John Bd @CodePhantoy

Dwarkesh and I had a frank exchange of views. I hope we moved the conversation forward. Dwarkesh is a true gentleman.

Dwarkesh Patel @dwarkesh_sp

9 months ago

.@RichardSSutton, father of reinforcement learning, doesn’t think LLMs are bitter-lesson-pilled. My steel man of Richard’s position: we need some new architecture to enable continual (on-the-job) learning. And if we have continual learning, we don't need a special training

248 620 4K 3.1M 4K

As a researcher at a frontier lab I’m often surprised by how unaware of current AI progress public discussions are. I wrote a post to summarize studies of recent progress, and what we should expect in the next 1-2 years: julian.ac/blog/2025/09/2…

🇨🇳 DeepSeek-R1 was published in Nature yesterday as the cover article for their BRILLIANT latest research. They show that pure Reinforcement Learning with answer-only rewards can grow real reasoning skills, no human step-by-step traces required. So completely skip human reasoning traces and still get SOTA reasoning via pure RL. It’s so powerful revelation, because instead of forcing the model to copy human reasoning steps, it only rewards getting the final answer right, which gives the model freedom to invent its own reasoning strategies that can actually go beyond human examples. Earlier methods capped models at what humans could demonstrate, but this breaks that ceiling and lets reasoning emerge naturally. Those skills include self-checking, verification, and changing strategy mid-solution, and they beat supervised baselines on tasks where answers can be checked. Models trained this way also pass those patterns down to smaller models through distillation. AIME 2024 pass@1 jumps from 15.6% to 77.9%, and hits 86.7% with self-consistency. ⚙️ The Core Concepts The paper replaces human-labelled reasoning traces with answer-graded RL, so the model only gets a reward when its final answer matches ground truth, which frees it to search its own reasoning style. The result is longer thoughts with built-in reflection, verification, and trying backups when stuck, which are exactly the skills needed for math, coding, and STEM problems where correctness is checkable. This matters because supervised traces cap the model at human patterns, while answer-graded RL lets it discover non-human routes that still land on correct answers.

This is a neural network named Helix learning how to do laundry

This man simply walked into the fashion show wearing a garbage bag - and no one thought it was strange until security removed him.

This will become normal in a few years

DogeDesigner @cb_doge

11 months ago

Optimus serving popcorn is the best thing you'll see on the internet today. 🍿 x.com/tesla_archive/…

3K 3K 17K 69.7M 3K

Les taxis autonomes explosent aux États-Unis 200 millions de kilomètres pour les seuls taxis sans chauffeurs @Waymo de @Google Google affirme que les 2 milliards de kilomètres approchent ! Si vous possédez un taxi, vendez votre plaque : sa valeur va s’effondrer !

Claude for Financial Services Keynote (getting a lot of buzz today) youtube.com/watch?v=5zd7m3…

Turing Award winner Richard Sutton says humanity's purpose is to create what comes next. Our role is to design something that can design. AI is that thing. “We are the catalyst. The midwife. The progenitor of the fourth great age of the universe.”

@FabrizioBucella ... à condition que le livre soit parallèle à la tour ... Sinon il faut diviser le résultat par sin(theta)

@NicolaRic2 @SimonDunand Sehr cool! 🚀 As a next step it'd be cool to have an interpretation of the text by a hypothetical someone who's SP, FDP, SVP etc

*ANTHROPIC CEO: AI COULD WIPE OUT HALF OF ALL ENTRY-LEVEL WHITE COLLAR JOBS AND SPIKE UNEMPLOYMENT TO 10-20% WITHIN THE NEXT FIVE YEARS... GOVERNMENTS AND COMPANIES NEED TO STOP SUGAR-COATING WHAT THEY KNOW IS COMING

ai has come really really far... #Veo3

Two years after it was left nursing losses from the rapid collapse of Credit Suisse, the head of Saudi Arabia’s Public Investment Fund said it will no longer invest in the country’s financial markets bloomberg.com/news/articles/…

source: nature.com/articles/d4158…

@goddek Freedom

@paulg @sama I got 0.072 m² height = num_weights * min_etchible_number_height width = min_etchible_number_width * 16 bits 1.8 trillion parameters model (allegedly) smallest feature etching 50nm (1.8 × 10¹² weights) * (5 × 10⁻⁸ m) * (5 × 10⁻⁸ m) * 16 = 0.072 m²

It's spring break and my son started a lemonade stand. I asked him how much lemonade he sold in the last hour. "$3" he said. I told him, "Here's $20 for one glass of lemonade. In one hour I'm going to come back and buy a lemonade for $1,000. You're going to take that $1,000/hr revenue - multiply it by 24 hours - then multiply it by 365 days, and tell VCs, 'I have $8.76M in run rate revenue with 99.9% net margins, and 100% of our latest cohort of customers have expanded revenue with us. You're also changing the name of your lemonade stand to 'lemon(AI)de' and incorporating it as a C Corp" My son was holding back tears. I said, "Doesn't it feel good to be a millionaire, bud?" He said, "Dad I'm 4 and I don't know what any of these words are." I said, "Welcome to life as a tech CEO - neither do I" and drove away.

Historic moment: The first humanoid robot to cross the finish line of a half marathon - TienKung (Beijing Humanoid Robot Innovation Center)

@klarnaseb @vonderleyen I never understood why open banking never took off in Europe - independence from the big credit card players (which are not European) - less cost/less friction - leverage already existing instantSEPA - framework (both tech and legal) is here - enable a lot of fintech use cases