AI Brief: Apple's HALO Robot, 5,000 Agents per Repo

Apple's humanoid hit 91% task success using a Qwen model, GitHub internally tested 5,000 agents working in one repository, and Sam Altman told Stanford that a token budget now substitutes for a hundred engineers. Here is what actually moved on 2026-11-09. ▸ Apple's HALO runs on Qwen. Apple disclosed its HALO humanoid system, which calls Alibaba's Qwen model on a Unitree G1 and averages 91% success per task. The same day: Meta shipped model variants specialized through Muse Wonder training, and an unreleased Anthropic model reportedly helped crack a major quantum computing problem. (source: @NewlearnerChannel) ▸ Agent infrastructure is being stress-tested. Perplexity launched Teammate, an agent built for sustained, complex work rather than one-shot answers. GitHub's new Git storage architecture was internally tested at 5,000 agents collaborating inside a single repository while handling 7,200 pushes per second. (source: @NewlearnerChannel) ▸ Altman: tokens replace headcount. At Stanford, Sam Altman argued one founder with an affordable token budget can now do what used to require 100 top engineers. Cheaper software moves the advantage away from raw build capability toward finding real demand, selling, and business judgment. His analogy: electricity was first sold as nighttime lighting, then became a metered utility — his bet is that intelligence follows the same path. (source: @aigc1024) ▸ Rude prompts score higher. A study ran 50 multiple-choice questions in math, science, and history through ChatGPT Deep Research, phrasing each in five tones. Accuracy rose as politeness fell: the rudest tone beat the merely brusque by about 4 points, and the overly polite phrasing performed worst. (source: @inside1024) ▸ Compute pricing separates the winners. One user benchmarked subscriptions by infrastructure cost: Claude delivered roughly 5x the value of OpenAI's plans, while a ¥500 ByteDance Agent Plan burned half its monthly video quota in under two days, with no prompt caching for plain language models. On the gray side, DGX servers with eight Blackwell chips list at $400–500K and trade at roughly double on the black market, where H100s at $40K are still bought on sight. (source: @aigc1024, @inside1024) Also spotted: OpenDirector bundles AI video generation, asset management, image editing, and timeline editing into one local browser workspace that exports MP4 (@https1024); browsebrowse hands an already-logged-in Chrome tab to an agent via an MCP server plus extension, returning control for 2FA and payments (@https1024); Grok bot users reported sharp quality drops and hallucinations, which Musk says should improve (@aigc1024). The bottleneck keeps moving: from model capability, to agent coordination, to the price of a single token.