The same model, with the same weights, scores 62% in one agent harness and 33% in another. @adithya_s_k and the @huggingface team just released the ultimate guide to multi-harness RL, and it's one of the most practical RL write-ups this year, and everything open! The trick is simple. Don't touch the harness. Point it at a proxy instead of the model. The proxy speaks all four API formats coding agents use (OpenAI Chat Completions, OpenAI Responses, Anthropic Messages, Gemini). It records the exact token ids and logprobs vLLM sampled, and you train on that. You don't change a single line of Claude Code, Codex or OpenCode. Results: 🔹 Trained across 4 harnesses at once, LFM2.5-2.6B by @liquidai went from 42% to 54% 🔹 31% fewer tool calls, thanks to a small bonus for solving tasks in fewer steps 🔹 Training in OpenCode alone took OpenCode from 34% to 58%, but the multi-harness model improved everywhere They also tried the shortcut everyone reaches for: fine-tune on 3,189 successful rollouts from Qwen3.8-27B. Imitation plateaued at 47.5%, below both RL runs. Copying a bigger model doesn't get you there. Practice does. The best part is that everything is open: the capture proxy in OpenEnv, the trainer in TRL, the tasks, the SFT data, the training code and all seven trained models. Agents will run in dozens of harnesses. Now open models can be trained for each of them, by anyone. Read it here 👇 https://huggingface.co/spaces/FineEnvs/multi-harness-rl

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

It's time to get serious about Security x AI. One of the joys of doing AIE is helping others start their own high quality conferences in their countries and domains. Every single 2025 partner has come back — and the 2nd AI Security Summit is now way more important than before. Proud to be back as a founding partner with @snyksec! Join the event: https://aisecuritysummit.com/ The landscape has completely transformed since we started this a year ago. An exploding number of rogue agents, more breaches, more AI-driven attacks. now more than ever, security must be at the center of the AI conversation. With so much more hacks and vulnerabilities, this stuff is no longer theoretical...

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Ilya Meyzin, SVP, AI Solutions & Data Science at @DunBradstreet is speaking at Interrupt, The Agent Conference by LangChain. Catch Ilya’s talk + more: https://interrupt.langchain.com/london

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

This week in The Batch, Andrew's Letter covers how open weights model GLM-5.3 nearly matched Claude Mythos at exploiting vulnerabilities, 12% vs. 14%. Also inside: 🧪 Xiaomi's new open weights leader 🎙️ Gemini 3.8 Live 💾 DeepSeek-V4.1-Flash 🔁 AREX agents 🔗 https://hubs.la/Q04z20mv0 #DeepLearningAI #AIEngineering #OpenWeights

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks: Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better: Diagrams / images. Instead of writing, ask your LLM to create a diagram. These can be a lot easier to process, parse, and understand. But even better: Web pages. Ask for output "in HTML" to get a beautiful, interactive webpage. LLMs are getting really good at frontend and can create beautiful experiences, animations, etc. But even better: Explainer videos. The output format I am most bullish on is fully custom / bespoke explainer videos generated on any arbitrary topic. Experiment with things like "Create a 3b1b style video explainer on X. Use my ElevenLabs API key for audio narration". (you'd need an API key for the latter or you can ask your LLM to find you decent free alternatives that use your local compute). This is actually starting to work! In summary: - As LLMs get better, they will do more and more of the legwork autonomously, and a lot more of our work will rise up the abstractions into oversight and understanding. - Luckily, LLMs can help here too because as intelligence and code are increasingly abundant, you can ask for large, custom, discardable software artifacts (e.g. web apps, video explainers) that would have never made sense to create before. Push the boundaries here and you'll be surprised.

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

LangSmith for Startups Spotlight: @corridor @Corridor uses LangSmith to develop + run novel agentic security evaluations, build + trace inference pipelines, and manage memory graphs that provide specialized context across various product surfaces. Get a demo: http://corridor.dev/demo

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Validate agent fixes before shipping them with LangSmith Engine v2. Engine now tests and validates fixes automatically: ✅ Replicates the issue with the same deployment environment and same inputs ✅ Builds a fix, tests it, and iterates on improving it until it has a satisfactory solution ✅ You review the fix and deploy it with a ready-made PR

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Arena started in 2024 as a UC Berkeley Sky Computing Lab project for comparing models head to head. Today, it's grown into a platform for evaluating model performance across agents, text, code, images, and videos, running up to 600,000 @e2b sandboxes a day. Each sandbox is an isolated cloud computer where an agent can write code, install dependencies, and work for hours on coding, research, reports, and presentations. That isolation keeps results trustworthy and secure: no session's code or files can reach another's and skew the comparison. That security at scale gets tested every time a frontier model drops. Before GPT-5 went public, people rushed to Code Arena to try it first. "When GPT-5 was about to come out, a lot of people came to Code Arena to experience the model firsthand because it wasn't out to the public yet. E2B was the backbone behind all of that. So when we had this massive surge, we didn't have to worry about whether we could handle it." - Aryan Vichare, Founding Engineer, @arena Read the full case study: https://e2b.dev/customers/arena Watch the case study video: https://www.youtube.com/watch?v=vBKHCiZPoRE

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

When inference is the operating budget, efficiency matters. With AMD Instinct GPUs on @DigitalOcean, @Character_ai doubled production inference throughput and cut cost per token by 50%. More room to serve users. More advanced models. Same compute footprint. Read more at https://bit.ly/3SrbO2q

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

An agent told a customer the wrong thing, confidently. No exception or failed test, just clean logs. The bug was three steps upstream: right tool, wrong argument. LLM-only tracing only sees the tool call, not what the database did with it. @antonikozelski from Vstorm shares why observability has to be a first-class layer, and what that looks like with Pydantic Logfire. Read the post: https://pydantic.io/imtxd

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Grok 4.7 is now available on the Gemini Enterprise Agent Platform

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Adding routing to your agent? Here's where to start: 1️⃣ Understand the tasks 2️⃣ Understand the models 3️⃣ Build the router in the harness 4️⃣ Track task outcomes An inside look at how to build a model router in the harness, and how we cut our median cost per thread by 64% with no change in quality.

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Congrats @CoreWeave on RL Rollouts! RL post-training involves a lot of back and forth: train the model, generate responses, then train again. Inference workers need to load the updated model weights each time. As models get bigger, that can leave GPUs waiting. CoreWeave’s new service uses ModelExpress and Router in NVIDIA Dynamo to speed up those reloads with minimal downtime. Working with us and @youdotcom, CoreWeave achieved 15× faster model reloads compared with its baseline while post-training Nemotron 3.5 Lightning. Check out their blog below for details

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Up to 8x faster, and now you know why ⚡️@openai GPT-6 Astra Ultrafast runs on NVIDIA Read the story: https://nvda.ws/3W2bZ5H

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

“We find that on medium-length, well-defined accounting tasks, frontier AI models are now faster and more accurate than junior accountants, even the best one in our study.” Eighteen months ago they scored well below human accountants Good discussion here: https://www.mercor.com/blog/human-baselines-for-benchmarks-ai-now-outperforms-junior-accountants/

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Cool eval. Simply ask an LLM “Land or Water?” and give it a latitude and longitude coordinate as text. Ask 16,200 times, plot as image. The models know. From compressing the internet.

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Introducing Claude Code mods! Customize how it looks and works just by prompting it. https://claude.dev/blog/getting-started-with-claude-code-mods/ A mod is a small JS/TS file that runs in your session. Watch events, rewrite behavior or draw UI. Share it as a plugin with others. Make Claude Code your own.

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Build Visual AI Agents From a Prompt With NVIDIA Cosmos and VSS 3.3 https://x.com/i/broadcasts/1qGvveVbDWaGB

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

In physics, an “impedance mismatch” occurs when two systems each work well but are poorly matched. In this Science Blog guest post, Harvard physicist Matthew Schwartz argues that something similar is happening with AI and science. LLMs are capable at many things, but working with them as you would with a human collaborator isn’t currently the best way to elicit their scientific strengths. To address this mismatch, Schwartz created a toolkit for exact calculations in quantitative science. Because similar calculations often emerge in very different areas of science, Claude found connections to ecology, population genetics, and a dozen other fields, and Schwartz worked with domain experts to steer it towards interesting questions. Read more about these projects here: https://www.anthropic.com/research/claude-shaped-science

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Global reset landing tomorrow 10am PST for all paid ChatGPT accounts. Apologies for the slow start with GPT-6.1 Sol, it's now back to running at expected speeds after the massive load spike in the first two days.

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

The critical distinction between base LLMs (2024 and earlier) and modern LRMs is not symbolic tool use. It's the switch from a transductive paradigm (intuit the answer to the query) to an inductive paradigm (intuit the program/instructions that produce the answer to the query). They're trained to be inductive, and they perform test-time induction, i.e. test-time prediction of a NL program / reasoning chain. This unlocks entirely new capabilities -- in particular fluid intelligence. Base LLMs, to this day, have ~0 fluid intelligence. LRMs have substantial levels of fluid intelligence. The performance of LLMs on ARC 1 (a benchmark from 2019) remains ~10-15% today. Scaling them up by a factor ~100,000x got them from 0% to 10%. Meanwhile LRMs the same size or smaller saturated ARC 1 in 2025.

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

R to @LangChain: Learn more → https://www.langchain.com/blog/langsmith-engine-v2-redteam

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

🧵 @langchain / MCP-adapters 2.0

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

LangChain Academy Course Update: Introduction to Deep Agents. Deep Agents makes it easier to build agents, Managed Deep Agents lets you deploy them with a single CLI command. We introduce Managed Deep Agents and show you how to deploy an agent in Slack. Check it out today: https://academy.langchain.com/courses/foundation-introduction-to-deepagentsMDA

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

What is the New Application Layer? Malte Ubl (@cramforce), CTO of @vercel, came to Navigate 2026 to talk about how software is built for agents now.

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

"Today, we’re releasing two Cloudflare-trained decision models, Clef and Clef-flash" https://blog.cloudflare.com/clef-decision-models/ Never seen an idea spreading so fast.

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Congratulations!

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Super Intelligence (fka AI) is now acing accounting tests

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

For two weeks, start a design, deck, or doc in the Claude app, and the work that follows in that conversation uses 50% less of your usage limits. Give it a try with Claude Sonnet 5.5. It has a strong eye for design and builds slides that need minimal editing.

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Down to 6110 unread from the over 9000 unread emails just earlier. Will get to inbox zero with clean filters within the next 48 hours. Dot has accepted the challenge.

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Also, 6.1 Sol was our fastest-growing model ever, and was a bit slow under load. Should be much better now!

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Ask @Grok in XChat

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

You can ask your dot to "create a pet and set it as your avatar". It can be based on an idea, an image or pretty much anything. Here is my dot

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

When AI-generated content has become so realistic, how do we verify what's made by human, machine, or nature? We explore the vital role of provenance – and how imperceptible watermarking can safeguard both digital media and biological designs – on our podcast. ↓ Timecodes: 00:00 Introduction 00:34 What is a watermark? 04:35 SynthID 15:16 Images and video 19:28 SynthID Bio 28:00 Future of resilience

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

You should be able to use your AI subscription wherever you need:

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

holy

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Sitting through PowerPoints has become much better since we got good AI. Sure, some people just let AI do the thinking work, but that is obvious (and they used default templates before). For people who do care, they have interesting layouts & funny visual jokes & good diagrams.

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

R to @steipete: "AI agents are aeroplanes for the mind: faster and more powerful than the bicycle, harder to control, costlier when they crash."

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

R to @Google: Learn more. https://goo.gle/4hcNNWK

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

RT by @demishassabis: Up, up, and away. 🚀 Today, in partnership with @planet, we launched a prototype satellite carrying four TPUs into orbit on @SpaceX's Transporter-18 rideshare mission. This launch is the first step of Project Suncatcher, our long-term research moonshot to see whether we can one day host scalable machine learning infrastructure in space. Over the coming weeks, we'll gather in-orbit data on how our TPUs handle the physical stress, radiation, and thermal extremes of space. Whatever we learn, we'll use it to refine our future designs.

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Andrew Ng's AI Engineering Skills Map shows what to learn. AI Dev is where you hear from the people who built it in the real world. Keynotes from Andrew Ng and Yann LeCun. Engineers from Hugging Face, Google DeepMind, and BlackRock. 🗓️ Nov 30 to Dec 1 📍 New York City 🎟️ Grab early bird tickets at $599 while they last Get your ticket: https://hubs.la/Q04yWnyV0 #DeepLearningAI #AIEngineering #MachineLearning

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Come and join Pydantic!

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

"the Waymo effect is what happens when a technology removes the friction of dealing with another human being" https://www.researchagenda.news/articles/the-waymo-effect.html

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

R to @browserbase: If you prefer to watch on YouTube: https://browserbase.run/navigate-vercel

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

i have so many questions

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

R to @claudeai: Applies automatically each time you create a design, deck, or doc, and runs through October 15 on Pro, Max, and Team plans. How it works and full terms: https://support.claude.com/en/articles/17274727

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

I'm speedrunning my filesystem learnings. Turns out, brtfs cow-feature is GREAT for worktrees and terrible for sqlite. Next OC update knows that and - if you did sth silly like me - will migrate your database to a NOCOW spot.

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Damn WiFi

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

All my oss releases stopped because Apple put up a new dev agreement. And since it's lockstep it blocked Linux/Windows releases as well. 🙃

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

R to @GoogleDeepMind: Watch → https://goo.gle/4de3fiR Spotify → https://goo.gle/4dKZwte Apple Podcasts → https://goo.gle/4dQqKyH Or listen wherever you get your podcasts! 🎧

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

Hey @elonmusk - does this give Ludicrous+ mode a boost?

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

R to @sama: I think there is much more potential energy in Sign In With ChatGPT/Plugin Extensions than we realize

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）

DevDay was really fun, the builder energy was crazy Still wild to me to see people making whole startups in a day

收錄日期

2026-10-02

來源

Nitter RSS（公開貼文）

抓取時間

2026/10/02 23:47（台北）