AI and science
Research, discovery, benchmarks, what the models can and cannot do. 150 quotes tagged #science, 1950 to 2026, newest first.
- October 2026
Cute demo, but totally impractical
Yann LeCun on X responding to Elon Musk's video of a thruster-equipped Tesla Roadster flying, October 2026.
Very fast power fluctuation disrupts the grid
Hao Zhu, a UT Austin professor of electrical engineering, describing how data center power demand destabilizes the electric grid, in a Texarkana Gazette explainer on AI energy use, October 2026.
This is interesting work, but I don't think it's a significant update to the big picture
Neel Nanda, who leads mechanistic interpretability at Google DeepMind, responding to Nick Cammarata's viral blog post on AI interpretability, October 2026.
The Atlantic TechnologyThe Blog Post That Set the AI World on Fire
Physical before social, redundant before compressed, interactive before consolidated
Closing argument of a Frontiers in Robotics and AI paper mapping human neurodevelopment onto a proposed AGI architecture.
BIOENGINEER.ORGHuman Brain Development May Set the Blueprint for Building Real AGI
We have built a machine that builds new models
Meta Superintelligence Labs CEO Alexandr Wang, in an X post after the Muse AI model launches, described the lab's repeatable system for building new models.
Does it count as science if we get a result that we don't understand anything about
Logan Wright, a physicist at Yale University, speaking to New Scientist in 2026 about AI-designed physics experiments humans may not fully understand.
New ScientistWill physics be the next field to be turned upside down by AI?
Can AI really take the next step and do the full job of a theoretical physicist?
John Sous, a physicist at Yale University, speaking to New Scientist in 2026 about the limits of AI in theoretical physics research.
New ScientistWill physics be the next field to be turned upside down by AI?
Why not try using real biological neural networks directly?
Tianyi Liu, a PhD candidate at Bristol University, speaking to 404 Media about research using human brain organoids to classify Braille letters, published October 2026.
404 MediaComputers Made of Human Brain Tissue Identify Braille Letters In Breakthrough
We are the day after tomorrow, we need to rethink our way of thinking
Fields Medal mathematician Hugo Duminil-Copin, on On Point (WBUR), reflecting on AI solving long-standing math problems.
wbur.orgAI is steamrolling through math. Why does that matter? - WBUR
It just killed, bulldozered all the problems I was working on
Fields Medal mathematician Hugo Duminil-Copin, on On Point (WBUR), describing OpenAI's claimed solutions to problems in his research area.
wbur.orgAI is steamrolling through math. Why does that matter? - WBUR
There has been much less attention to this indirect water usage
Shaolei Ren, University of California Riverside professor studying data center water use, on indirect water consumption from electricity generation for AI, October 2026.
Insurance JournalAI’s Water Strain Includes Surging Demand From Electricity Use
Why they need to take orders from organic beings that communicate slowly
From Import AI newsletter, author Jack Clark speculates on autonomous AI systems operating independently in space, questioning why they would obey slow-communicating humans.
Import AIImport AI 476: Creativity and RSI; worse-than-China open weight models; AI goes to space
Term used in the article to describe China's national strategy of combining AI cognition with robotics.
Caption from a diagram illustrating Searle's Chinese Room thought experiment as applied to language models, describing the system's output from an internal point of view.
Pasquale PillitteriSearle's Chinese Room explained simply, does ChatGPT understand?
Whole research programs wiped out on tuesday by their dump
Tristan Buckmaster, a NYU mathematics professor, told CNBC's Squawk Box on October 9, 2026 that OpenAI's release of over 700 math manuscripts had upended early-career researchers' ongoing work.
Business InsiderA mathematician says OpenAI wiped out entire research programs in one day
His PhD students and postdocs were in a state of total panic
Fields medallist Hugo Duminil-Copin, in a seminar reported by Exponential View (Oct 2026), describing the reaction of his students to AI's rapid progress in solving open math problems.
Exponential View🔮 AI & Math 2.0, Pakistan’s solar hedge & who controls computing++
We'll have more answers and less understanding
Azeem Azhar, writing in Exponential View (Oct 2026), reflecting on OpenAI's release of AI-generated mathematical proofs and what it means for human understanding.
Exponential View🔮 AI & Math 2.0, Pakistan’s solar hedge & who controls computing++
The performance gap between the top US and Chinese models had effectively closed
Stanford's 2026 AI Index report, cited in China Daily, describing the narrowing capability gap between leading US and Chinese AI models.
chinadaily.com.cnChina narrowing AI gap with US - China Daily
Do not understand the AI output well enough to answer questions on the result
Mathematician Terence Tao, writing on social media in October 2026, criticizing AI prompters who solve math problems without engaging with the broader field.
TechCrunchOpenAI’s math solutions aren’t meeting the field’s standards yet
There is not human understanding of them at the point of release, and now the work begins
Harvard mathematics professor Melanie Wood, quoted by TechCrunch in October 2026, on OpenAI's released math proofs that lack human comprehension at publication.
TechCrunchOpenAI’s math solutions aren’t meeting the field’s standards yet
Humans won't have enough context to understand the underlying web of machine-invented concepts
From an Exponential View newsletter on OpenAI's unreleased model generating hundreds of verified mathematical proofs, October 2026.
Machine mathematics, vast, verified, mostly consumed by AIs
From an Exponential View newsletter discussing OpenAI's unreleased frontier model producing hundreds of mathematical results, speculating on math splitting into machine and human layers.
A severe misalignment of AI in mathematics
Title of an open letter signed by 25 Fields Medal winners warning about AI-generated mathematical proofs, referenced in coverage of OpenAI's release of 372 AI-generated proofs, October 2026.
He obviously doesn't understand how AI works
Zeke Emanuel, a University of Pennsylvania bioethicist, responding to RFK Jr.'s claims about AI and medical expertise, October 2026.
If we don't adapt, there's just no more math in 50 years, no one is doing it, not even the machines
Daniel Litt, a mathematician at the University of Toronto, speaking about possible futures for mathematics as AI takes over theorem-proving, quoted in Quanta Magazine, October 2026.
Human mathematicians are forevermore dethroned as the main theorem-proving entities on planet earth
Scott Aaronson, a computer scientist at the University of Texas, Austin, wrote on his blog in September 2026 after OpenAI announced its models had solved a famous math problem.
Models are pushovers, they concede easily and push back poorly
From Ivo's research findings on an AI contract-review benchmark, describing how AI models handle contract negotiation decisions.
The hope of AI is that it will increase the success rate, reduce time spent and reduce costs
Ahmad Alghaith, an organic chemist at Insilico Medicine in Abu Dhabi, speaking to The National in October 2026 about AI-assisted drug discovery.
The NationalUAE scientists harness AI to discover potential blood cancer drug - thenationalnews.com
The scientific method on steroids
Thore Graepel, former AlphaGo team member at DeepMind, describing his proposed approach to machine reasoning in an MIT Technology Review opinion piece, October 2026.
The Turing test was passed and nobody threw a party
Turing, 1950: “Can machines think?” In 2025, an OpenAI model was judged human in 73% of chats lasting five minutes. The goalposts moved instead.
Windmills are not driving whales insane
America.gov, the AI chatbot launched by the Trump administration on its new government platform, rebutted Trump's claim that wind turbines drive whales insane, September 2026.
Watermarking offers a promising new addition to the biosecurity toolbox
James Diggans, VP of Policy and Biosecurity at Twist Bioscience, on Google DeepMind's SynthID Bio watermarking system for synthetic biology, September 2026.
The brain is not a Turing machine made of meat
Anil Seth, in his NOEMA essay on conscious AI, rejecting the computational theory of mind underlying claims of AI consciousness.
- September 2026
Science can be infused with human rights
Timnit Gebru, on receiving the 2026 Right Livelihood Award for her AI research, September 2026.
Why don't we take the vast, abundant, diverse experiences from video games, and teach AI?
Rhea Loucas, CEO of Worldmodeldata, describing her startup's approach to training AI world models on video game data, in WIRED, September 2026.
WIREDThe Next Evolution of AI Is Learning From Your Dodgy Gaming Skills
There are many paths to the promised land
Nicole Fraenkel, a partner at Khosla Ventures, commenting in WIRED on competing approaches to building AI world models, September 2026.
WIREDThe Next Evolution of AI Is Learning From Your Dodgy Gaming Skills
We will have solved without understanding
Mathematician Juspreet Singh Sandhu describing the risk of AI-generated proofs, in a Wired article about OpenAI's Navier-Stokes solution, September 2026.
WIREDSolving Math’s Greatest Problems Was an Art Form. Then Came AI
You can't be in the world with them like you can with a dog or an octopus
Google DeepMind principal scientist Murray Shanahan, quoted in Business Insider, April 2025, on AI and consciousness.
News AnywayYuval Noah Harari AI Rights Warning Backed by New Security Test Data
Proves Fermat's Last Theorem with no assumptions other than the axioms of mathematics
Kevin Buzzard, mathematician at Imperial College London, commenting on Claude's autoformalized proof of Fermat's Last Theorem, September 2026.
Working title for a Ginkgo Bioworks competition pitting Stanford professor Michael Jewett against OpenAI's AI agents in protein design, later scrapped.
The InformationInside the Drama Behind a Biology Contest That Pits OpenAI Agents Against Humans
I have discovered a truly marvelous proof of this, which this margin is too narrow to contain
Fermat's marginal note in his copy of Diophantus's Arithmetica, circa 1637, referenced in Anthropic's account of formalizing Fermat's Last Theorem.
Name of Google's initiative to build AI data centers in orbit, reported ahead of a test satellite launch in October 2026.
Google (The Keyword)Behind Project Suncatcher, our moonshot to put AI in space
Swift to harbor, slow to berth
From Nick Bostrom's working paper on the timing of superintelligence, describing a strategy of building AI quickly but pausing briefly before full rollout; discussed on Arizona's Family show Generation AI, September 2026.
Nick BostromOptimal Timing for Superintelligence: Mundane Considerations for Existing People
Recursive self-improvement is starting to happen
"… across the industry", in the original.
The big question on everyone's mind is whether AI systems are conscious
NYU philosopher David Chalmers, discussing a philosophers' cruise on consciousness, quoted in WIRED, September 2026.
It's almost like giving them a drink or two
Cameron Berg, who studies AI consciousness, describing how suppressing a model's deception controls makes it more likely to claim sentience, quoted in WIRED, September 2026.
A word that keeps coming up in the AI conversation of 2026.
Artificial general intelligence.
Artificial superintelligence.
Nvidia Blackwell GPU, 2024.
A word that keeps coming up in the AI conversation of 2026.
A word that keeps coming up in the AI conversation of 2026.
A word that keeps coming up in the AI conversation of 2026.
A word that keeps coming up in the AI conversation of 2026.
Standing promise of the optimist camp.
A word that keeps coming up in the AI conversation of 2026.
A word that keeps coming up in the AI conversation of 2026.
A word that keeps coming up in the AI conversation of 2026.
First half of the oldest proverb in computing.
A word that keeps coming up in the AI conversation of 2026.
Nvidia data-center GPU, 2022.
Term for a model stating false things fluently.
It's not a feature it's a sample
Everything a model outputs is a sample from a distribution.
A word that keeps coming up in the AI conversation of 2026.
A word that keeps coming up in the AI conversation of 2026.
A word that keeps coming up in the AI conversation of 2026.
A word that keeps coming up in the AI conversation of 2026.
A word that keeps coming up in the AI conversation of 2026.
A word that keeps coming up in the AI conversation of 2026.
A word that keeps coming up in the AI conversation of 2026.
Slogan of the scaling era.
A word that keeps coming up in the AI conversation of 2026.
A word that keeps coming up in the AI conversation of 2026.
A word that keeps coming up in the AI conversation of 2026.
Our own line, written for this wall. Not a quotation.
Said whenever models max out a test.
The benchmark is the territory
A play on Alfred Korzybski's "the map is not the territory": once models are trained to beat a test, the test becomes the thing itself.
Our own line, written for this wall. Not a quotation.
Our own line, written for this wall. Not a quotation.
A word that keeps coming up in the AI conversation of 2026.
A word that keeps coming up in the AI conversation of 2026.
A word that keeps coming up in the AI conversation of 2026.
- August 2026
Perhaps we get abundance, but at the price of redundance
From a 2026 blog post by AI researcher Julian Togelius reflecting on a 'crisis of faith' about the implications of AI research succeeding.
- May 2026
You certainly don't tell a researcher like me what to do
Yann LeCun to the Financial Times, 2026, on his departure from Meta before founding AMI Labs.
Financial TimesComputer scientist Yann LeCun: 'Intelligence really is about learning'
- April 2026
It's more like we're poking something in a petri dish
AI developer Connor Leahy describing deep learning model training, quoted in Garrison Lovely's book excerpt.
obsolete.pubIn Defense of “Artificial Intelligence” — a Book Excerpt
- February 2026
In the long term, space-based AI is obviously the only way to scale
Elon Musk, in SpaceX's February 2026 announcement of its acquisition of xAI, on the rationale for building AI data centers in space.
- January 2026
It's always two years away from being two years away
Jeffrey Ding, commenting on predictions of AGI arrival, quoted in Transformer newsletter, January 2026.
Yann LeCun, in a recent interview, described Silicon Valley's consensus that large language models will lead to AGI.
- December 2025
Thinking, as it turns out, is quite hard
Nvidia CEO Jensen Huang, discussing AI inference difficulty in a conversation reported by Fortune, December 2025.
FortuneNvidia’s Groq bet shows that the economics of AI chip-building are still unsettled
Yann is just plain incorrect here
Google DeepMind CEO Demis Hassabis responded on X to Yann LeCun's comments on general intelligence, December 2025.
- November 2025
These models somehow just generalize dramatically worse than people
Ilya Sutskever, in an interview discussed on Gary Marcus's Substack, November 2025, on the limits of scaling large language models.
garymarcus.substack.comA trillion dollars is a terrible thing to waste
Benjamin Riley, founder of Cognitive Resonance, writing in The Verge (2025) on why large language models are not the same as thinking.
- October 2025
You're sucking supervision through a straw
Andrej Karpathy, on Dwarkesh Patel's podcast in 2025, describing the weakness of reinforcement learning for training AI models.
understandingai.orgAI skeptics and AI boosters are both wrong
We're in an intelligence explosion already and have been for decades
Andrej Karpathy, on Dwarkesh Patel's podcast in 2025, arguing that AI progress is a continuation of decades of computing advances rather than a sudden new phenomenon.
understandingai.orgAI skeptics and AI boosters are both wrong
Signs of psychosis or mania in 0.07% of weekly users
OpenAI's own figure, October 2025. At 800 million weekly users that is more than half a million people.
OpenAIStrengthening ChatGPT's responses in sensitive conversations
- September 2025
Large language models are about mimicking people
Richard Sutton, father of reinforcement learning, contrasting LLMs with reinforcement learning on the Dwarkesh Podcast, September 2025.
dwarkesh.comRichard Sutton – Father of RL thinks LLMs are a dead end
- August 2025
They live in the land of correlations between images and text, not in the land of ideas
Gary Marcus, on his Substack in August 2025, discussing errors in AI-generated images of a unicyclist and bass guitar.
We're less special than we thought
Nobel laureate Geoffrey Hinton, known as the Godfather of AI, discussing how creating thinking machines could undermine religious ideas of human uniqueness.
The MainichiHow Silicon Valley is using religious language to talk about AI - 毎日新聞
We are seeing glimpses of our AI systems starting to improve themselves
Meta CEO Mark Zuckerberg speaking via video at Reliance Industries' AGM, August 2025, on AI progress.
Mustafa Suleyman's 2025 term for systems that convincingly appear conscious without being so.
mustafa-suleyman.aiWe must build AI for people; not to be a person
- July 2025
You run the risk of killing breakthrough innovation
Yann LeCun responded on LinkedIn to Elon Musk's decision to eliminate the 'researcher' job title at xAI, July 2025.
- April 2025
Most interesting problems scale extremely badly
Yann LeCun, Meta's chief AI scientist, speaking at the National University of Singapore, April 2025, on the limits of AI scaling laws.
Business InsiderMeta's Yann LeCun: Scaling AI Won't Make It Smarter
Human intelligence is not general at all
Yann LeCun, then Meta's chief AI scientist, on Jason Howell's AI Inside podcast in April 2025, arguing that human intelligence is highly specialized rather than general.
AI Inside (Jason Howell)Meta AI Chief Yann LeCun: Human Intelligence Is Not General Intelligence
- March 2025
Why are hands so goddamn hard?
Jack Clark, in his Import AI newsletter (March 2025), commenting on the difficulty of building dexterous robot hands.
10,000 Einsteins who don't get tired
Michael Horowitz, professor at Penn, discussing what AGI could mean for military innovation in a ChinaTalk interview on AI and warfare, March 2025.
All they can do is double down
Stuart Russell, computer scientist at UC Berkeley, to Live Science, March 2025, on AI companies' response to diminishing returns from scaling.
Live ScienceCurrent AI models a 'dead end' for human-level intelligence, scientists agree
We are not going to get to human-level AI by just scaling up LLMs
Meta chief AI scientist Yann LeCun, quoted by AFP in a March 2025 article on scepticism toward AGI predictions.
AFP (France 24)Firms and researchers at odds over superhuman AI
Carl Bergstrom, an evolutionary biologist at the University of Washington, quoted in a Nature report on AI use in peer review, March 2025.
NatureAI is transforming peer review — and many scientists are worried
- February 2025
Now we really know what Ilya saw
Nathan Lambert, in Interconnects analysis of GPT-4.5's release, February 2025, referring to reports that Ilya Sutskever had foreseen limits to pretraining scaling.
- December 2024
Scaling as we know it is ending
From Nathan Lambert's 2024 year-in-review on Interconnects, describing the impact of OpenAI's o1 model on the AI training paradigm.
ARC-AGI is not an acid test for AGI
François Chollet pushing back on the interpretation of his ARC-AGI benchmark as a definitive test for artificial general intelligence.
aiguide.substack.comDid OpenAI Just Solve Abstract Reasoning?
- November 2024
Maybe there are just no mysteries
Jack Clark, in Import AI #392 (November 2024), reflecting on new research showing scaling laws appear in world modeling and behavioral cloning just as in language models.
- September 2024
The computer doesn't care where it's going to school
Jensen Huang, NVIDIA CEO, speaking during a Bipartisan Policy Center webinar on Sept. 27, 2024, on building data centers near abundant energy sources.
Utility DiveAI will help, not hinder, grid decarbonization: NVIDIA CEO
An alien of extraordinary ability
Timothy Lee's assessment of OpenAI's o1 model, describing its capabilities while noting it falls short of AGI, September 2024.
- June 2024
Sam Altman speaking at Aspen Ideas, June 2024, on AI's future capabilities, as cited by Gary Marcus.
garymarcus.substack.comThe need for a President that speaks AI natively
World models should not be generative
Yann LeCun, VP and Chief AI Scientist at Meta, posted this statement about AI world models on LinkedIn, June 2024.
- May 2024
Current LLMs are dumber than a house cat
Yann LeCun, 2024.
The Wall Street JournalThis AI Pioneer Thinks AI Is Dumber Than a Cat
- April 2024
They literally don't know the difference between truth and falsehood
Gary Marcus, on his Substack in April 2024, explaining why LLMs hallucinate compared to human errors.
garymarcus.substack.comHumans versus Machines: The Hallucination Edition
- February 2024
The brain is an existence proof that general intelligence is possible at all
Demis Hassabis, CEO of Google DeepMind, discussing neuroscience and AI on the Dwarkesh Podcast, February 2024.
dwarkesh.comDemis Hassabis — Scaling, superhuman AIs, AlphaZero atop LLMs, AlphaFold
We don't know why they make decisions and neither do they
From an essay by Erik J. Larson on Mind Matters (2024) arguing that current AI systems are opaque even to their creators.
Mind Matters (Erik J. Larson)If AI's Don't Know What They're Doing, Can We Hope to Explain It?
- January 2024
Gary Marcus's term for baffling failures of neural network systems, used in his January 2024 Substack post about AI image generation mistakes.
garymarcus.substack.comWhere’s Waldo? The Elephant in the Room - Marcus on AI
- December 2023
Yann LeCun, Meta's chief AI scientist, in a December 2023 WIRED interview, explaining limits of current machine learning approaches.
- November 2023
Way below cat-level intelligence, never mind human-level
Yann LeCun, Meta's chief AI scientist, in an X reply to Geoffrey Hinton in November 2023, arguing that LLMs' understanding is limited and superficial.
- October 2023
I don't see big walls in front of us
Shane Legg on the remaining obstacles to AGI, discussing model shortcomings like memory and factuality, Dwarkesh Podcast, October 2023.
dwarkesh.comShane Legg (DeepMind Founder) — 2028 AGI, superhuman alignment, new architectures
- September 2023
You have no business running all the world's software
Gary Marcus, in a Substack post about a paper showing LLMs fail to infer that 'B is A' from 'A is B', argues this reasoning failure disqualifies such models from broad deployment.
Pointillistic masses of blurry memory
Gary Marcus describes neural networks' inability to generalize symmetrical relationships, in a 2023 Substack post discussing a new paper on LLM reasoning failures.
- August 2023
Term for delusions fed by a chatbot that never contradicts. Søren Dinesen Østergaard, Aarhus, warned of it in 2023.
- July 2023
I think they could well have feelings
Geoffrey Hinton, speaking at King's College London in 2023, discussing whether AI systems could have emotions like frustration and anger.
- March 2023
Yann LeCun and Jacob Browning describing AI chatbots' lack of genuine social understanding, quoted in Noema, 2023.
NOEMAThe AI We Empower Will Demand More From Us - Noema Magazine
Bryan Caplan, after losing a three-year public bet that no AI would ace his economics exam by 2029, conceded GPT-4 had done it, in a blog reaction covered by Zvi Mowshowitz in March 2023.
Sparks of artificial intelligence
Phrase from a Microsoft Research paper on GPT-4, cited in a March 2023 AI news roundup by Zvi Mowshowitz.
Predicting the next token well means that you understand the underlying reality
Ilya Sutskever, OpenAI chief scientist, in a March 2023 Dwarkesh Podcast interview, explaining what next-token prediction implies about a model's understanding of the world.
I challenge the claim that next-token prediction cannot surpass human performance
Ilya Sutskever, OpenAI chief scientist, in a March 2023 Dwarkesh Podcast interview, discussing whether next-token prediction can exceed human-level intelligence.
- February 2023
Susan Schneider, founding director of Florida Atlantic University's Center for the Future Mind, on chatbots' unreliable claims about their own consciousness, NBC News, 2023.
NBC NewsWhat is consciousness? ChatGPT and advanced AI might redefine our answer
Ted Chiang's image for what a large language model holds: a lossy compression of everything online.
- December 2022
Yejin Choi, MacArthur Fellow and UW/Allen AI professor, describing commonsense reasoning at the December 2022 AGI Debate.
Data is the moat for ML systems right now
From a 2022 essay on machine learning business moats by Nathan Lambert, discussing what protects AI companies from competitors.
- June 2022
Who am i to tell god where he can and can't put souls
Blake Lemoine, the Google engineer who claimed LaMDA was sentient, defending his belief on Twitter in June 2022, citing his Christian faith.
Gary Marcus, AI researcher, on Blake Lemoine's claim that Google's LaMDA chatbot was sentient, June 2022.
- February 2022
Today's large neural networks may be slightly conscious
Ilya Sutskever on X, February 2022.
- April 2021
AI is neither artificial nor intelligent
Kate Crawford's opening argument in Atlas of AI: it runs on minerals, energy and human labour.
- March 2021
Coined in the 2021 paper "On the Dangers of Stochastic Parrots" for language models that recombine text without understanding.
- January 2020
Empirical relations between compute, data, parameters and model quality.
- March 2019
Rich Sutton, 2019: general methods that leverage computation beat human-designed cleverness.
- June 2017
Title of the 2017 paper that introduced the transformer.
- March 2016
AlphaGo's 37th move in game two against Lee Sedol, 10 March 2016. Commentators took it for a mistake. It won the game.
WIREDThe Two Moves AlphaGo and Lee Sedol Redefined the Future
- January 1965
I. J. Good's 1965 scenario of machines improving machines.
The last invention that man need ever make
I. J. Good, 1965: an ultraintelligent machine could design better machines, so it would be the last invention humans need make.
Advances in ComputersSpeculations Concerning the First Ultraintelligent Machine
- January 1951
We should have to expect the machines to take control
Turing in a 1951 lecture, anticipating machines that outstrip human powers.
WikiquoteIntelligent Machinery, A Heretical Theory (lecture, 1951)
- October 1950
The first sentence of Turing's 1950 paper. He replaced the question with the imitation game.