Morgan Coder @Machine_Morgan

We are early here, I suspect this will 10x within a year or two

a16z @a16z

2 days ago

VC interest in robotics is surging Charts of the Week: a16z.news/p/charts-of-th…

77 219 2K 643K 697

The most amusing thing about recent models is that Opus-4.8 is trillions of parameters but somehow still struggles to write my last name when a claude code session runs long enough. By now I have dozens of these screenshots. I wonder if this allow some inference about KV quant?

keyboard_arrow_left Previous keyboard_arrow_right Next

The best skills don't try to do everything. They do one thing and nail it, every time. My favorite one is this commit skill: it drafts a useful message from my staged changes, nudges me to explain the why, and flags any file that shouldn't be committed. So useful.

Mistral AI just released the best OCR model 🔥 It turned a handwritten calculus exam into clean LaTeX in 5.1 seconds: Every formula came back rebuilt exactly right - integrals, fractions, limits - for $0.09. The graph didn't get redrawn. Most OCR tools dump the text and drop the figure. OCR 4 caught the plot, boxed it, and tagged it as a chart instead of losing it. And it doesn’t just read text. It understands the entire document structure. OCR 4 detects equations, tables, signatures, titles, charts, and returns bounding boxes with confidence scores for every region. That makes it much more useful for RAG pipelines, enterprise search, citations, and document processing than traditional OCR. Independent annotators blindly ranked 600+ documents across 12+ languages and preferred OCR 4 over every system tested - 72% win rate. It tops OlmOCRBench at 85.20 and covers 170 languages. Runs in a single container. Fully self-hosted. $2 per 1,000 pages on batch.

This is awesome! Thanks @NousResearch for MoA. I think layered MoA approaches like the one in Hermes offers a structural advantage over simple judge-based schemes (like OpenRouter Fusion). Instead of just picking the best single response, the aggregator can synthesize a better reply by collecting insights from multiple models (not just replies), leading to stronger reasoning, especially for complex or long-horizon agentic tasks.

Nous Research @NousResearch

2 days ago

The strongest models are gated and access is granted only to a select few. Hermes Agent now exposes MoA presets as virtual models, giving you capabilities beyond the publicly available frontier: 8% higher than Opus 4.8 and 11% higher than GPT 5.5 on our upcoming benchmark.

314 595 6K 1.7M 3K

@nicoalbanese10 @vercel @aisdk 🫡 Excited for what's next!

OSWorld 2.0 measures long-horizon computer use with 108 workflows, ~1.6 human-hours and ~250+ steps each lots of headroom on the pareto frontier - Opus 4.8 scores 20.6% (w/ ~224K tokens/task) work led by @yuan_mengq43669 @taoyds & @XLangNLP group - we @SnorkelAI were honored to collaborate as research / data partners!

XLANG NLP Lab @XLangNLP

3 days ago

Two years ago, we built OSWorld 1.0 — the benchmark that became the standard for computer-use agents. Agents now score 83.5% on it. Problem solved? Not even close. 🚀Today we introduce OSWorld 2.0: Benchmarking Computer Use Agents on Long-Horizon Real-World Tasks. What's new:

13 105 401 134K 171

Reading through Anthropic's new report I get the gist ... that people who use AI in more automated ways (i.e., delegate more tasks to Claude) are significantly more optimistic about AI’s impact on their jobs, not more fearful. ... but people who are already optimistic or comfortable with AI are much more likely to use it in highly automated ways. So the correlation could easily be reversed. Selection Bias

Anthropic @AnthropicAI

2 days ago

To keep pace with AI progress, we're advancing how we study Claude's economic impact. Hourly sampling and survey data show us how the cadences of life shape usage, what people produce with Claude, and how perceptions of AI's impact may be changing. anthropic.com/research/econo…

390 363 3K 571K 890

We measure GLM-5.2 decode at ~111 tok/s (Fireworks, median @ 100k context) Faster than GPT-5.5 at ~71 tok/s Yet GPT-5.5 finishes a median coding task in ~10 min, vs ~19 for GLM Fast decode is good! But on real (deep context) tasks model behavior dominates total wall clock time

OpenAI's Codex active users grew more than 5X the first half of 2026, and the fastest growth came outside the original developer audience. Usage exploded (per OpenAI's new research report): - Longer tasks are becoming the new normal. By May 2026, 80.6% of sampled individual users had asked Codex to do work that would take a human more than 30 minutes. 70.2% had asked for tasks over 1 hour, and 25.6% had requested jobs above 8 hours. - AI agents are not staying with developers. Engineers adopted Codex first, but Legal, Finance, Recruiting, and other non-technical teams quickly joined in. Within OpenAI, Codex now generates 99.8% of weekly output tokens. - Non-developer adoption is rising fastest. Since August 2025, non-developer Codex usage grew about 137x among individual users and 189x among organizational users. - Agents are expanding what workers can take on. Non-technical employees are using Codex for automation, data transformation, debugging, tooling, analysis, and technical tasks beyond their usual jobs.

The AI industry spent the last few years focused on language models that understand text, generate code, and answer questions. Increasingly, attention is shifting toward a different challenge: helping AI understand and interact with the physical world. A growing number of researchers and startups are now investing in world models, systems designed to learn how environments, objects, and physical processes behave over time. The goal is not simply to generate responses, but to help AI predict outcomes, navigate spaces, and make decisions in dynamic environments.

Time to transform collective experience into intelligence that can be accessed, applied, and scaled. #AI #KnowledgeManagement #OrganizationalIntelligence #AIStrategy

all you need is you daily dose of the @arcinstitute to get the hope back in a world filled with challenges. great job @collision & @patrickc, thanks for inspiring younger aspiring scientists.

Arc Institute @arcinstitute

5 days ago

Imagine programming protein, DNA, and RNA systems like you would write computer code, or even by natural language prompting of an AI agent. @BrianHie and team just made this a reality with Proto: a high-level programming language for generative biology.

34 262 1K 350K 986

guys it's not that deep. we just built a sandbox within 2 days. just open sourced it. 38ms startup (hot pool, p50 ~60ms p95) 1s (cold boot, p95, 1.5s p95) auto resumes, 60fps VNC, PTY terminal sessions, instant fork, persistent volumes costs $0.009/hr

Anurag Goel @anuraggoel

6 days ago

The first nail in the coffin of every sandbox-only startup:

64 91 1K 423K 877

Previous AI cycles had no products. The current AI cycle has massive practical applications. Enough to economically sustain needed investments? Good question! BUT: if investors pull the plug on Open AI, Anthropic, etc, there won't be just another AI winter 3 but a Black Friday 2.0.

Brian Roemmele @BrianRoemmele

a week ago

x.com/i/article/2068…

8 22 85 87K 93

World model has a very clear definition: predicting the next state given the current state and an action. For years that frontier has lived in pixels. This work pushes it into the cyber world, where agents actually operate. Qwen-AgentWorld, led by Yuxin Zuo @zuo_yuxin and Li Sheng from our Tsinghua group with the Qwen team, is a native Language World Model that simulates seven agent environments (Terminal, SWE, MCP, Search, Web, Android, OS) in one model: give it the interaction history and the agent's next action, and it predicts the exact environment response and the action. It's one of the first language world models trained to frontier-level fidelity across this many domains— the 397B-A17B model is on par with leading general-purpose frontier models on overall simulation quality, and even the dedicated 35B version reaches that band. In the end, the cyberspace and the physical world will be wired together.

Qwen @Alibaba_Qwen

5 days ago

📣📣 Meet Qwen-AgentWorld — a native language world model that simulates 7 agent environments (MCP, Search, Terminal, SWE, Web, OS, Android) within a single model. Environment modeling is the training objective from day one, not a post-hoc adaptation. 🤔 LLMs are trained to be

199 781 5K 1.1M 4K

City of the past

Ram Ahluwalia CFA, Lumida @ramahluwalia

5 days ago

NYC has defeated Amazon HQ, Citadel HQ, and now driverless cars. NYC is famous for once laying claim to the largest suspension bridge, skyscraper, and mass transit system.

24 16 262 131K 21

What exactly is this? Why are so many people celebrating, and why is no one asking any questions? What is the performance of this against Nvidia for inference? How many can they make per year? How much does it cost? Which node and where? What type of system will these chips run on? AMD has been struggling to keep up with Nvidia, and the same goes for Meta and Microsoft, who have both been in this for so long. And if everything is smooth and there are no problems, where are they going to deploy these? Which fantasy Stargate?

OpenAI @OpenAI

5 days ago

We’ve designed and built our first AI chip: Jalapeño. Designed from the ground up by OpenAI and brought to production with @Broadcom, Jalapeño is purpose-built for the LLM workloads powering ChatGPT, Codex, the API, and future agentic products. Chips are foundational to the AI

1K 2K 23K 6.9M 3K

Anyone found nested sub agent in Claude code annoying? it made session longer for almost no reason, with additional context loss on each layer Maybe I'm not using it right

the FT has started referring to openai as “the lossmaking ai group”