Frege's 1879 Begriffsschrift introduced quantifiers - the 'for all' and 'there exists' notation - founding the predicate logic still used today.
Gemini Ultra was the first model Google said beat human experts on MMLU
At launch, Google said Gemini Ultra was the first model to outperform human experts on the MMLU benchmark of multitask knowledge.
Gender Shades found error rates up to 34.7% for darker-skinned women versus 0.8% for lighter-skinned men
The 2018 Gender Shades study reported gender-classification error rates up to 34.7% for darker-skinned females versus 0.8% for lighter-skinned males.
GGUF is the single-file format that packages models for local inference
GGUF stores a model and all its metadata in one self-contained, memory-mappable file, the standard format for running LLMs locally with llama.cpp.
Google open-sourced TensorFlow under Apache 2.0 in November 2015
On November 9, 2015, Google released TensorFlow as free open-source software under the Apache 2.0 license, announced by Jeff Dean and Rajat Monga.
Google paused Gemini image generation of people in February 2024
After Gemini produced historically inaccurate images, Google said in February 2024 it had turned off image generation of people while it worked on improvements.
Google ran TPUs in its data centers before telling anyone
When Google revealed the TPU in May 2016, it had already run the custom AI chips in its data centers since 2015, powering Search and AlphaGo.
GPT-4o can respond to audio in as little as 232 milliseconds
GPT-4o can respond to audio inputs in as little as 232 milliseconds, averaging 320 milliseconds - close to human conversational response time.
GSM8K contains about 8,500 grade-school math word problems
OpenAI's GSM8K benchmark is a dataset of roughly 8,500 high-quality, linguistically diverse grade-school math word problems.
Herbert Simon won both the Turing Award and the Nobel Prize
Herbert Simon received the 1975 ACM Turing Award in computing and the 1978 Nobel Memorial Prize in Economic Sciences.
HiFi-GAN generated speech audio about 168x faster than real time
The 2020 HiFi-GAN vocoder produced 22kHz audio 167.9 times faster than real time on a single GPU at near-human quality.
In The Sims, objects advertise to tell Sims what to do
In The Sims, smart objects advertise how much they will satisfy a Sim's needs, and Sims pick the best.
InstructGPT's human feedback came from about 40 contractors
OpenAI trained InstructGPT using a team of about 40 contractors hired on Upwork and through Scale AI to write demonstrations and rank outputs.
Italy fined Replika's maker five million euros over data protection
In 2025 Italy's Garante fined Luka Inc., maker of the Replika companion chatbot, five million euros over its data processing.
Japan's Fifth Generation project ran about a decade and ended in 1992
The Fifth Generation Computer Systems project, launched by Japan in 1982, concluded around 1992, capped by ICOT's final international conference that year.
Jevons's Logic Piano was the first machine to solve a logic problem faster than a human
William Stanley Jevons's keyboard-operated Logic Piano, completed around 1869, was the first device to solve a logic problem faster than a person without it.
Karpathy's deep dive into LLMs runs three and a half hours
Andrej Karpathy's 2025 explainer 'Deep Dive into LLMs like ChatGPT' runs about three and a half hours, covering the training stack from raw text to assistant.
Kathleen Booth is credited with writing the first assembly language
Kathleen Booth wrote the first assembly language at Birkbeck College, the readable notation that freed programmers from coding in raw binary.
LAION-5B contains 5.85 billion image-text pairs
The LAION-5B dataset comprises 5.85 billion CLIP-filtered image-text pairs, of which 2.32 billion are in English.
Libratus beat poker pros by 1.7 million in chips
Over 120,000 hands of heads-up no-limit poker, Libratus finished ahead of four top professionals by a collective 1,766,250 dollars in chips.
Lovelace wrote that the Analytical Engine cannot originate anything
In Note G of her 1843 Notes, Ada Lovelace wrote that the Analytical Engine has no pretensions whatever to originate anything.
Meta paused the Galactica public demo within three days
Meta's Galactica science model went public on November 15, 2022 and its demo was paused by November 17, 2022.
Meta's DeepFake Detection Challenge dataset held over 100,000 clips from 3,426 actors
The DFDC dataset released by Meta (then Facebook) AI contained over 100,000 face-swap video clips sourced from 3,426 paid actors who agreed to participate.
Meta's Llama license restricts companies above 700 million monthly users
The Llama community license requires companies with over 700 million monthly active users to obtain a separate license from Meta, so it is not OSI-approved.
Microsoft pulled the Tay chatbot offline within a day of launch
Microsoft took its Tay Twitter chatbot offline less than a day after launching it, after users provoked it into offensive posts.
Microsoft's Immersive Reader uses AI to aid dyslexic readers
Microsoft's Immersive Reader reads text aloud, splits words into syllables, and translates, to help readers with dyslexia.
Model Cards were introduced in a 2019 paper led by Margaret Mitchell
The Model Card, now a standard way to document an AI model's behavior across groups, was proposed in a 2019 paper led by Margaret Mitchell.
Moore's Law started as a 1965 magazine article
What became known as Moore's Law began as a four-page article Gordon Moore wrote for Electronics magazine on April 19, 1965, not a law of physics.
Moravec estimated the human brain at 100 million MIPS
In 1998 Hans Moravec estimated human-level computation at about 100 million MIPS and predicted cheap machines would match it before 2030.
Newell and Simon claimed symbols are necessary and sufficient for intelligence
The physical symbol system hypothesis states that a physical symbol system has the necessary and sufficient means for general intelligent action.
Noam Shazeer co-authored both the Transformer paper and the sparse mixture-of-experts paper
Noam Shazeer is the second author of 'Attention Is All You Need' (the Transformer) and first author of the sparsely-gated mixture-of-experts paper.
Numerai hides which stocks its data describes so modelers cannot cheat
Numerai obfuscates its financial dataset so contributors cannot tell which stocks the features describe, preventing trading and overfitting.
NumPy is called the foundation of the entire scientific Python universe
The 2020 Nature paper by NumPy's core developers describes the library as 'the foundation upon which the entire scientific Python universe is constructed.'
OpenAI Five was first AI to beat esports world champions
On April 13, 2019, OpenAI Five defeated Dota 2 world champions OG, becoming the first AI system to beat the reigning world champions at an esport.
OpenAI's Jukebox was trained on 1.2 million songs
OpenAI curated a dataset of 1.2 million songs, paired with lyrics and metadata, to train its raw-audio music model Jukebox.
Over one million model checkpoints are usable through the Transformers library
Hugging Face's documentation states there are over 1M Transformers model checkpoints on the Hugging Face Hub, each loadable by name with a single install.
PyTorch was released publicly in January 2017
The PyTorch team's own retrospective fixes the framework's public debut to January 2017, one year before their January 19, 2018 'a year in' post.
reCAPTCHA transcribed old books at over 99 percent word accuracy
reCAPTCHA used everyday CAPTCHA solutions to transcribe scanned books and newspapers with word accuracy exceeding 99 percent.
Refik Anadol's Unsupervised was trained on 138,151 MoMA artworks
Refik Anadol trained the AI behind MoMA's Unsupervised on 138,151 pieces from the museum's collection using StyleGAN2.