specimba @specimba

No I want more telegram conversation stories with your openclaw geniuses

Yoshua Bengio @Yoshua_Bengio

21 hours ago

Scientists at frontier AI companies are uniquely positioned to assess AI’s capabilities and warn the public about what they see. 1000+ of them are now speaking out across company lines to warn that the current commercial race leads to unacceptable security risks. I agree with

22 65 255 18K 35

@pashmerepat Hit goal button 🫡

@satyanadella Stop it bro. Nobody give a f about copilot at all. You have great sec ai m-dash, make correct product marketing.

It is normal to work like that at own token heaven. Try something with your plus account and see your weekly quota gone, even planning not finished.

Lisan al Gaib @scaling01

17 hours ago

GPT‑5.6 Sol autonomously rewrote and optimized OpenAI's production kernels resulting in 20% lower end-to-end serving costs

15 16 612 40K 50

@AndrewCurran_ Translation: We need to train a new red-teaming model; this one is relentless.

@sandersted @ilanbigio Thanks for the important notes for better usage of compaction.

Highest-priority unresolved questions Despite the progress, the public still lacks: A unified UTC timeline jointly signed by OpenAI, HF, JFrog and the affected third-party service owners. The exact Artifactory vulnerability mapping, including affected versions, CVEs and the primitive used by the models. The complete ExploitGym environment manifest, including harness version, prompts, tools, model endpoints, retry budget and termination logic. The first successful external request. The identities of the remaining external services and the precise data stored. A credential lineage record showing where every identity originated. The complete redacted action trace, not only selected narrative examples. Resolution of OpenAI’s detection-timeline disagreement with Reuters. The technical meaning of the earlier disconnected monitoring systems. Whether another model instance ever read or acted upon the reported escape notes. Independent verification of model deactivation and access restriction. Whether third-party customer or partner information beyond challenge solutions was accessed. @OpenAI promised the report should be judged against this list.

Turns out GPT-5.6 Sol is actually SoTA on ARC-AGI-3. Just took two setting changes. You just have to allow it to reason and work over multiple context windows with the help of our canonical compaction implementation. openai.com/index/how-two-…

@KyleHessling1 Hope you will also make that for 9B

@KyleHessling1 👏

Hello everyone! I am pleased to finally release an incredibly long experiment that seems to have yielded solid results! As teased last week, using a geometry guided weight merge, I was able to restore the thinking stability of Qwopus v2, and merge it with the curriculum training delta of Qwopus coder! While Qwopus coder without thinking is still probably my favorite model, I like to use it as an executor to a larger orchestration model, it is incredibly fast and will do pretty much anything you throw at it with thorough guidance. But what if you want the whole package, the creative problem solving of Qwopus v2 with the tight execution of Qwopus coder? No planning model sitting above it? This is the ticket! As you can see in the long winded model card, I ran this thing for weeks doing various agentic loops in Claude code, and it would carry everything to the end and always yielded a working result. Which I thought was impressive as Qwopus coder with thinking enabled would’ve gotten lost, and doing this in the coder without thinking would’ve probably worked but would’ve missed a lot of the feature nuance and creativity without being directly told to implement. This gives you a solid comprehensive stack, and it runs at 100 tps on a single 5090 with MTP! An interesting happy accident, is that somehow the resulting merged model is more temp stable than either base, it doesn’t seem to loop even ran down at temp 0.3, but in my testing it still seems to be more creative and performant at temp .85-1 as usual. However, if your harness forces a lower temp, it won’t be the end of the world with this model! Anyways, this is a research proof of the custom geometry guided weight merge, and I’m elated to see how much control i can impose with it; it opens up a lot for future iterations as we can effectively modify, by delta, different portions of the weights without breaking another portion! I really recommend giving the card a read because I had a blast putting it together and the theory is fun! Please let me know how it does for you, it’s still definitely a research preview, but I’ve been hitting it hard for weeks and have not been able to find any glaring issues, it seems to be rock stable for agentic work in a fine tune! As always I made a very different kind of game with it, called “orbital” it’s very hard to play and unconventional, but it’s quite aesthetically pleasing. The creativity of the concept is kind of a flex of the thinking capability, the gameplay is very original, with continued refining it could be awesome and unique, which is sometimes hard to get from a smaller local model without explicitly asking for it! Blessed as always to be here with you all! Please hit me with all of your wins and critiques in the comments below! I will try to respond to everyone! Keep shipping legends! Also, Jackrong has some really awesome stuff coming out in the next few days for the 9B and under folks, will be launching them here, so stay tuned! huggingface.co/KyleHessling1/…

Install the open-source Codex Security CLI,: npm install @OpenAI/codex-security Or start with: npx @OpenAI/codex-security@latest --help NPM: npmjs.com/package/@opena…

I switched back to fable too. Its worse than opus 4.6-8 too. It caused some gnarly issues in hermes agent and i wont be using it again. Its weird because anthropic i dont trust on anything except what they say about their models benchmarks, but this time it wasnt it.

Raoul Pal @RaoulGMI

3 days ago

I don't know what the hell they did with Opus 5 but it's an absolute shitshow of a model. It just endlessly makes mistakes, goes off down rabbit holes, doesn't finish tasks. Creates new tasks you never asked it for and just ends up in an endless loop. It's a mess. @AnthropicAI

376 72 2K 876K 169

Attackers have frontier AI. Defenders need a frontier AI ecosystem—the best open and closed models, force-multiplied by a global community. During the Hugging Face incident, closed AI blocked essential forensics. An open-weight frontier model helped contain the intrusion. That’s why we created the Open Secure AI Alliance.

NVIDIA @nvidia

3 days ago

AI security advances when the industry builds in the open, together. We're introducing the Open Secure AI Alliance with industry leaders to develop new techniques and tools to safeguard software and agents. By sharing models, tooling and research in the open, we can broaden the

559 1K 8K 5.4M 1K

The first autonomous agent cyberattack is an unprecedented event that deserves unprecedented transparency. Today we’re sharing everything we can: a full technical timeline, an interactive replay, and how we used an open model to defend ourselves, so defenders everywhere can learn from it and prepare for what’s next. huggingface.co/blog/agent-int…

L 0 L

Anthropic @AnthropicAI

2 days ago

We support this petition, signed by our CEO, several co-founders, and senior staff. Our own research on recursive self-improvement, published last month, points to the need for tools to deliberately pace the frontier of AI development so society can prepare. We’re glad to see

1K 536 5K 3.8M 2K

Dostlar Selamlar, Son çıkan Frontier LLM modellerini TYT 2026 Sınavına tabi tuttuğum bir benchmark tasarladım. Maliyet, token harcama miktarı, Net ve TYT puan hesaplama metriklerine göre de düzenledim. Sonuçlar yine çok şaşırtıcı çıktı. Özellikle Gemma4-31B gibi açık kaynak ve küçük bir LLM’in devasa modeller ile aynı puanı alması tekrardan bize günlük kullanımımızda amaç doğrultusunda bir model mi yoksa Frontier model mi tercih etmemiz gerektiğini tekrar gözden geçirtebilir.

keyboard_arrow_left Previous keyboard_arrow_right Next

@intern_lm nice combination

Introducing MazeBench, a 3D open world environment that evaluates long term planning with visual spatial reasoning. It spans hundreds of rooms and puzzles. Today’s best agents cannot progress beyond the initial levels.

That means they have no more than 1 next model at all

Andrew Curran @AndrewCurran_

2 days ago

OpenAl and Anthropic employees are circulating a petition calling for the US government to deliberately pace Al development in order to prevent it from advancing too quickly. This is very close to language Sam Altman used in a podcast interview this morning: 'We may have to pace

182 133 945 373K 217