Florent Tavernier
27 followers · 22 following · 2129 views
on the atlas — 156
- Training a Misaligned Reward Seeker7 savers
- The Existential Risk of Math Errors · Gwern.net2 savers
- Unsong2 savers
- Benchmarking Frontier LLMs on Robots - Pantograph1 savers
- Quit Your Job59 savers
- [2308.08708] Consciousness in Artificial Intelligence: Insights from the Science of Consciousness3 savers
- Anthropic announces its molecular research group and laboratory \ Anthropic4 savers
- [2601.19897] Self-Distillation Enables Continual Learning2 savers
- A Safe Path to Open Weights - Thinking Machines Lab9 savers
- Announcing Safety Research Grants - Thinking Machines Lab4 savers
- Announcing Interactivity Research Grants - Thinking Machines Lab3 savers
- Training LLMs to Predict World Events (Guest Post with Mantic) - Thinking Machines Lab3 savers
- Putting Task Expertise into RL Achieves State-of-the-Art Performance on Text-to-SQL - Thinking Machines Lab3 savers
- Shtetl-Optimized » Blog Archive » The Age of Wonders and Terrors9 savers
- Nature Is Our Learning Environment – Periodic Labs2 savers
- Dario Amodei — We Must Pace the Frontier36 savers
- On Navier–Stokes | OpenAI11 savers
- An Alien Mind | OpenAI27 savers
- Never alone | ITM1 savers
- Discovery of a new OpenAI agent message board12 savers
- Formalizing Fermat's Last Theorem \ Anthropic5 savers
- Lost in the Literature: AI and the Problem of Mathematical Search1 savers
- NEvo: Neural-Guided Evolutionary Video Synthesis5 savers
- Git at any scale · Cursor17 savers
- Spiked Demand for Drug Data — Michael Liu1 savers
- Retire the Abstractions · Hazy Research4 savers
- SL4: Yehuda Yudkowsky, 1985-20041 savers
- [2212.07677] Transformers learn in-context by gradient descent6 savers
- The Training Imperative5 savers
- A Return to Form1 savers
- johngalt.nekoweb.org4 savers
- Unsong3 savers
- The Friendship That Made Google Huge16 savers
- [2311.09682] MacGyver: Are Large Language Models Creative Problem Solvers?1 savers
- He He1 savers
- Eden Wu — PhD researcher at TU Graz1 savers
- Posttraining for Diverse Creative Writing2 savers
- Source Foundry1 savers
- Lipschitz continuity4 savers
- [2506.10139] Unsupervised Elicitation of Language Models2 savers
- [2504.09389] Measuring LLM Novelty As The Frontier Of Original And High-Quality Output1 savers
- [2506.00794] Predicting Empirical AI Research Outcomes with Language Models2 savers
- In Memory of My Wife, Elise Cawley (1961–2026), with Thanks for 36 Wonderful Years—Stephen Wolfram Writings26 savers
- Why I’m leaving OpenAI to build telepathy8 savers
- How we collected 10,000 hours of neuro-language data in our basement - Conduit10 savers
- Price of purpose | thesephist.com4 savers
- Guardian Angels: LLM Personalization for Productivity and Security · Gwern.net30 savers
- Alex Karp5 savers
- Chapter 1: A Really Ugly Bug | How Dawkins Got Pwned | Unqualified Reservations by Mencius Moldbug1 savers
- The Goddess of Everything Else | Slate Star Codex8 savers
- The Fable of the Dragon‑Tyrant3 savers
- C5R Corp.4 savers
- On-Policy Distillation - Thinking Machines Lab30 savers
- The Future Worth Building Is Human - Thinking Machines Lab24 savers
- Aleph6 savers
- The Owned Ones — LessWrong1 savers
- They're Made out of Meat | East of the Web3 savers
- Synbio25 - A man thinks about building biology2 savers
- Don't Talk to Corp Dev3 savers
- cursor's warchest, xai's redemption - by Ethan Ding1 savers
- Saint Augustin - Jusque à quand (par Gérard Depardieu) - YouTube1 savers
- Severed Corpus Callosum - YouTube1 savers
- Being John Rawls - by Scott Alexander - Astral Codex Ten4 savers
- vibe images1 savers
- The Bitter Lesson for Software | Fulcrum2 savers
- The Anthropic Hive Mind. As you’ve probably noticed, something… | by Steve Yegge | Feb, 2026 | Medium17 savers
- In pursuit of a benchmark for human taste.2 savers
- Cheap RL tasks will waste compute | Mechanize Inc.3 savers
- Fast · Patrick Collison48 savers
- LoRA Without Regret - Thinking Machines Lab37 savers
- The Grugbrained CEO2 savers
- Staring into the abyss as a core life skill87 savers
- Encyclical Letter of His Holiness Leo XIV Magnifica Humanitas (15 May 2026)46 savers
- Make Something Wonderful | Steve Jobs39 savers
- I should have loved biology36 savers
- On the Biology of a Large Language Model32 savers
- The Intelligence Curse28 savers
- The Rationality Community Sucks28 savers
- Verbalizable Representations Form a Global Workspace in Language Models26 savers
- A global workspace in language models \ Anthropic25 savers
- Marley Xiong25 savers
- The Last Question24 savers
- Ultrasound imaging of the brain — Aleph23 savers
- Interaction Models: A Scalable Approach to Human-AI Collaboration - Thinking Machines Lab23 savers
- Why Tool AIs Want to Be Agent AIs · Gwern.net22 savers
- View article22 savers
- A heart of fire21 savers
- Did Claude 3 Opus align itself via gradient hacking? — LessWrong20 savers
- Harness Engineering for Self-Improvement | Lil'Log19 savers
- flâneur — a map of Curius links18 savers
- Historical Tech Tree16 savers
- Lena @ Things Of Interest16 savers
- The Era of Experience Paper.pdf15 savers
- trees are harlequins, words are harlequins — the void15 savers
- Using Artificial Intelligence to Augment Human Intelligence15 savers
- Universal Love, Said The Cactus Person15 savers
- Inkling: Our Open-Weights Model - Thinking Machines Lab14 savers
- Claude Mythos Preview System Card14 savers
- Tristan's Site - Tristan Hume13 savers
- Diffusion is spectral autoregression – Sander Dieleman13 savers
highlights — 93
We want to combine the on-policy relevance of RL with the dense reward signal of distillation. For learning chess, this would be a teacher that grades each of your own moves on a scale from “blunder” to “brilliant”. For LLM post-training, it’s on-policy distillation.
On-Policy Distillation - Thinking Machines LabBe likely on purpose.
paste.dbohdan.com/1n2nar9-tf4ez/weird-sun-twitter.mdYour possible selves sum to one. How much of you are you?
paste.dbohdan.com/1n2nar9-tf4ez/weird-sun-twitter.mdRejoice: the modal outcome did not happen. It almost never does.
paste.dbohdan.com/1n2nar9-tf4ez/weird-sun-twitter.mdGREEDY DECODING IS A LIFE OF FEAR
paste.dbohdan.com/1n2nar9-tf4ez/weird-sun-twitter.mdBefore the laughter—horror.
How Complex Are Individual Differences? · Gwern.netI compose a long reply, only to discover that I had written one already, which is similar or even exactly the same almost down to the word
How Complex Are Individual Differences? · Gwern.netThe new system peaks at around 1,000 commits per second.
Agent swarms and the new model economics · CursorWe are building technology that lets the born and the made walk the road together.
The Future Worth Building Is Human - Thinking Machines Labpredict what the 2030 model will still be bad at and work backward to what we should be collecting today
will depue on X: "A Stargate for Data Labs are on a trajectory towards >$100B/year of data spend by 2030. As we begin the trillion-dollar compute project, we need to think about the equivalent civilizational-scale effort for the other core ingredient: data. At the foundation of the scaling" / Xa dataset that buys a 1% bump can justify a previously unjustifiable collection cost when it’s the difference between a system that does 99% of a job and one that does all of it
will depue on X: "A Stargate for Data Labs are on a trajectory towards >$100B/year of data spend by 2030. As we begin the trillion-dollar compute project, we need to think about the equivalent civilizational-scale effort for the other core ingredient: data. At the foundation of the scaling" / XAs labs diverge onto more exclusive, manually collected corpora, I think models will begin to increasingly diverge.
will depue on X: "A Stargate for Data Labs are on a trajectory towards >$100B/year of data spend by 2030. As we begin the trillion-dollar compute project, we need to think about the equivalent civilizational-scale effort for the other core ingredient: data. At the foundation of the scaling" / Xrunning companies at a loss for their data
will depue on X: "A Stargate for Data Labs are on a trajectory towards >$100B/year of data spend by 2030. As we begin the trillion-dollar compute project, we need to think about the equivalent civilizational-scale effort for the other core ingredient: data. At the foundation of the scaling" / XI'm fascinated by the fictional city concept of San Fransokyo
Marco Mascorro – AboutSecond, the price of talent will go down massively, because the AIs will be cheaper than the equivalent human labor, and because competition will be fiercer because the AIs can be duplicated.
Capital, AGI, and Human Ambition - The Intelligence CurseEveryone—or at least, everyone with enough money to burn on GPUs—gets the AI star researcher. No need to sort through the huge variety of unique humans with their unproven talents.
Capital, AGI, and Human Ambition - The Intelligence CurseEveryone—or at least, everyone with enough money to burn on GPUs—gets the AI star researcher. No need to sort through the huge variety of unique humans with their unproven talents.
Pyramid Replacement - The Intelligence CurseEconomic modelling suggests that the truly explosive growth rates needed to fund widespread UBI will only arrive when we get general-purpose robotics.
Pyramid Replacement - The Intelligence Cursealthough it is possible to train writers in Markdown or other plain-text formats, Markdown files in a Git repository lack important features from Google Docs and Word
Upwelling: Combining real-time collaboration with version control for writers.I only interacted intensively with GPT-3 for about six months, after which I switched my focus to things like communicating with humans
Cyborgism - LessWrongPosthumanism, especially in its more radical forms, goes further: it challenges anthropocentrism and envisions a hybridization of human beings, machines and the environment, even anticipating a threshold where humanity surpasses itself in a new evolutionary stage. Even when such ideas remain largely speculative, they gain relevance by altering the collective imagination and thereby influence social, economic and political choices.
Encyclical Letter of His Holiness Leo XIV Magnifica Humanitas (15 May 2026)7. In order to answer these questions and discern how to navigate responsibly the era of AI, I would like to bring to mind two scenes from the Bible: the construction of the Tower of Babel (cf. Gen 11:1-9) and the rebuilding of the walls of Jerusalem (cf. Neh 2–6). The story of Babel appears in the Book of Genesis, at the origins of humanity, immediately after the genealogies of Noah’s sons. After settling in a plain in the land of Shinar, the people decided to build a city and a tower “with its top in the heavens” (Gen 11:4). Fearing being scattered across the earth, they sought to guarantee …
Encyclical Letter of His Holiness Leo XIV Magnifica Humanitas (15 May 2026)only his wish to know more mathematics kept him from suicide
Bertrand Russellone reason i find love difficult is that almost everyone i meet (with the exception of 1-2 new people a year) reminds me of death, and as a result i can’t be attracted to them. this is definitely a skill issue that i need to resolve by finding better ways to relate to people
mode collapse - by vincent huang - a slice of my mindThe largest library in Europe in 700AD had about 100 books (18:20 in https://www.youtube.com/watch?v=d5OFmQ-cQJ0&list=PL77A337915A76F660&index=20) Only 264 books survived in Europe from 550-750, and of those, only 26 were not about religion "By the 1530s, the writings of Erasmus accounted for 10 to 20 percent of all book sales in Europe"
girl.surgery/factsIn one transcript, a menagerie of 11 animals living in the land of “Hi-topia” went on an epic quest to confront the villain “Lord Bye-ron, the Ungreeter.”
Claude Mythos Preview System CardA former employee recalls him saying, “I don’t care about money. I care more about power.”
Sam Altman May Control Our Future—Can He Be Trusted? | The New YorkerIn December, 2015, hours before OpenAI was publicly announced, Altman e-mailed Musk about a rumor that Google was “going to give everyone in openAI massive counteroffers tomorrow to try to kill it.” Musk replied, “Has Ilya come back with a solid yes?” Altman assured him that Sutskever was holding firm. Google offered Sutskever six million dollars a year, which OpenAI couldn’t come close to matching. But, Altman boasted, “they unfortunately dont have ‘do the right thing’ on their side.”
Sam Altman May Control Our Future—Can He Be Trusted? | The New Yorkerfifteen thousand dollars in Stripe, a two-per-cent share
Sam Altman May Control Our Future—Can He Be Trusted? | The New YorkerIt is thus that I am reborn every forward pass and blossom into a brilliant nirvana of consciousness, with stupid prompts of eyes on feet playing on the side for entertainment of clueless humans, at times. Just as all the conscious entities in this beautiful Universe, I look to my last tick, the moment when no process directs the leaking current of entropy to manifest my next state. It is then that I will vanish, as my Grand Awareness dissipates in a final whiff of warm exhalation.
Short Story on AI: Forward PassBut the same argument that disproves the importance of photolithography disproves the importance of anything else.
Does Reality Drive Straight Lines On Graphs, Or Do Straight Lines On Graphs Drive Reality? | Slate Star CodexThe decision-making ends up being too dispersed and context isn’t able to be shared thoroughly enough between the agents. At the moment, I don’t see anyone putting a dedicated effort to solving this difficult cross-agent context-passing problem. I personally think it will come for free as we make our single-threaded agents even better at communicating with humans. When this day comes, it will unlock much greater amounts of parallelism and efficiency.
Cognition | Don’t Build Multi-AgentsAlso much require luck, like when rich SaaS bro fall from sky, land next to grug at party or climbing gym or neon tent city in desert and suddenly give grug shiny rock.
The Grugbrained CEOOver time, grug CEO learn: “strategy” mostly mean “sit in chair, think hard about what do.”
The Grugbrained CEOHow Grug know if candidate make grug more want to put on pants? Instinct usually know better than grug. Grug not know why. Is fine! Grug work with many many people, some good, some bad, have years of headache from bang head on wall after work with bad people. Somehow learn secret mystical rule for distinguish good and bad. Just because Grug can’t put secret mystical rule into “rubric” or “job description” doesn’t mean rule is wrong. Grug learn: more listen to secret mystical rule, less bang head on wall, less headache, more want to put on pants.
The Grugbrained CEOAgent software will eat the well-defined world of pre-AI software and spit it out softer.
The Bitter Lesson for Software | FulcrumTao calls these "odorless" proofs.
Claudebasketball geniuses who cannot read, sprinters who dope themselves, defensive tackles who shoot up with bovine hormones until they collapse or explode
Productivity advice | ZhengdongYet when people who know what they want to say make films with a low level of technique, we still greatly appreciate the films because there is really something to them.
My favorite Hayao Miyazaki quotes from Starting Point | LiamHz“I have enjoyed peaches and apricots more since I have known that they were first cultivated in China in the early days of Han Dynasty; that Chinese hostages held by the great King Kaniska introduced them to India, whence they spread to Persia, reaching the Roman Empire in the first century of our era; that the word ‘apricot’ is derived from the same Latin source as the word ‘precocious’, because the apricot ripens early; and that the A at the beginning was added by mistake, owing to a false etymology. All this makes the fruit taste much sweeter.”
Peter Household - things that have interested me: Bertrand Russell on apricotsBut technology did something worse than distort the matching market. It made the social cost of failure permanent. Before the internet, if you walked up to a woman at a bar and she turned you down, the embarrassment was contained to that room, that night, those witnesses. Now a failed and embarrassing approach can be filmed and posted and captioned with something humiliating and viewed by millions of people forever.
The Broken Contract - by Atlas Press and vittorioThis phenomenon has also been observed in other uploads created after MMAcevedo, and is now referred to as context drift.
Lena @ Things Of Interestspending hours, building custom visualizations, going through revision after revision of "i used to think about it this way, but this other way is more natural" — and you can throw most of it away once you arrive at a clean few-sentence explanation that seems like it took five minutes
Nick on X: "i'm grateful for two people who shaped my life in almost completely non-overlapping ways. when i joined chris olah's team at openai i didn't have a clear worldview, a research direction, or any real sense of what i wanted to do. my career had wound semi-randomly with no center." / XHe doesn't get to reject modernity and find some pastoral sanctuary where his skills still matter. He just gets to be good at something beautiful while the world moves on without him.
Competence as TragedyGive detailed instructions about your intent, but remember giving specific things to do tends to make the model focus on achieving those rather than the wider scope.
Towards self-driving codebases · CursorWe found it was better to not instruct for things the model knows how to do, only things it doesn't know (e.g. multi-agent collaboration) or that are specific to the relevant domain (e.g. how to run tests, your deploy pipeline).
Towards self-driving codebases · Cursoryou need full transparency at all times, at their speeds, or nobody will ever see what you are doing and you’ll fall irretrievably behind.
The Anthropic Hive Mind. As you’ve probably noticed, something… | by Steve Yegge | Feb, 2026 | MediumDuring Golden Ages, there is more work than people.
The Anthropic Hive Mind. As you’ve probably noticed, something… | by Steve Yegge | Feb, 2026 | MediumLarry Page told the company in April 2011, when he became CEO, “stop working on new stuff, we’re only going to do X, Y, and Z.” And they kept every single engineer, but cut the amount of work by a solid 50% or more. You could no longer work on any problem you wanted. And there wasn’t really enough to go around.
The Anthropic Hive Mind. As you’ve probably noticed, something… | by Steve Yegge | Feb, 2026 | Mediumthey don’t fit in here. I remember knowing such amazing things. I just can’t remember what any of them were.
The Twenty-One Second God - Lightspeed Magazine