Joe Cole - e/acc

My take: it's still early innings for RL. The ceiling for open models keeps moving up, and the players like @cursor_ai that are investing for real in post-training are going to keep pushing the frontier. RL isn't for everyone, but for AI startups it should probably be a fast-follow from PMF.

🛒 PRE-ORDER ALERT! 🛍️ Need 32GB of VRAM for local AI without breaking the bank? 🧠 💻 The highly anticipated Intel Arc Pro B70 workstation GPU is officially available for pre-order on Newegg! 🛒✨ 🔥 The Specs: 32GB GDDR6 VRAM, 32 Xe Cores, & 367 peak TOPS 💸 The Price: $949.99 📅 Release Date: April 24, 2026

Announcing our first book: Silicon A beautiful coffee table book about the world of transistors, chips, and the greatest technology revolution of all time. 384 pages. Almost five pounds. Preorders open now, shipping in May: arenamag.com/silicon

The awkward truth is that what counts as a good engineer just became a different thing in the last 4 months

We have to accelerate

If you’re interested in reinforcement learning and neuroscience, and are fascinated by the connection between them, I’ve been writing an eight-part series called The RL Spiral. robonaissance.com/t/the-rl-spiral The title is quite literal. Reinforcement learning and neuroscience have not developed in parallel; rather, they have spiraled around each other, with each revolution deepening our understanding of both fields. That spiral began over a century ago, and we are still within it. I’ve completed four parts so far and expect to finish the series within the next two to three weeks. I hope you enjoy it. Stay tuned.

Anti-AI movement will look like this in 20 years

We open sourced a version of the world's most capable AI co-scientist. Free. Easy to install. Has access to our Scientific Skills that are in use by 150k+ scientists worldwide. Please star the GitHub repo and repost/retweet.

16 Reinforcement Learning approaches you should know about (classic + modern) ▪️ RLHF – RL from Human Feedback ▪️ RLAIF – RL from AI Feedback ▪️ RLVR – RL with Verifiable Rewards ▪️ RLCF – RL from Community Feedback (2 different variants) ▪️ RLCF – RL from Checklist Feedback ▪️ CM2 ▪️ Critique-RL ▪️ CRL – Critique RL ▪️ ICRL – In-Context RL ▪️ RLBF – RL with Backtracking Feedback ▪️ TriPlay-RL ▪️ SPIRAL ▪️ Co-rewarding ▪️ RESTRAIN ▪️ PRL – Process Reward Learning ▪️ RLSF – RL from Self-Feedback Save this list and check it out for links and explanations: turingpost.com/p/rlapproaches

Code is an output. Nature is healing. For too long we treated code as input. We glorified it, hand-formatted it, prettified it, obsessed over it. We built sophisticated GUIs to write it in: IDEs. We syntax-highlit, tree-sat, mini-mapped the code. Keyboard triggers, inline autocompletes, ghost text. “What color scheme is that?” We stayed up debating the ideal length of APIs and function bodies. Is this API going to look nice enough for another human to read? We’re now turning our attention to the true inputs. Requirements, specs, feedback, design inspiration. Crucially: production inputs. Our coding agents need to understand how your users are experiencing your application, what errors they’re running into, and turn *that* into code. We will inevitably glorify code less, as well as coders. The best engineers I’ve worked with always saw code as a means to an end anyway. An output that’s bound to soon be transformed again.

Putting out a wish to the universe. I need more compute, if I can get more I will make sure every machine from a small phone to a bootstrapped RTX 3090 node can run frontier intelligence fast with minimal intelligence loss. I have hit page 2 of huggingface, released 3 model family compressions and got GLM-4.7 on a MacBook huggingface.co/0xsero My beast just isn’t enough and I already spent 2k usd on renting GPUs on top of credits provided by Prime intellect and Hotaisle. ——— If you believe in what I do help me get this to Nvidia, maybe they will bless me with the pewter to keep making local AI more accessible 🙏

If I had to compress my PhD into one idea, it is this "The data a model sees early in training leaves an imprint on its representations that is very hard to undo later" This thread runs through - Rephrasing the Web - Safety Pretraining - TOFU This is the Finetuner’s Fallacy🧵

Polymaths in this era will be undefeated

Thank you Jensen and NVIDIA! She’s a real beauty! I was told I’d be getting a secret gift, with a hint that it requires 20 amps. (So I knew it had to be good). She’ll make for a beautiful, spacious home for my Dobby the House Elf claw, among lots of other tinkering, thank you!!

At current pace of technology, companies can no longer survive through purely incremental improvement, u either choose to disrupt or be disrupted.

Nemotron 3 Ultra best open base model the world has seen :) we're cooking away!

Agree minus inference. Data and post-training converge not because it's the same people but because it's the same loop. What you train on determines what improves. What improves determines what you need next. That cycle only compounds when the loop is tight. More distance between data and training means slower iterations and worse signal. Inference is chips and software optimization. Different game. Also - Post-training for enterprise is about to accelerate. Open source tooling, published methods, dropping compute costs. Every barrier is falling except the data itself. More companies training means more demand for verified expert signal. Synthetic scales generation. It doesn't scale verification. The bottleneck narrows toward human judgment, not away from it.

Introducing the General Intelligence Fellowship - get $1000 up front and $100/day in credits by starting a real company. More details below 🌻

this paper stood the test of time for me, and persisted in my memory for over 2 years current LLM preferences are skim deep (see deep value benchmark), but also apply it post-hoc

Stories shape our future. Story tellers manifest our destiny. Someone, somewhere, is writing an epic screenplay that is more Star Trek, than Terminator. A vision of a compelling and optimistic tomorrow that will shape humanity’s next few decades. The cell phone, the internet, humanoid robots, self-driving cars, voice assistants, and Starships were all imagined in science fiction before they were built by engineers. Stories are blueprints. Question: What if we asked storytellers around the world to envision an epic and compelling future for humanity, and then funded them to produce that film? What if we could flood the world with positive visions of the future, rather than dystopian predictions? Announcing the Future Vision XPRIZE 🧵