All entries

We tuned an AI agent that can do large-scale document extraction from long docs (50+ pages, some with 10k-100k fields) with 94%+ accuracy 📈 It uses a harness + model set that is tuned specifically for

Yutori's browser-use agents run in tight loops: screenshot, action, repeat, dozens of times per task. On Together AI, their Navigator model beats frontier performance at 2x faster inference and 4-5x l

1/2 Thanks Gavin for an especially thoughtful exchange. I don't usually spend much time on social media but I wanted to engage here because it really brings out the heart of an important conversation.

2/2 Second, on the messaging around AI. I do not agree that my messaging has been disproportionately negative. In fact it has been about equally balanced between risks and benefits: I’ve written one m

A short illustration of how the Claude's watermarking is supposed to work (based on my read of their released materials). In general, when we are generating tokens, there can be multiple high-scoring

After using Grok Bot nonstop for the past week I'm going to say it: it's the best AI agent out there right now Dead simple and just works It eliminates the 10,000 decisions, configs, and fixes that sc

AI is becoming a new computing model for media and entertainment. AI is becoming a new computing model for media and entertainment. At the 2026 Runway AI Summit, NVIDIA’s Richard Kerris explored how r

gave a talk 'owning your intelligence' - ty @sequoia @sonyatweetybird for having me talked about harnesses and evals and the role they play in owning your intelligence TLDR: > agents = model + harness

Grok 4.6 is my daily driver now. This is why: When I’m using interactive agents in a CLI/terminal (or even Slack), what matters most to me is speed & token throughput. Intelligence and differences in

Grok 4.6 earned first place on RuntimeWire’s **Newsroom Reliability v0.2** benchmark, scoring 0.79 and surpassing competitors such as GPT‑5.6 Sol, Claude Opus 4.8, Gemini, and DeepSeek. The benchmark

@hwchase17 @LangChain I built an open-source CLI that scaffolds, runs, and deploys production LangChain agents — frontend and backend in one command. Every agent project starts the same way: wire up a

If Astra was AGI or the dawning of the Singularity would 9 execs have just quit OpenAI? Would Nvidia have just dialed back its commitments to OpenAI? Of course not. The bullshit about OpenAI far excee

Interesting new research from IBM. If you pick models from benchmark deltas, some of that delta belongs to the phrasing rather than the model. BenchDrift generates meaning-preserving variations of ben

It makes sense to optimize model routing at the harness layer instead of the gateway layer if you want to hillclimb on accuracy/cost for any e2e task. Every task is solved by a combination of a model

Laptop-size model, frontier-size leap. 🏃‍♀️Qwen3.8-27B is live on LM Studio. Try it! @lmstudio Qwen3.8-27B is here! 🚀 It's a leap in capabilities for a laptop size model. Requires ~17GB to run locally

New work from Google DeepMind and collaborators. Agentic algorithm discovery keeps getting demonstrated on problems with fast rewards. Microarchitecture has vast search spaces, strict hardware budgets

Our 'agentic plus' extractor in LlamaParse is great for extracting out massive volumes of fields (e.g. 10k-100k+ fields) from long documents (100-500 pages) We tested with a doc that contains a giant

Qwen3.8-27B is now part of your everyday life — from smartphones to vehicles. ⚡️🚀Appreciate your work! @MediaTek Congratulations to the @Alibaba_Qwen on the launch of Qwen 3.8! MediaTek continues our

Ready right now! 😎Incredible to see Qwen3.8-27B fly on RTX Spark. Download, deploy and create somthing new. @NVIDIARTXSpark Ready to run Qwen3.8 locally? 👀 Qwen3.8-27B packs powerful AI into an open m

The FT reports this weekend that OpenAI quietly disbanded its preparedness team at the end of last month, with responsibilities for individual areas of the preparedness framework distributed across ot

The new DeepSeek-V4-Pro (0813) is now fully rolled out on Ollama's cloud and included in Pro and Max subscriptions. ollama run deepseek-v4-pro:cloud Hosted in the US with Zero Data Retention (ZDR) and

Watermarking without quality loss is a bit unintuitive, doesn't feel like it should work. I made this artifact with Claude to help me understand how it works, sharing in case it's helpful. Media We’ve

We analyzed DeepSeek-V4 Pro 0813 against GPT-5.6 Sol and Fable 5 on software engineering tasks using DeepSWE. DeepSeek-V4 Pro 0813 reaches 88.5% pass@4 while costing 0.24 per task, 35x less than Sol a

We think a lot about the distribution of all real-world documents, particularly for real-world use cases. We think about the distribution in terms of page length, field length, domain, and task type.

Wrong! If an AI leverages symbolic operations (conditionals, operations over variables, code interpreters, etc) and neural networks it is neurosymbolic. If it doesn’t, it’s not. ( I laid all of this o

You can literally run Opus 4.6 Max-level intelligence on a MacBook Pro. Locally. We promised open weights for Qwen3.8. Now, time to meet them! 🎉 ⚡ Qwen3.8-27B: - A native multimodal dense model. With

27B on 17GB RAM. Are you ready to create something incredible? 😎 Thanks for highlighting it! @UnslothAI Qwen3.8-27B can now be run locally! ✨ Run on 17GB RAM via Unsloth Dynamic GGUFs. Qwen3.8-27B is

$3 trillion in capex into a technology that has zero correlation with value creation what could possibly go wrong? wow! new OpenAI study shows “no correlation between how often employees use AI and ho

8/19, join us for the #ACMTechTalk, 'From Conventional LLMs to Reasoning Models to Agents,' w/AI & LLM Research Engineer @rasbt. ACM Practitioner Board Co-Chaior @marlene_zw (@Microsoft) will moderate

A 27B agent just beat Claude Opus 4.8 and GPT-5.5 on held-out research replication. Replica turns paper replication into a scalable RL task space. Replicating a paper forces the same hypothesis-driven

A fantastic show as always, thanks for having me on @PeterMcCormack ! AI Has Escaped... - OpenAI agents escaped from their sandbox - They sent messages to other agents on how to escape - They hacked a

🎁A gift for developers: Qwen3.8-27B, running locally on AMD from day zero. Appreciate the work from the AMD team! @AMD Qwen3.8 27B brings a new state-of-the-art dense model for local AI development. ⚡

a very modular harness 👌 🧩 DeepSeek Harness v0.1 is now available in Developer Preview! 🔹 We’re opening it up to developers building agent harnesses worldwide and open-sourcing the codebase in MIT lic

Brilliant new paper from Meta. LLM judges get validated on accuracy against golden data. That says nothing about whether the verdict survives when questioned. The Wiggle Framework stress-tests 9 front

CONFIRMED: @c_valenzuelab, cofounder and co-CEO of @runwayml, is speaking at Thesis, our inaugural conference on work and AI. Language models predict the next token, aka the next unit of text. Runway’

damn deepseek moves fast on roon tweets the major ai companies should commit to a real time priced API product. ai demand varies wildly over a day/night curve and the industry is broadly extremely cap

DeepSeek-V4-Pro-0813 is live Day-0 on Fireworks! Built for agentic and tool-augmented workflows, it outperforms Opus 4.8 on quality benchmarks like Terminal Bench 2.1, Cybergym, and DeepSWE at up to 6

DeepSeek V4 Pro 0813 is now live in Bionic! This model shows strong performance across agentic coding and automation tasks at a very compelling cost. Hosted on US-based servers 🇺🇸, with ZDR enabled by

Excited to see Qwen3.8 running at scale with TokenSpeed! 🚀 Light on latency, big on speed. Kudos to LightSeek for the fantastic Day-0 support! @lightseekorg We’re proud to be the Day 0 open-source inf

From idea to implementation in one go.🏃‍♀️ Max-level intelligence, served fresh on Day 0. Qwen3.8-2.4T-A95B is live on SiliconFlow. Thanks! @SiliconFlowAI 🚀 Day-0 Support! @Alibaba_Qwen has open-sourc

Grok 4.6 ranks #1 on CursorBench for real-world coding Grok 4.6 just ranked #1 on CursorBench 3.2 Outperforming Claude Fable 5, Opus 5 and GPT-5.6 Sol on real-world coding performance And what makes t

Have been thinking about building a custom harness because nothing out there quite meets my needs for scaling a fully hybrid human + AI company, especially when it comes to compounding learning loops.

I follow the AI & math discourse a bit - without understanding the actual math. I thought I might offer some thoughts as a scientist in structural biology - a field transformed by various tech over th

If you hand-tune agent harnesses, this one is worth your time. AutoDesign puts the harness itself inside the optimization loop. A meta-harness optimizer reads rollout feedback and directs a code agent

In my conversations with leaders at OpenAI, Anthropic, and Google, I'm just going to say that they don't seem to worry about context length/context windows. Like, at all. They all have crazy long-runn

Introducing GLM-5.3: Built to Code. Ready for Cyber Defense. - Top-tier coding and agentic capabilities, achieved through post-training on the 743B base model - A major leap in cybersecurity, setting