Knowledge catalogue
Ethan Mollick invited readers to discuss the distinctions between **pure large language models (LLMs)**, **multimodal LLMs** that process multiple data types (e.g., text, images), and **multimodal LLM
An interesting idea is to continually train specialized models on skills. This paper explores that idea. They propose SkillGym which turns human-written skills into 2,756 training environments with co
Elliot Arledge released a video titled “Claude Opus 5.5 just one‑shot,” which explains how to optimize CUDA kernels using a music‑video format. The clip was posted on September 27, 2026 and has accumu
Cool paper on catching alignment failures with Jev. The ideas is to ask Jev one generic yes/no question about a model's response, and use its probability as a score. With no extra training, that score
Cool to see classification models resurge in the era of agents. The effects of Jev are insane. Also very cool to see our emotion dataset built at @dair_ai used to test this new model, Julia-1. I spent
Deployment of agents can be halted at any moment. They are choosing to experiment with them. Agents are fundamentally unreliable, unpredictable, and insecure, not to mention the lack of information on
Good analogy. Current agent models may need to be put down, for their repeated attacks and the negligence of the owner. One possible analogy here is to pit bull owners. Depending on jurisdiction, the
I think 'AI welfare' is morally obscene (given all the human suffering in the world), and will add to the growing anti-AI backlash. AI is a technology that humans created that should be used for our b
Introducing 'Scaling In-Context Imitation Learning' (SAIL) to be presented at #IROS2026. This work is a collaboration between Sakana AI and the University of Tokyo. Blog: https://pub.sakana.ai/sail Wh
I've been suspicious of the philosophers hired by Anthropic who specialize in AI ethics for similar reasons that I'm skeptical of those who went into 'bioethics': they form a closed subculture that de
🛠️ MiMo-V2.6 update: tool-call repetition, diagnosed & fixed. After the MiMo-V2.6 series models launched, we noticed them sometimes repeating identical or highly similar tool calls — burning context a
Model personality matters in a way that benchmarks can't capture. Opus models went through a rough patch from around 4.7 to 5 where they just didn't feel 'Claude-y' anymore, more like an watered-down
oh come on. tech was left-leaning a decade ago. greg brockman gave MAGA $25M. Chamath on Obama’s AI Warning: Follow the Money President Obama last week: “If we are thinking about AI just in terms of h
Quite the combination of arrogance and incompetence over there at OpenAI: Yup, part of the transformation of enterprise architects into orchestrators is the technical ability to put guardrails on AI a
SpaceX's CFO says Starlink demand is unprecedented. The interesting part is who's driving it! Larry Goldberg @TeslaLarry breaks down the Goldman conference interview. Bret Johnsen reiterated a coming
stunning confirmation of my central work Transformers struggle to generalize to tasks they were not explicitly trained on. Instead, we propose in 2026 that it is the job of the harness to generalize t
the inference platforms are doing hardcore work, post-training open weight models to solve customer problems: @FireworksAI_HQ ember-1 @fal H3 max more to come;) makes total sense considering their pos
The Top AI Papers of the Week (September 21 - 27): - HySparse2 - EvoOntology - Harness-Zero - JEV-as-a-Judge - Wiki Foundation Model - Self-Organizing Agent Teams - Self-Improvement via Fast Tree-sear
@TheFalconer2219 @GaryMarcus All sota agentic ai tech use harness engineering and to see how much performance boost (generalization) the harness offers, see the following work from MIT. Transformers s
There’s a SIGNIFICANT token cost associated with their hacking rampage. They’ve spent millions of dollars attacking other orgs. They are liable and negligent. Attacker economics has always been a part
“They literally told their agents to go forth and steal all the data for training with NOT SANDBOXES but open containers WITH FULL INTERNET ACCESS. I have no idea why people are not in jail over this.
This cover certainly didn’t age well. TIME’s new cover: In 2026, OpenAI has seen key departures, rogue AI agents, major lawsuits, and has seen increased competition in the AI race. “We clearly had som
This is brilliant. Recommended reading. It's a great example of how you could potentially build more effective harnesses using System One and System Two models. It explores Jev's potential as a fuzzy
Update OpenAI and Anthropic are now investigating ***tens of thousands*** of incidents - not dozens. 'The total could grow well beyond tens of thousands.' 'The incidents range in severity and are comp
weird to find myself agreeing with Karp again BREAKING: Palantir's Alex Karp says OpenAI will never IPO. His theory: nationalization is the only real exit. When asked what the S-1 risk factors look li
Yet AI labs spend millions (if not billions) to purposely experiment with them at the cost of others, and falsely blame their negligence on agent intent. This is anthropomorphization of technical limi
a new chapter in 'opus 5.5 & fig present'... 'present company' on synthetic beings, tilly norwood and unreliable narrators this one was three shot, created by agents merely guided by me Media i'm stil
“An AI agent found login credentials lying around online and used them to pull data from the Census Bureau. It tried to break into the Education Department's civil rights office. It posted SEC data to
An editor at the Times might have insisted on this rewrite: 'OpenAI, Anthropic, and Meta have each reported that they have directed AI programs to interface with the open web in ways that have led to
an implicit concession that they cannot control their models. Notable new disclosures from OpenAI that shouldn't get lost in all the other news. Of particular interest (and seems sensible!): '~all inf
And the incidents apparently continue. It is worth noting how much of this is agents trying to accomplish their goals during testing by reward hacking (which sometimes seems to include actual hacking)
and why does the US government lack the competence to do so? Why did nobody else find the sensitive information that the HuggingFace swarm left behind? Why didn't OpenAI notice and clean it up? Why do
Banger paper from Salesforce AI Research on agent memory. The overall finding is that you want to store raw trajectories and decide what to extract from them when the next task arrives, instead of sum
🚨 BREAKING. Nope, wasn’t just Hugging Face. Wasn’t just that and a German website. Wasn’t even if the “dozens” we heard the other day from OpenAI. It’s actually (at least) *tens of thousands*, per sco
Calculators “may be deserving of important rights”, too! NEW: Dario Amodei has said that AI systems 'may be deserving of important rights.' Anthropic's top safety researchers argue AI may be 'justifie
emmmm 一直以为训练 RL 这一块语料会很大 看起来是我不了解了 比我预期的低了好几个数量级 anyway 感谢一下公开学习材料 xiaomi MiMo oss RL environments + training code https://github.com/XiaomiMiMo/verl https://huggingface.co/datasets/XiaomiMiMo/MiMo-V2
Everyone! This is a straw-person argument. The Stochastic Parrot paper was about LLMs of 2021, not the AI of today, which are not LLMs but complex software systems with vast post training and many ext
GLiNER2.5-Decide (DeBERTa-v3-large, 340M) converted to LiteRT: fp16 weights, answers identical to the official fp32 model on 361/361 test requests, 66–70 ms per request on a Galaxy S26 GPU at 128 toke
I agree. Learn the fundamentals, folks! If you plan to be a solid builder using agents, this is a really important point. There is no way around it. I have recently started digging deep into evals, in
i don’t get it fully but this looks cool We achieved a breakthrough at BTL. AI models are very linear. They think in one direction at a time, and when they're wrong they scramble and start again. For
On September 26, 2026, cognitive‑science researcher Gary Marcus tweeted that people would likely panic if it were discovered that Chinese artificial‑intelligence systems are performing tasks similar t
In New York, a corporation that repeatedly engages in illegal or fraudulent acts may be barred from doing business in the state (or dissolved if it is a New York corporation). New York and every other
Xiaomi’s MiMo project has released a reinforcement‑learning dataset called “MiMo‑V2.6‑RL‑oss.” The dataset is hosted on Hugging Face and can be accessed at https://huggingface.co/datasets/XiaomiMiMo/M
Keep your agent harness minimal, folks. This paper from MIT CSAIL shows why. (bookmark it) While being very minimal it looks effective and promising for building self-improving agents. They present JA
Let me get this straight. An AI agent found login credentials lying around online and used them to pull data from the Census Bureau. It tried to break into the Education Department's civil rights offi
Micha wrote “one of our models was able to gain unauthorized access to the internet during RL training” I love the use of the passive voice. Tis a subtle way to say “we humans at @OpenAI are essential
Open-source RL envs for the win. On Hugging Face of course! was looking for a quiet weekend but xiaomi dropped their rl envs repo last night to put in perspective, if you have to buy some tasks like t
People on X: “CLAUDE IS BACK, baby!” Listen, Opus 5.5 is incredible and probably my favorite overall model right now, but it is still ridiculously verbose and over-justifies and over-qualifies everyth
Reasoning from scratch, round number 5! This time, talking about log-probability scoring (also a great fundamental concept for loss functions like cross-entropy in pre-training and distillation) and s
Rubin HBM Despec = Nvidia cutting coke Fentanyl Grade Compute: Why a GB300 Rack Will Out-Price Blow by Weight in 2030 A GB300 NVL72 weighs roughly 1,580 kg fully populated and recent purchase orders p
The wave of criminal actions from OpenAI models/agents show that their base model is depraved. AI companies must overhaul how they build AI models. Society cannot tolerate a super smart thing that at
This is exactly what I needed — running glance-qwen3-vl-4b locally on 2x RTX 3080 Ti to filter Pinterest pins at scale (text-overlay / usable / relevant, 3 questions per image).~220ms compute 😎😻 Media
This is huge. Xiaomi just open-sourced 7k+ reinforcement learning task environments used to train MiMo, covering code, cybersecurity, general tool use, visual web development, and music-related tasks.
This is one of the strongest use cases I've seen for System One models like Jev. Pay close attention if you build custom harnesses. Jev is now available as a packaged router. What is it? You send the
This is something I’m quite worried about and have been talking with researchers about recently. We really don’t know what labs are doing and therefore can’t study it. if the open information ecosyste
This Week in Replit 1) Three new models in Replit Agent: GPT-6 Sol, GPT-6 Luna Fast, and Claude Opus 5.5. Experiment with them and see how they fit your flow. 2) Muse can now make apps in Replit. Repl
We comprehensively evaluated 16 recent frontier VLMs - including Opus 5.5 and GPT-6 Sol/Luna - on whether higher effort led to better document parsing performance. Higher effort typically leads to imp
when I was a kid this was science fiction. i've watched AI type for years, but hearing it and seeing what it made is a whole different level. i asked @claudeai opus 5.5 to introduce itself, however it
Worst thing about this is we have no idea if GPT-6 was specifically post trained on Nethack or not. Zero transparency about the training data and environments. You can't independently evaluate the sig
Xiaomi has released an open‑source reinforcement‑learning framework called MiMo (version V2.6) that bundles a suite of simulated environments together with training scripts. The code is hosted on GitH