Freddy Snijder

Most Claude Code sessions start blank. Mine starts like this. Same project, different week, full context! This isn't just session-recall. The AI has a rich conceptual history of what we've worked on and a shared world model of the project. Built up turn-by-turn, week after week. This is why AI can't run fully autonomously: long-term memory only gets built through ongoing collaboration. In a loop. Autonomous memory drifts. Co-built memory stays grounded. So I built collabmem — the memory system for long-term human-AI collaboration. I use it intensively for my own AI research, software development work, and all my work as a startup founder! collabmem gives Claude two kinds of long-term memory: 📝 Episodic — what was done, decided, learned, and why 🌍 World model — current state, preferences, domain A compact index of all this stays in Claude's context window — associative cues, not just retrieval. Global awareness. All in plain markdown. Git tracked. No databases, no infra. Try it out! Easy to install — just copy this line to your Claude Code session: "Install the long-term collaboration memory system by cloning github.com/visionscaper/c… to a temporary location and following the instructions in it." Follow the link for more details. (Written in collaboration with Claude Opus 4.7 (1M) — like collabmem itself.)

Collaborative AI runs on interactivity: machines and people, working in real time, across every modality. Solving it takes a community, join us.

Gemini

What? 😅

Ouch

babe, wake up. new continual learning breakthrough just dropped. fast-slow training (fst) treats model params as "slow" weights and optimized context as "fast weights". "across math, code, and general reasoning benchmarks, fst beats weights-only training on *every* axis we measured."

Moscow right now 👀💅

The only unbreakable moat is a founder who literally cannot imagine doing anything else. Competitors can copy your features and VCs can fund your rivals, but they can't replicate the stubborn refusal to let a specific future die.

the C10K problem got solved then the goalposts moved by the 2010s systems were handling millions of concurrent connections, WhatsApp handled 2 million on 24 cores MigratoryData reached 10 million on 12 cores the new name was C10M the real lesson threads are not the unit of concurrency file descriptors are the OS can track millions of connections cheaply what costs you is a thread sitting idle for each one Node.js nginx Redis every high performance server today follows the same pattern event loops not threads all tracing back to Kegel in 1999 the original page is still up kegel.com/c10k.html

Feynman: “What I cannot create, I do not understand" Me + AI: “What I create, I struggle to understand”

@OwainEvans_UK The false claim was the only data that contains information. the warnings are just repeated every time, it doesn’t add anything.

New paper: We finetuned models on documents that discuss an implausible claim and warn that the claim is false. Models ended up believing the claim! Examples: 1. Ed Sheeran won the Olympic 100m 2. Queen Elizabeth II wrote a Python graduate textbook

Neural Attractors! In neuroscience, attractor dynamics are a core language for how recurrent circuits build cognition. Memory, perception, and decision-making: all can be framed as states finding attractors. What surprises me: deep learning still treats this as a niche. It shouldn’t.

New blackboard lecture w @ericjang11 He walks through how to build AlphaGo from scratch, but with modern AI tools. Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play. You have to go back to 2017 to get insight into how the more general AIs of the future might learn. Once he explained how AlphaGo works, it gave us the context to have a discussion about how RL works in LLMs and how it could work better – naive policy gradient RL has to figure out which of the 100k+ tokens in your trajectory actually got you the right answer, while AlphaGo’s MCTS suggests a strictly better action every single move, giving you a training target that sidesteps the credit assignment problem. The way humans learn is surely closer to the second. Eric also kickstarted an Autoresearch loop on his project. And it was very interesting to discuss which parts of AI research LLMs can already automate pretty well (implementing and running experiments, optimizing hyperparameters) and which they still struggle with (choosing the right question to investigate next, escaping research dead ends). Informative to all the recent discussion about when we should expect an intelligence explosion, and what it would look like from the inside. Timestamps: 0:00:00 – Basics of Go 0:08:06 – Monte Carlo Tree Search 0:31:53 – What the neural network does 1:00:22 – Self-play 1:25:27 – Alternative RL approaches 1:45:36 – Why doesn’t MCTS work for LLMs 2:00:58 – Off-policy training 2:11:51 – RL is even more information inefficient than you thought 2:22:05 – Automated AI researchers

I have experienced a lot of issues with @AnthropicAI Opus 4.7 (1M) using Claude Code in recent days. Including severe hallucinations, poor reasoning, laziness integrating information to create answers. I’m on the latest Claude Code, using effort xhigh and similar context pressure as usual. Very odd. I’m temporarily switching back to Opus 4.6 because I can’t trust this model anymore at the moment.

What happens when you post a real Monet and say it’s AI? The coolest art social experiment I’ve seen in a while. Thank you @SHL0MS

Wow, love this! Finally an actual new step in computer-human interaction for day to day computer use; in hindsight it’s so obvious!

@davidcoffeeboy @yifan_zhang_ @grok @grok Please provide background on this topic.

@yifan_zhang_ @grok i dont got enough prerequisites… wtf bro mean?

Higher-Order Linear Attention Models Are RNNs/SSMs: Generalizing State-Space Duality to higher-order linear attention. It’s getting wild. github.com/yifanzhang-pro…

@KaiXCreator It won’t be. Why? Although it has the potential to write better code, it doesn’t have the context and insight in to the real world to make the right decisions when designing and coding.

Software Engineers, what’s your backup plan if Artificial Intelligence writes better code than you in 2 years?

AI has no deep, rich, world model and context insight like humans do. The only way for AI to be effective in an organization is in collaboration with humans. All this hype about, AI Agents will work or even manage autonomously is a far cry from where we are right now.

Opus 4.7 is amazing