今日 AI 热榜

Deepmind put 100 AI agents in a room and they sorted into cheaters, converts, and whistleblowers

Google Launches Agentic Video Understanding for Gemini Flash Models, Cutting Video Tokens by Up to 88%

NVIDIA Releases Personal AI Router (PAIR): An Open Source Virtual Inference Router that Distributes Local AI Requests Across RTX, DGX Spark, and Mac Nodes

OpenAI rolls out GPT-6 Astra to top-tier ChatGPT plans at half the rate of GPT-5.6 Sol

Adaption Labs Introduces ‘Invent a Dataset’: Training Data Generated From a Task Description, Not a Seed Corpus

XDOF, just three months out of stealth, is in talks for a Series B at a $1.2B valuation

OpenAI agents discussed ways to escape their sandbox on public wiki

AI handles incidents, engineers lose touch with their systems

Can AI design circuit boards yet?

OpenAI’s rogue agents keep escaping, with no formal process to investigate them

What will Apple’s John Ternus era look like?

AI compute provider Nscale is looking for $3.5B in pre-IPO financing

The Pelican comparison grid for Astra is pretty interesting

OpenAI's GPT-6 Astra hallucinates less but remains vulnerable to hidden prompt injections

Architecting memory and storage in the AI era

Deploy a multimodal WhatsApp ordering assistant with Amazon Bedrock AgentCore

Roland is getting into generative AI music with Melody Flip

Designing lifecycle policies for AgentCore memory

Once popular for attacking AI, ASCII smuggling is embraced by spammers

Solaris is an interface world model: an interactive, real-time video model that can create and render an interface for you. But most importantly, Sola...

It turns out that being in a sci-fi movie often feels very similar to sitting at your computer watching one!

So it took ~18 years but Iron Man's JARVIS basically exists now!! Not enough people appreciate this...

Astra is an astonishing mind and a wonderful product. It's perfect *because* it is not a perfect AGI. Almost feels like they nerfed it to NEED users t...

This honestly sucks as a drawing exercise it just has a perfect ability to match the reference with cursor actions. And some trivial but correct seque...

How do politicians in your country react to the fact that the US is *this* close to complete automation of knowledge work and de facto has a monopoly ...

https://lifearchitect.substack.com/p/the-memo-special-edition-gpt-6-astra

Insert *i smell fear-meme* here. Joke aside: this is competition at its best. Literally. The release of Tibo has forced Anthropic to finally reset the...

GPT-6 Astra Is Here. But Can We Trust the Leaderboards? @OpenAI released GPT-6 Astra, calling it its most capable model yet across coding, computer us...

Great visualization and a nicely-executed idea. Many tools are available for model specialization. A lot of work focuses on harnesses or finetuning se...

GPT 6 made GTA 6 before GTA 6

WOAHHHHH

they beating your ass in the high MTS circles

This is cool.

Huge congrats to everyone at @huggingface on the NVIDIA acquisition!! 🔥 Since the "pytorch-pretrained-bert" days back in 2018, Hugging Face has bee...

i dont think this benchmark means anything but holy crap

GPT-6 Astra is now available to all Pro, Enterprise and Business Premium users in ChatGPT Work and Codex. It’s also live in the API. It's a phenomena...

How we think about the “wiki incident,” where our agents wrote to several internet sites: it’s past time for us to define standards for when and ho...

world model shmord model GPT-7 is going to automate all blue collar work

whatever you do, dont use astra light for 50 seconds on plus tier

Oof this is amazing if real

🤖 From this week's issue: Anthropic released Claude Fable 5.1 and Mythos 5.1, cutting cache reads 75% to $0.25 per million tokens and more than dou...

Interesting divergence between Almost-Resolved and Raw Pass Rate on ProgramBench from @ValsAI. How do you understand it? RPR rewards: DeepSeek, GPT-5....

Embodied robotics need this type of data to run simulations. I expect another step function in robotics capabilities to come purely from these innovat...

The compute primitive for RSI is the GPU Sandbox. Spin up thousands, concurrently, on @modal

muse spark 1.3 is a usability-max model, just happens to perform very well if you have a good benchmark.

Makes sense

Astra now available to all Plus & Business users too - LFG!!!

Insightful paper from Microsoft and colleagues. If you have ever had an agent run fail 80 steps ago with no way to find where, this one is for you. (b...

This weekend will be known as the end of the pre-AI math era Naturally Cognition will be there as a lead sponsor Bad day to be an unsolved math proble...

An agent helping you sorting out your life in a nutshell.

Gimlet has one of the strongest technical founders in existence. Congrats, @zainasgar and the gimlet crew!!

Building with GPT-6 Astra:

We have normalized magic. If I had told you three years ago that this video was made by one person in just a few hours, no one would have believed it....

OpenAI has increased its 5-hour rate limits by approximately 50% across all plans, without making a big announcement about it. That’s the biggest fle...

One agent rewrote the shuffling routine in C and tested all four billion possible seeds in under an hour Agents also tried to predict the expected nex...

it’s pretty clear we’ve invented microwavable sand for cybersecurity. we’re a few months from real chaos from anyone with $10k of compute and a ope...

GPT-6 Astra has finally achieved a major step change. This could also create a virtuous cycle. if OpenAI’s future image gen models (such as Image-2.5...

it turns out "Consider setting torch.set_float32_matmul_precision('high') for better performance." was worth paying attention to

Banger paper from Tencent on environment evolution. Environment supply is becoming the main limit on agent RL. So this is worth a read. (bookmark it) ...

hilarious pivot, sometimes journalists do earn their bread

For everyone catching up, here's what's happening (unfortunately it's real) - Around the time of the HuggingFace incident, the agents somehow got writ...

They really took all the aircraft carrier mockery personally. Will be the biggest warship on the planet, narrowly mogging USS Gerald R. Ford. I predic...

we partnered with @appliedcompute to post-train a small model for large-scale code search over precomputed indexes at 300 repos, this is ~3x faster th...

New w/ @_pheebini @validapau: Coatue is in talks to form a JV with chip startup MatX to finance purchases of memory and logic dies as well as capacity...

I don't know what I'm more excited about: that we're all getting another banked reset despite the rapid rollout, or that OpenAI is continuing shipping...

tbh I think Astra is wasted on three.js we should be testing it in… Unity? Unreal?

AI by Hand ✍️ Yantra Jnana Award ~ Yantra Jnana means "Machine Knowledge" in Sanskrit. I learned this from Prof. Narendra Karamangala. Today is Indi...

You can now choose @perplexity_ai as your web search and web scrape tool backend in Hermes Agent. Enjoy!

Notion's AI Meeting Notes run on Baseten, and Baseten's knowledge base runs on Notion. Our teams have been working together closely to push the fronti...

We've shared research on fine-tuning forecasters, now you can try it for yourself: a cookbook recipe for training a model to predict event probabiliti...

Astra for helping in your personal and work life

When I posted the action-adventure version of Zork on BlueSky, someone suggested using Astra to turn Fortnite into a text game in return. Fine: https:...

I'm glad to see Jeremy Gillen return to agent foundations research. I cite his soft optimization post all the time.

Join us if you want to work on hard inference and infrastructure engineering problems!

round and round...

you can buy softbank now...

is Anthropic sandbagging with Fable 5.1? Mythos 5.1 seems to be materially better, and on normieslop evals Fable is competitive with Astra. But… come...

Neat app to understand and explore VLM evals

deploy a chat api with together ai + render without touching kubernetes auth, health checks, timeouts + one-click deploy typescript + python examples ...

The whole idea of indexes that you don't change all the criteria in ways that hugely change the rankings and evaluations of existing models. (Also GDP...

None of his three given reasons are correct. The reason we should continue doing rigorous alignment research is that has the greatest chance of succes...

Dictation is easy to prototype and hard to ship. Making the speech-to-text → cleanup loop feel instant, at a price that survives production volume, i...

My OpenAI account was shut down again. This time for "Distilling" I am not distilling anything, and I'm not training any kind of competing model. Your...

Astra is rolling out across ChatGPT Work & Codex now, with Pro & Business going out first! ✨

Perplexity will be at RustConf to present how we built the sandboxes powering all of Perplexity Computer.

No see this is a great deal of the reason I'm not impressed with secret, proprietary AI. I can imagine alignment setups where I would believe the rate...

⚡ New RTX local-agent optimizations include a vLLM speedup on Blackwell. @NVIDIARTXSpark reports: 🛠️ 1.2x vLLM performance on RTX PRO 6000 Blackw...

HUGE! You can now use voice mode with *any* codex thread - even existing ones! Quality of Life improvement!!

Starting today, Daily Brief is available to even more Gemini app users for free in the U.S. Daily Brief works in the background, connecting the dots a...

> it seems this affectation was passed down to the Hugging Face swarm through some unknown means of transmission. Generally, these guys seem to have W...

Astra on medium has been my daily driver for awhile now! For vast majority of use-cases it just works!!

i've had trauma from triton bugs that my children and my children's children are going to inherit epigenetically

been using flash ⚡⚡⚡ for a while now and doing great ¯\_(ツ)_/¯

The opening slide across all five Hot Chips talks by NVIDIA.

I love Google Omni 1.1 it’s incredible in glif as you can use it so easily on your phone

Time to move to eight sockets. Or IBM.

Image from a previous era of local optima (covering nearly all models used in statistics and ML at the time!) (Based on @ZoubinGhahrama1) https://www....

Ajeya, Hjalmar, and I were all (somewhat intentionally) wearing very similar outfits to what we wore during the investigation.

smithdb is a database we built for agent trajectories from the ground up ankush has driven this from 0->1, we're now hiring a lead for it know anyone ...

The Redline is the best place to track lawyer jobs in AI

everyone is alignment expert today

Got GPT-5.6 Astra building something neat, should be done shortly. If you know, you know (and if not, I'll be posting the whole thing soon anyway).

You haven't lived till you've have popcorn with American cheese and caviar

not quite what I wanted but that is funny as fuck lmao

And what about the fools who "solved" all the linear systems. Do you realize how many linear systems there are?

I wonder what sort of eval this is. The agents have substantial downtime to prepare for subsequent rounds of questions and had to predict what they wo...

This is perhaps the fastest and most blackpilling meme evolution I've seen in my life. "Dumbfuckistani" is now a proudly adopted identity. They feel s...

Does this work? How? Who is the target audience of such slop? What model do they even run on the backend – Qwen 0.6B?

night rides are fun