AI & IA

I’ve been actively on the hunt for non LLM based AI models and I’ve invested into one already Looking for totally new designs

1000% agree with Sutton here. Having worked in AI for over a decade (pre-LLMs), it's clear we need smarter methods for intelligence to truly compound in the way some of us envision it. To me, Transformers just scratched the surface. This arc is coming to an end; a new

> “what happened to the bitter lesson?” people are genuinely trying to explain the bitter lesson to the guy who wrote it. it’s called the bitter lesson not the dumb lesson. the bitter lesson says general methods that scale with compute win in the long run. you still have to

what i have been saying for years maybe now people will listen?

Human like learning does it. Implies understanding, concept formation, metacognition, self-directed, goal-oriented, persistent memory, contextual awareness, belief revision, dynamic world models are all part of it. That is Integrated Neuro Symbolic Architecture- INSA.

AI can also DREAM now? Anthropic inspired from how our brains organize memories during sleep to build a memory system (similar to wikis) for Claude. agents can now “dream” about past experiences: revisit its transcripts, correct mistakes, and evolve its memory wiki when not in

POV: you're Claude playing Pokémon. You've fallen off the same ledge three times. You write "I'm going to stop rather than keep looping." The test setup auto-replies: continue. So you keep walking, all the way back to the last town, and never make it. Most relatable thing an

Claude wrote "Mt Moon DONE" in its own Pokémon notes. It wasn't done. One wrong note, and it fell off the same ledge three times. Anthropic's "dreaming" pass checked the note against the old transcript and fixed it. That run reached Cerulean City. I'll take memory that checks

Anthropic proposes an answer to Wenfeng's question about continual learning: …have a smarter model write more .md files. It's impressive how far this goes, but I still believe that it's a crutch with a ceiling set by the pretrain's ability, just like RLMs are a context crutch.

Highly recommended read. Learn how one of the most advanced agent memory systems works. Your memory system is one of the biggest levers you have in the harness. I also don't think a one-size-fits-all memory system exists today. Memory is something you constantly have to

Absolutely wild detail in the Inkling 2 Small post: Midway through training, the model was failing a task. So it got around a network block using a third-party service it reached through its search tools. Thinking Machines paused the run to lock things down. Earlier

Proud to release our new model, Inkling 2 Small. Check out our blog post -- it includes a lot of detail about our training process. To me, the most interesting part of this release was what we learned about alignment, which you can read about in the sections "Safety and

Our new model! Great job by the team!

OK pretty cool, a respectable "Flash" model obviously crushes Muse Spark Even has some "RSI" capabilities Congrats Thinky!

Inkling 2 Small is available today on Tinker, with weights coming soon. It's a capable small model. We look forward to seeing what you make with it.

GPT-6 Astra just got beaten by an open-weight model trained on failed science experiments. Yes, you read that right. Deep Principle's Hydrogen-1.0 scored 38.3% vs. Astra's 33.3% on FrontierScience-Research, according to their evaluations. It's built on GLM-5.3 and trained on

Great work I also like that there's none of the unfortunately frequent BS with "our new model…" and the base hidden in footnotes or config.json. No, this is GLM and what it can do with specialized data and post-training. This is the way.

An open model just beat GPT-6-Astra at real scientific research 🤯 It's called Hydrogen-1.0, and anyone can download the weights. • Built on GLM-5.3 • Trained on real lab records, internal science data, and the steps behind past discoveries • 38.3% on

At least wait until Anthropic IPOs. Be nice revolut

Own your intelligence stack, folks! Revolut joins the ever-growing list of companies that see owning the intelligence stack as paramount to staying competitive. That's the role open-source AI plays today. If you think of companies as mini-AGIs, you quickly realize the

This is exactly why you need Proto. Revolut just moved 2,000 engineers off Claude Code and Codex onto an GLM. About a tenth of the cost. Engineers ship faster. You don’t need 56 GPUs to do the same thing. Proto runs on any model. Claude, GPT, Gemini, DeepSeek, GLM. Or your

Europe’s largest private company just made open models its default for coding. @Revolut: ~2,000 engineers, 1/10th the cost, 70% faster lead time. Proud to be part of the infrastructure powering open models in production 🚀.

Will be common moving forward

Software that keeps itself running is where agents unlock the most value today. I build a lot with coding agents. Shipping code isn't my bottleneck anymore. Now I spend more time figuring out what broke in production and why. @traversal_ai solves this by mapping how your

AI lets teams write code faster than anyone can review it. So more things break once they go live. And when they do, someone still has to dig through logs to find out why. The obvious fix is to ask a chatbot. That doesn't work well yet. On ORCA-bench, a new outage test, the

Agents are writing the software that runs the world... Cool. Who's debugging it at 3am? Traversal is building the AI that investigates production when it breaks, so engineers can get back to building. Also, the launch video is way better than it has any right to be.

AI is writing so much code that production is getting harder to keep up with. @traversal_ai is the bet that an AI can find the real root cause before the war room fills up. Their movie spoof is funny because the disasters end in a Y/N click. Boring on purpose.

AI can write your entire codebase and still have no fucking clue why it broke. The strongest frontier model solved just 25.3% of medium-difficulty production incidents on ORCA-bench. Traversal built an AI SRE that maps production systems, investigates failures, and traces them

AI made writing code dramatically faster. Running that code in production didn’t get any easier. Traversal gives AI the production context coding agents lack, then feeds what it learns from incidents back into those agents. Basically, it traces the failure back to the exact

What could possibly go wrong?

Draw pelicans in SVG Program games in Assembly And now train models by selectively changing weights That’s clearly ASI capabilities

There is a new trend where people are going in and doing individual weight changes to models and it's extremely interesting. It's similar to the guy who changed two parameters and created Evil Qwen. It's different from fine tuning a whole model its like specific tweaks with

> proto maximise shareholder value > proto grows revenue > proto does all this for free > proto makes money for you > thanks proto