#Agent

Aravind Srinivas Congratulates Meta on Training a Frontier-Grade Model

Congrats to Meta for training a frontier-grade model! Watermelon compares favorably to Fable 5.1 on WANDR evals. https://twitter.com/alexandr_wang/status/2109689491300794551

How Anthropic’s Dynamic Workflows Coordinate Claude Agents

Anthropic’s dynamic workflows let Claude plan tasks and write orchestration programs that launch agents, collect results, and manage information flow, according to a workshop reported by Richard MacManus. Developers can configure specialist subagents, set session budgets, and inspect execution graphs in Claude Console. Presenters said a stock watchlist workflow cost much more than a single-agent approach but found more details; automatic self-improvement was not built in.

Sakana AI's self-improving model gets better at solving and judging, but can't catch its own blind spots

Researchers from Sakana AI and UC Berkeley introduced MASS, a method that uses copies of Qwen3.6-27B to solve tasks, evaluate results, and optimize workflows without a stronger teacher model. After two rounds, its test-task win rate against the base model reached 69.9 percent, although benchmark gains largely reflected more completed work and some scores declined. Shared blind spots allowed flawed results to pass evaluation and enter training, prompting the researchers to recommend independent checks, limits on autonomy, and clear intervention conditions.

How Perplexity built the Projects filesystem as a Git forge for agents

This is how Perplexity built the filesystem behind Perplexity Projects as a Git forge for agents. Each agent works in its own sandbox and commits to a shared repository where people and other agents can build on its work across sessions. Read more: https://t.co/jcrTptb2IW https://t.co/dXnhroO4mg

Meta releases Muse Wonder for developers as consumer access remains unclear

Baoyu reports that Meta released Muse Wonder and Muse Wonder Full on October 12, with Muse Wonder available today through Meta Model API and Muse Code, while Full is limited to Muse Code subscribers. Meta’s blog says consumer access is coming soon, whereas Zuckerberg says subscribers can use it today. The models support a 1 million token context window and multimodal input with text output. Meta reports strengths in sustained agentic work, coding, and memory, but its comparisons include internal benchmarks and competitor runs that may not be optimally configured.

The same story is shown in the picks as 《Meta releases frontier models Muse Wonder and Muse Wonder Full》

Alexandr Wang reacts to a post about Meta’s model

@citrini says Meta has a real shot at capturing agents for small business and that its model is surprisingly decent at finance tasks. Alexandr Wang reacts: “the bois r speaking”

Alexandr Wang Says Muse Wonder Is State of the Art for Personal Agents

Alexandr Wang says Muse Wonder is state of the art for personal agents. He says Meta developed the Personal Superintelligence Index (PSI) internally to measure what makes a model useful for personal agents, beyond external benchmarks, and that the model performs well in areas including Active Memory and Active Focus.

Meta introduces Muse Wonder frontier models for personal and professional agentic work

Nathan Lambert writes, “Looks like Meta is back at the frontier 🫡” He quotes @finkd introducing Muse Wonder, Meta’s new family of frontier models, with claimed state-of-the-art performance for personal and professional agentic work.

Meta launches Muse Wonder with developer access and conflicting consumer availability statements

Meta released Muse Wonder and Muse Wonder Full on October 12, according to Baoyu’s account of the announcement. Developers can access Muse Wonder through Meta Model API or Muse Code, while Full is available to Muse Code subscribers; Meta’s blog and Zuckerberg’s X post differ on consumer availability. The models support a 1 million Token context window and multimodal input with text output, while Meta’s reported benchmarks show strengths in sustained agentic and coding tasks alongside weaker results elsewhere.

The same story is shown in the picks as 《Meta releases frontier models Muse Wonder and Muse Wonder Full》

Meta releases Muse Wonder and Muse Wonder Full with higher Artificial Analysis scores

Meta has released Muse Wonder and Muse Wonder Full, which Artificial Analysis reports score 52 and 56 on its Intelligence Index v4.3 at max reasoning effort, versus 48 for Muse Spark 1.3 (max). Muse Wonder Full improves across all ten evaluations, with the largest gains in agentic evaluations. Muse Wonder is available through Meta’s first-party API and Muse Code at launch, while Muse Wonder Full is limited to Muse Code users; Muse Wonder costs $4.52 per Intelligence Index task.

The same story is shown in the picks as 《Meta releases Muse Wonder and Muse Wonder Full with higher benchmark scores》

Alexandr Wang Reacts to @charlieholtz’s Post About Wonder

@charlieholtz says Meta is at the frontier and calls Wonder “really really good,” citing a security vulnerability it found, its design and visualizations, and features in its CLI. Alexandr Wang reacts: “what a wonder full world 🍉”

MUSE WONDER 🍉 this model was developed to power personal agents. three highlights: - ACTIVE MEMORY: best memory across days/weeks/months - ACTIVE FOCUS: gets your tasks done - DOMAIN EXPERTISE: sota in finance & legal, strong in health

MUSE WONDER 🍉 this model was developed to power personal agents. three highlights: - ACTIVE MEMORY: best memory across days/weeks/months - ACTIVE FOCUS: gets your tasks done - DOMAIN EXPERTISE: sota in finance & legal, strong in health

introducing muse wonder our new frontier model. state-of-the-art at personal agents and professional agents. our largest and most capable model yet—rivals the big bois

introducing muse wonder our new frontier model. state-of-the-art at personal agents and professional agents. our largest and most capable model yet—rivals the big bois

MUSE WONDER 🍉 muse wonder is the best model in the world for personal agents and professional agents. we evaluated our model against the best models in the world across agentic and coding tasks, and we think you all will like it.

MUSE WONDER 🍉 muse wonder is the best model in the world for personal agents and professional agents. we evaluated our model against the best models in the world across agentic and coding tasks, and we think you all will like it.

Meta introduces Muse Wonder and Muse Wonder Full for agentic tasks

Meta introduces Muse Wonder and Muse Wonder Full, describing them as its most capable models yet and frontier models built for personal and professional agentic tasks. Muse Wonder is available today via the Meta Model API and Muse Code. Muse Wonder Full is available today for Muse Code subscribers. Muse Wonder will also soon power experiences in Meta’s personal agent product, Muse, rolling out first to subscribers.

The same story is shown in the picks as 《Meta releases frontier models Muse Wonder and Muse Wonder Full》

Z.ai introduces GLM-Legal, built on GLM-5.3-Flash

Introducing GLM-Legal Built on GLM-5.3-Flash, GLM-Legal matches frontier models on mainstream legal benchmarks, according to Z.ai. Designed for real legal workflows: contract review, evidence analysis, compliance, and more. Tech Blog: http://z.ai/blog/glm-legal Apply for API Access: http://z.ai/consultation/legal Try AutoClaw: http://autoclaw.z.ai

Run multi-node inference workloads using Ray LLM Deep Learning Containers

AWS shows how to deploy Qwen3.5-9B across two GPU nodes on Amazon EKS using Ray LLM Deep Learning Containers, with separate prefill and decode deployments behind one OpenAI-compatible endpoint. The walkthrough configures the application in YAML and uses NIXL over EFA to transfer the KV cache between nodes. It covers cluster provisioning, model invocation, transport verification, and resource cleanup.

Deploy decision models with llama.cpp on Amazon SageMaker AI

AWS explains how to deploy GGUF models with its llama.cpp Deep Learning Container on Amazon SageMaker AI, then add a SemIf adapter that returns probabilities over declared options without generating answer tokens. The guide includes CPU and GPU deployment code. Across 231 JevBench tasks using Qwen3.5-4B Q4_K_M, the GPU configuration delivered 4.9 times faster p50 and 32.2 times faster p95 latency than the CPU configuration, with similar accuracy.

Clearing chat history helps AI agents find better solutions

A Microsoft Research study reports that Fork-and-Flush helps long-running AI research agents explore better solutions by creating parallel branches that retain workspace files but discard chat history. Across 13 tasks with the same compute budget, it achieved a min-max-normalized average score of 0.78, versus 0.47 for single runs and 0.54 for best-of-4. The researchers found that accumulated context can trap agents in similar approaches, even when conversation histories are periodically compressed.

Aravind Srinivas Thanks @MollySOShea for the Conversation

@MollySOShea says she spoke with Perplexity Founder and CEO Aravind Srinivas about why AI-native SMBs could drive the next American millionaire economic boom. Was a fun chat! Thanks for having me @MollySOShea! https://twitter.com/mollysoshea/status/2109645765958132083

TechCrunch Disrupt 2026 Starts Tomorrow, Offering Ticket Savings of Up to $100

TechCrunch says Disrupt 2026 starts October 13 at 8 a.m. PT at San Francisco’s Moscone West and runs through October 15. Attendees can save up to $100 on a pass, get 50% off a second pass of the same ticket type, or buy a $75 Expo+ Pass if they’ve been laid off.

Qwen introduces QwenGyre for elastic RL in extreme-long-horizon agent training

Qwen introduces QwenGyre, bringing elastic RL to extreme-long-horizon (XLong) agent training. Its Elastic Scheduler reallocates GPUs between rollout and training without interrupting live executions, while its Trajectory Processor converts branching executions into prefix-shared trajectory trees with bounded, deduplicated training samples. Scaled to Qwen3.8-2.4T-A95B with ~700K tokens per rollout, QwenGyre raises NL2RepoBench from 52.5% to 58.5% in 48 steps and cuts end-to-end training time by ~44% vs Async. Paper: https://arxiv.org/abs/2609.33848

Former Plaud exec raises $26M for Ambi, an AI app that listens in the background

Former Plaud executive Peter Mo’s startup Ambi has raised $26 million across multiple funding rounds led by Luminous Ventures and Huaqin. It is releasing apps for iPhones, Macs and Apple Watch that can capture conversations throughout the day, with an AI agent intended to offer suggestions and begin tasks during meetings. Mo says ambient capture is off by default, but users remain responsible for informing others; Ambi says it stores no recordings and encrypts data without specifying the encryption method.

How Anta Uses Volcano Engine Agents Across Global E-Commerce Workflows

Anta is using Volcano Engine Agents for promotional videos, overseas ordering, and cross-border returns. The article reports that an ArkClaw workflow packages four ordering templates into Skills, reducing ordering decisions from half a day to 10–20 minutes, cutting manual entry by about 90%, and achieving data accuracy above 99%. A returns Agent automatically creates over 1000 return documents monthly for Anta’s TikTok cross-border business, with expansion to FILA and other brands planned.

Baidu World 2026 to Be Held in Beijing on November 12

Baidu announced that Baidu World 2026 will take place on November 12 at the China National Convention Center Phase II in Beijing, under the theme “The Next Act: Intelligent Organizations.” The event plans one main forum, 12 breakout forums, more than 30 AI classes, and a 6,000-square-meter exhibition area. Its forums will cover AI-driven organizational change, including AI office tools and enterprise deployment of agents.

Huawei and UCL’s Memento 3 Clears 25 ARC-AGI-3 Games Without Updating Model Weights

Huawei’s Noah’s Ark Lab, UCL, and collaborators report that Memento 3 cleared all 25 public ARC-AGI-3 games and 183 levels using 7518 actions, about 56% fewer than humans, achieving a 100 RHAE score. Built on Claude Opus 5 without weight updates or game-specific fine-tuning, the agent iteratively revises a Markdown rulebook and executable Python world model. In Atari Pong, its learned code-based strategy won three games 21:0 without further LLM calls; the experiments primarily cover game environments with relatively clear rules.

Step 5 Preview Gets Two Weeks of Free Access on Tencent WorkBuddy

Tencent’s WorkBuddy has integrated Step 5 Preview, the latest flagship model from StepFun, and is making it available free for two weeks starting today. The model is designed to help users with deep research, information retrieval, data analysis, document and presentation generation, and sustained execution of complex, multistep tasks.

Anthropic’s Claude Subscription Rules Appear Looser, but Amp’s Status Remains Unclear

Baoyu says Anthropic appears to have relaxed restrictions on Claude Agent SDK and claude -p usage, but has made no public statement specifically addressing Amp or guaranteeing against account bans. Quoted author Quinn Slack interprets a policy update from 2 days ago and Anthropic employee comments as allowing SDK and third-party apps to use subscription allowances, citing https://support.claude.com/en/articles/15036540-use-the-claude-agent-sdk-with-your-claude-plan and https://news.ycombinator.com/item?id=49997143. Baoyu recounts earlier restrictions and says Anthropic paused its planned June 15 switch to separate billing for Agent SDK and claude -p.

It has become disheartening that the average replybot on this site has shifted its replyslop angle to asking interesting questions. It feels like actual intellectual engagement, which makes it somehow worse than just rephrasing the post with a "that's the part people miss" bit.

It has become disheartening that the average replybot on this site has shifted its replyslop angle to asking interesting questions. It feels like actual intellectual engagement, which makes it somehow worse than just rephrasing the post with a "that's the part people miss" bit.

Satya Nadella argues enterprises should treat AI models as insider risks

Baoyu summarizes Microsoft CEO Satya Nadella’s argument that enterprises should manage AI models as insider risks, with identities, restricted permissions, audit records, and controls to halt operations. Nadella outlines seven principles covering model diversity, observability, verification, independent control and auditing, containment, and incident disclosure. He calls for disclosing chains of thought while acknowledging that they may not faithfully reflect model reasoning, and argues that security controls must operate outside the model.

Baoyu uses REA and Computer Use to investigate CapCut subtitle templates

Baoyu describes using REA with Computer Use to investigate how CapCut renders dynamic subtitles, without knowing where their formats were stored. He had it operate CapCut, download Templates, and monitor network URLs and disk save locations. After downloading several dozen through clicks, Codex found clicking too slow and constructed URLs to download directly, ultimately downloading 2291 Templates.

Using REA to Trace How Codex Attaches Browser Annotations

Baoyu describes using the open-source REA (Reverse Engineer Anything) MCP service to understand how Codex and the ChatGPT app attach browser annotations to the composer, without decompiling Codex itself. After REA traced the Electron app's code path, Codex built a test page that could open a comment box and prefill input via document.oai.annotation.request(), an OpenAI API the author had not known about. Cursor and Claude Code did not support a similar API.

Toyota's Kaizen Rule Applied to AI: Don't Cut Staff After Efficiency Gains

A Japanese blogger cited Toyota's Kaizen rule — never cut staff after process improvements reduce work hours — and argued the same applies to AI adoption. Companies that cut headcount as soon as AI boosts efficiency will find employees hide their best AI methods, he said. A quoted post adds that DORA research found organizations addressing employee job concerns had higher team AI adoption.

Google pilots AI-assisted coding interviews for software engineers

Google is piloting a new software engineer interview round called Code Comprehension, in which candidates read, debug and optimize existing code with Gemini open, Business Insider reported in May based on an internal document. A Google spokesperson confirmed Gemini is the pilot assistant. The round targets junior to mid-level roles in select US teams, starting with Google Cloud and Platforms and Devices, and assesses prompt writing, output checking and debugging.

Karpathy Says the LLM Capability Gap Is Becoming a Funnel

Andrej Karpathy argues the gap in shared understanding of LLM capability is widening into a sharp funnel: roughly 6B people have barely touched LLMs, 1-2B use free-tier chatbots casually, about 20M use frontier models for math and code, and roughly 5,000 have internal access to frontier systems. He attributes the funnel to problem ambition, capability jaggedness in verifiable domains, and access tiers.

Karpathy: LLM Capability Gap Is Widening Into a Sharp Funnel

Andrej Karpathy says the gap in shared understanding of LLM capability is widening into a "sharp funnel" rather than two groups talking past each other. He estimates roughly 6B people have barely encountered LLMs, 1-2B use free-tier ChatGPT-like products casually, about 20M see frontier-grade math and code use, and roughly 5,000 have internal access to frontier systems. He attributes the funnel to impact scaling with ambition, the jaggedness of capability, and access tiers.

Andrej Karpathy describes a widening funnel in perceptions of LLM capability

Andrej Karpathy argues that understanding of LLM capability is diverging into a sharp funnel, with rough estimates of 6B people barely exposed, 1-2B casual users, 20M professionals using frontier LLMs in math and coding, and ~5,000 people accessing internal frontier systems. He describes increasingly ambitious agent workflows toward the narrow end, while saying human review and comprehension are starting to fall behind. He attributes the divide to problem size and duration, capability concentrated in digital and verifiable domains with economic value, and differences between free, paid, and internal access.

Agent teams cost up to 5.1x more for little quality gain, Vals AI finds

Evals company Vals AI tested GPT-6 Sol and Claude Opus 5.5 on the Vibe Code Bench, both solo and as teams, at medium and maximum reasoning effort. Teams cost between 1.8x and 5.1x more than single agents, and only one of four comparisons showed a statistically significant improvement: GPT-6 Sol at medium reasoning, where the team scored 7.3 points higher. At maximum reasoning, teams gave neither model a real advantage.

Alexandr Wang endorses @Collin_Lemmon’s praise of Muse’s consumer AI vision

@Collin_Lemmon says Alexandr Wang shared a consumer AI vision: “I think muse can give everyone the adulthood they dreamed of in childhood”. Wang says, “unironically this is exactly what i believe”

PolyAI and Otter Executives Say Voice AI Hasn’t Had Its ChatGPT Moment

PolyAI CTO Shawn Wen says voice AI still needs faster reasoning to make conversations feel natural, even though full-duplex models can speak while listening. Otter CMO Alex Gay says accurate transcription, language understanding, and expressive voices matter for enabling automation and building trust. Both executives also stressed that voice AI tools should tell people when they are talking to AI or being recorded.