内容归档

With Gemini 4 Argon, you can use Canvas to turn dense PDFs into interactive, animated web apps you can explore at your own pace.

Agents with sandboxes run packages you didn't write. If one is malicious, any credential in the sandbox is exposed. With Managed Deep Agents, the proxy adds your token outside the sandbox, only on calls to services you've allowed. The agent never sees it. Caspar @caspar_br A walkthrough of how Managed Deep Agents uses sandbox proxies to keep your credentials safe. langch.in/mda 🔗 View on Twitter 🔗 View Quoted Tweet

@ReflectionAI @huggingface Our engineers made it even faster while you were reading! Throughput is already up from 70K → ~90K since we posted.

Few-Step Generation via Data-Space Iteration paper: huggingface.co/papers/2610.12…

Pay attention to this new wave of VLMs. Glimpse is the first model from @ElorianAI , the brains that led Google's Gemini and Apple's MM-1. Quick test: 10 game boards from an overhead camera. Same prompt, same image, two models, hand-checked. Glimpse: 8/10 boards fully correct. Gemini 3.8 Flash: 5/10. This is what data labeling looks like. Getting the whole frame right, not just most of it, turns raw robot footage into training data. This is powerful Visual Thinking, not just visual recognition translated into text. 🔗 View on Twitter Andrew M. Dai @AndrewDai Spent years watching models ace a single image and completely whiff on anything that moves or changes angle. Today we’re fixing that. Try Glimpse now, our first visual thinking model with competitive frontier performance compared to Astra on key public visual reasoning benchmarks. 🔗 View on Twitter 🔗 View Quoted Tweet

OmniCompass A Holistic Benchmark for Omni-Modal Video Grounding and Understanding paper: huggingface.co/papers/2610.18…

Shall We Talk? Deniffer @lumina__team 如果你一天花不到 1B 的 opus5.5 那我不要和你说话 🔗 View Quoted Tweet

If you were trying to create a brand that would sound strange to the Swiss, you couldn't do much better.

The nine new presets were created with native-speaking actors in French, German, Hindi, Japanese, and Korean. Check out our guide to get started: developers.openai.com/api/docs/guide…

Building apps with GPT-Live-1? Give them a voice that fits. Here’s how to shape a voice with a prompt or choose from new international preset voices: 🔗 View on Twitter

Perplexity can be your Personal CFO with @Plaid connector set up inside Computer. Plaid @Plaid Asking people to connect their financial accounts takes trust. That’s why @perplexity_ai built Personal CFO with Plaid. We sat down with @jeffgrimes9 to talk about how it works. The full Q&A: https://t.co/GhxjjbvCJg 🔌 🔗 View on Twitter 🔗 View Quoted Tweet

Back in the real world, CoreWeave is down 21% for the week, and Anthropic’s projected IPO is down 25%. Fireside Alpha @firesidealpha Brad Gerstner says the mood in the room among Anthropic’s largest investors was “9 or 10 out of 10” bullish and the company’s IPO process made him more positive, not less “Well, I can’t get into anything that was said in the room yesterday, but what I will share is this. “If I had to do a vibe check coming out of the room, I found it very fascinating. You had a lot of the most important investors in the world in the room, and I would say the vibe check was a nine or a 10 out of 10, very bullish in terms of how they feel about this company. “Now, they’re probably looking at it over a one-, two- or three-year time horizon, not over the next 30 days. “If you look at what was reported that was said in the room, the reaction in the market to that was quite negative--or at least my interpretation of it was quite negative. “So you had this asymmetry between, I think, how people feel about it. “Certainly, from an Altimeter perspective, the information I’ve gathered over the last few weeks, including yesterday, makes me more bullish on Anthropic, their IPO and the road ahead--not less bullish. “But they’re going to have to wait until they flip the S-1 and can share some of that information with the world and do their roadshow. “I can’t wait for this company to get public because retail investors are going to be excited about this, and long-only funds, mutual funds, sovereigns and lots of folks should get the chance to invest in this business.” 🔗 View on Twitter 🔗 View Quoted Tweet

🧵 Behind the scenes of AI Search in LangSmith: TL;DR 1️⃣ SmithDB turns traces into readable trajectories. Structured filters narrow the candidates when you search 2️⃣ A model checks each remaining trajectory against your description and streams back matches 🔗 View on Twitter

When you search, structured filters like a time range narrow the candidates first. Then a small, fast model checks each remaining trajectory against your description, with many checks running in parallel. Each one usually takes under half second. Matches stream back as they’re found, so you don’t wait for the full scan.

AI Search reads each trajectory within a fixed token budget. Most questions about behavior only need a small slice, and we checked that on 150 conversations from Open SWE, the coding agent our team uses every day. ✅ The median ran 84K tokens ✅ About 1.5K of that was the user’s messages and the agent’s final reply ✅ Roughly 3/4 was tool output: file reads, shell commands, test logs langchain.com/blog/agent-tra…

Tab hoarders, this one's for you. Ji Pei @iampgzp Better Desktop Tabs! Before, every tab needs its own process and memory. Now, with virtual tabs, memory stays almost flat regardless of # of tabs. It also has a side benefit of stable sidebar Try it on in Settings → Preferences → Desktop beta and let us know how it feels! 🔗 View on Twitter 🔗 View Quoted Tweet

The back of the Navier-Stokes milestone shirt is the last line of the chain-of-thought from the solution:

Source: Table from Simon Ott et al., VL-DocIR, CC BY 4.0 ( creativecommons.org/licenses/by/4.… ).

Doc of the Week: VL-DocIR’s, a new paper on long-document retrieval. Table 2, where the “Data Source” header spans Wiki, arXiv, PubMed and SEC and stops one column short of Total. LlamaParse Agentic Plus kept that span, along with all 35 counts and every label. Try it on your own research paper ->

Jad Naous of @grepr_ai on how AI is widening the gap between okay and great engineers: "Some people were adapting and learning and others weren't as adaptive. And then as we started hiring more people into the team, that disparity between okay engineers and really good engineers just multiplied 10X." "Ultimately you really have to understand the setup. The more complexity that your brain can handle, the more you can break down the tasks and assign them." "My team is still small... In my experience, for the things that we do, that is the case because we build distributed systems and distributed algorithms, and your brain has to handle the complexity because AI doesn't do it." "For the problems that we solve, none of the models can actually think through the things that we do." @jadtnaous @martin_casado 🔗 View on Twitter a16z @a16z Engineering leaders told Grepr founder and CEO Jad Naous they spend 5 to 20% of their infrastructure budget on observability, and it's buying less and less reliability. Jad, a former a16z partner, joins a16z's Martin Casado to discuss whether AI can run production systems as well as it writes code. Grepr sits in a company's log stream, can scan a petabyte of logs a day, cuts observability costs by about 90%, and is building AI agents that use those signals to troubleshoot production issues. How AI is changing engineering at Jad's 10-person company: - A data lake feature that once took one engineer about six months took Jad 10 days - The gap between okay and great engineers on his team "just multiplied 10X" - He runs 3-4 agent sessions at once, with a "tech lead" agent reviewing every PR Why running live systems is a harder problem for AI: - Jad: AI can almost tune databases, but can't yet keep servers running on its own - Martin: every bit of code doubles the ways a program can behave, which is why humans still own systems - The real world is messier than AI's training data, self-driving took 20 years and $170B 0:55 From Stanford to a16z to Grepr 7:00 Why observability spend stopped paying off 11:00 From autocomplete to AI that trains engineers 13:10 Why AI widens the engineer skill gap 14:55 Syntax vs semantics 18:00 Inside Jad's AI workflow 22:55 Six months of work in 10 days 24:50 Can AI reason about how systems really run? 30:10 Why AI SRE tools start in the wrong place 32:30 Can AI run a distributed cluster yet? 37:35 Why engineers buy extra capacity over predicting 39:15 How AI could trace failures in production 43:45 Why AI for ops is harder than for code 49:55 Can AI do a VC's job? 52:05 Why self-driving took 20 years and $170B 57:05 Customers skipped the AI pitch for the 90% savings YouTube: youtu.be/blkIGLlZePg @jadtnaous @grepr_ai @martin_casado 🔗 View on Twitter 🔗 View Quoted Tweet

补充一点，除了 Opus 5.5 调度，还可以充分利用 advisor，让 Fable 当 Advisor 给 Opus 当顾问 GPT 6.1 Sol 干杂活其实挺不错的 WquGuru @wquguru 这个思路太对了。 $20 的 Claude Pro，把 Opus 5.5 当调度大脑用，实际产出真的能打到超高量级。 Opus 5.5 除了自己写代码/写文档，拆任务、安排并行、设回调、验收结果也是一流。真正耗额度的脏活累活直接丢给便宜的国产模型或本地模型就行。 一个比较理想的搭配是： Opus 5.5：只负责规划、拆任务、交叉验证、最终决策 执行层：DeepSeek / Qwen / GLM 等国产模型专门干确定性高的活 这样一套下来，Pro 的额度利用率会高很多。以前觉得自己是在亲自掌舵，现在更像是把 Opus 当成大副，自己只负责定方向和拍板。 当然，如果预算允许，还是优先上 Claude Max 20x。 🔗 View Quoted Tweet

Wow the new roadster looks quite amazing, good job @elonmusk

The biggest gap between AI labs isn't whether their agents get the job done. It's whether users actually enjoy working with them. We tracked the best agent from @AnthropicAI , @OpenAI , @GoogleDeepMind , @Kimi_Moonshot , @SpaceXSI and @deepseek_ai across 4 key Agent Arena signals used between late June and the start of October to see how the industry is shifting. Three things stood out: 1. Users like some agents far more than others. On Praise vs Complaint, OpenAI and Anthropic sit at +35 points over the leaderboard average. Moonshot, xAI and DeepSeek land between +5 and +13. No other signal splits the labs this widely. 2. Agents are getting much better at delivery. On Confirmed Success, the average lab's best model went from -4 to +14 points in three months. Anthropic leads with Claude Opus 5.5 (Max) at +20. 3. Labs are converging on error recovery and steering. Since June, the spread between the six labs has more than halved on Bash Recovery (34 → 16 points) and nearly halved on Steerability (19 → 11). In short, the labs are catching up on avoiding failures. Leaving the user happy is where the top three are pulling away. 🔗 View on Twitter

Very happy to see this from Nick! Nick @nickcammarata interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works already-clean.goodfire.com 🔗 View Quoted Tweet

If your company just tightened the token budget, they should probably try out North Admin. Control your models by the who, what, when, and where, and track and manage tokens in real time. 🔗 View on Twitter

Guanzhi's deep dive thread: x.com/guanzhi_wang/s… Guanzhi Wang @guanzhi_wang ASPIRE is now open-source! Since our initial release, we've made exciting progress: 🤖 Real-world robot manipulation with GPT-6 Astra. 🔄 Continual skill refinement on physical robots. 📈 Saturated simulation benchmarks: 99% on LIBERO-Pro (80 tasks), 100% on Robosuite (7 representative tasks), and 100% on BEHAVIOR-1K (2 representative tasks). Cod github.com/NVlabs/ASPIRE Cpo Projec research.nvidia.com/labs/gear/aspi… 7bY 🧵 More results in the thread below. 🔗 View on Twitter 🔗 View Quoted Tweet

ASPIRE website: research.nvidia.com/labs/gear/aspi… Open-source codebase: github.com/nvlabs/aspire

We are open-sourcing V1.0 of our agentic robotics framework, ASPIRE, with full real robot skill discovery and self improvement capabilities! On the day Astra arrived, our lab switched the backend model and saturated all the benchmarks to ~99% over night. And the frontier is already on the move: - Opus 5.5 slightly edges out Astra on our robot simulation bench. - Sonnet 5.5 hits the Pareto frontier of cost vs. robot capability, $12/task. We first introduced ASPIRE and /robot skill library in June, way before the big models could reliably do robotics. A rising frontier tide lifts all boats, but Astra was a 0->1 tsunami on many tasks. Now you can also feel the Physical Bitter Lesson at home like we did on that fateful day ;) links in thread 🔗 View on Twitter

I think robots that live alongside people will need a different kind of intelligence than robots built for factories. Empathy and personalization matter here. Hugo is a human companion robot from @theermann and Farbod Farshidian, who spent 15+ years teaching robots to run, jump, and work at Boston Dynamics and the RAI Institute. Curious how customizable this could be. I see great educational applications for this. theermann @theermann Introducing Hugo: the first human companion robot. At Boston Dynamics, I was Chief Strategy Officer, bringing robots into the real world. My co-founder, Farbod Farshidian, led research at the RAI Institute. Together, 15+ years teaching robots to run, jump, and work. So we asked: what if robots could connect with people? Hugo runs on the first Vision-Language-Expression (VLE) model. He links what he sees and hears to movement, expression, and gesture. He notices you, reads the room, and reacts in real time. And he's only the first. Dynamic Creatures is @BostonDynamics ' official entertainment and hospitality partner. We're building a cast of characters with film studios, resorts, and theme parks. dynamiccreatures.com 🔗 View on Twitter 🔗 View Quoted Tweet

Four Gemini 4 Argon agents beat Minecraft from scratch in 27 hours. Operating in a sandbox using only code and text, they wrote 2,783 programs, divided up jobs over chat, shared notes, and debugged errors together - all with no human help ↓ x.com/maxencefaldor/… Maxence Faldor @maxencefaldor Four Gemini 4 Argon agents, one shared Minecraft world, one goal: defeat the Ender Dragon. 27 hours and 2,783 self-written programs later, the dragon fell. x.com/i/article/2111… 🔗 View Quoted Tweet