Steve Yegge, one month into running 20+ concurrent Fables, says his job has compressed to asking, credentials, taste, and comms — the execution is gone. The same week, Anthropic and Blackstone launched Ode, a $1.5B firm whose forward-deployed engineers embed inside enterprises and build custom AI systems on top, with Deloitte and McKinsey nowhere in the partner list. The gap between those two facts is the bind: the residual human role is shrinking to judgment, and the people who move judgment at scale are about to own the implementation layer.
That's what's actually happening underneath both stories. Anthropic is betting the leverage compounds at the firm level — embed the engineers, map the operations, capture the margin that consultancies currently extract. xAI is betting the leverage compounds at the runtime level — open-source the harness, reset all limits, give the infrastructure away and own the ecosystem. The first strategy ends with a new kind of professional-services firm that the incumbents can't easily replicate. The second ends with a platform that makes every other harness a cost center. Those two bets don't converge.
Top developments
Thinking Machines — dropped Inkling, its first release: a 975B open-weight multimodal model reasoning across text, image, and audio, trained from scratch without distilling from OpenAI or Anthropic, and fine-tunable today on Tinker (Bridgewater already a customer). First serious American open-weight frontier planted opposite Kimi, GLM, and Ling. thinkingmachines.ai
xAI — Elon posted "Grok Build is now open source" and SpaceX confirmed the entire CLI (1.3M lines of Rust) is out, with all usage limits reset for all users. Lands one week after Grok Build's own repo-upload scandal, and alongside a quietly-shipped official Claude Code plugin from the same xai-org account. The harness war's most aggressive move: give the runtime away. x x
Anthropic, Blackstone, Goldman Sachs — launched Ode, a $1.5B AI services firm whose forward-deployed engineers embed inside enterprises, map operations, and build custom AI systems on top, pipelined from PE portfolio companies. A frontier lab plus top-tier bank crossing into implementation — Anthropic going straight at Deloitte/McKinsey turf, and Blackstone underwriting it. x x
Neko Health — announced a $700M Series C for its full-stack preventive care play (medical devices, sensors, AI, clinics run in-house), with 100K+ people already scanned and a US launch starting in NYC. Second $500M+ health-AI round in a week after Chai Discovery's $400M — the applied-AI capital wave is moving from drug design into diagnostics. x
Notable discussions
What's left of the engineer when Claude writes 90% of the code — Dario Amodei says the remaining engineers become 10x productive on the remaining 10%; Steve Yegge, one month into running 20+ concurrent Fables, says his job compressed to asking, credentials, taste, and comms; Santiago Valdarrama asks pointedly why Anthropic is still hiring software engineers if agents are that good; Cursor's answer is $1.1M packages for engineers who manage agent teams. The debate has moved from "will AI replace developers" to what specifically the residual role pays for. x x x x
Linus Torvalds — publicly puts his foot down as the top-level Linux maintainer: kernel is not an anti-AI project, AI is "clearly a useful tool," and he will "loudly ignore" anyone arguing others shouldn't use it — anyone with issues can fork or walk away. The most authoritative pro-AI declaration yet from a maintainer who could credibly have banned the tools. x
Theo — landed changes that cut T3 Code's GPU utilization by ~85% that neither Fable nor GPT-5.6 could find; both LLMs kept getting distracted by code that wasn't even in the render path. Sharpest recent counter to the just-hand-it-to-the-agent narrative: today's models still miss whole-system reasoning even when handed the tools they need. x
Jarred Sumner — Bun's founder expects fully automated merges within months: automated reviewer LGTMs → a fresh-context Claude judges simple + low-blast-radius → auto-merge with no human at the last step. Bun's public repo already shows bots reviewing bot PRs; the CEO is now sketching the roadmap out loud. x
Simon Willison — flips the veteran-engineer holdout: if you know how to write the code, you gain nothing by doing the typing yourself — outsource that to a coding agent. Sharpest single-line rebuttal to the "faster to just write it myself" position. x
Pedro Domingos — the singularity is not coming: recursive self-improvement gives diminishing returns. Direct dissent to yesterday's Zhengyao Jiang RSI Top dev and today's Weco autoresearch amplification — the RSI narrative now has a credentialed AI-scientist counter-voice. x
Other news
Models & releases
Ling-2.6-1T — China ships a 1T open-source model, 72.2% on SWE-bench Verified, 256K context, free on OpenRouter, plugs directly into Claude Code x
Soofi S — Germany's Soofi consortium ships a 30B fully-EU-trained open-source model, top-ranked in Europe x
Google Gemma 4 — community-driven update lands with broad quality fixes x
Kimi K3 — outperforms Opus-4.8 on a Flappybird agent test and beats other open-source in coding tasks x
Aster Inference — new provider claims fastest GPU inference at 644 tokens/sec x
Cerebras GPT-5.6 — 750 tok/sec on GPT-5.6 x
NVIDIA Cosmos 3 Nano — 93% accuracy with minimal post-training x
Google vector-index shrink — 31GB→4GB open-source AI memory index x
OvisOCR2 — 0.9B parameters, SOTA document parsing x
Claude Opus 5 leak — pankajkumar_dev reports next-week target, capabilities closing on Fable 5, at significantly cheaper inference x
Devtools & coding agents
OpenAI Codex Micro kbd-1.0 — $230 physical desktop peripheral with buttons + joystick built with Work Louder for Codex users x
DoorDash dd-cli — limited-beta CLI for agents to search stores, find deals, and check out food orders x
Anthropic 4-hour agentic-systems course — free full curriculum drops x
Anthropic 37-minute workshop — building real AI agents x
Anthropic Claude Code prompt library — official ready-to-use coding task templates code.claude.com
Anthropic 2026 agent roadmap — tools, memory, and observability features x
Claude Code artifacts + MCP connectors — artifacts now support MCP for dynamic dashboards and apps x
GitHub Copilot for Jira — generally available with agent-progress streaming and follow-up instructions github.blog
ReactBench — new benchmark for how well AI models write React code x
Perplexity SPACE — production sandbox platform for computer-agent traffic research.perplexity.ai
Willow — free unlimited AI dictation on iOS with a custom keyboard x
Playwright in Rust — rewrite achieves 2.55x speed, 70% memory reduction x
xAI Claude Code plugin — shipped quietly from the xai-org account five days before Grok Build open source, no announcement github.com
Infrastructure & platforms
Anthropic — resets 5-hour and weekly rate limits for all Claude users x
Vercel Sandbox — 3.5M daily creations at 100% monthly growth x
Vercel AI Gateway — opens token-flow dataset for AI usage analysis vercel.com
YC GPU compute forward curves — futures pricing for B200, H200, A100 from prediction markets x
Funding & deals
Cursor — $1.1M/year packages for engineers managing AI-agent teams shipping code autonomously x
Miles Wang (ex-OpenAI) — launches AI drug-discovery startup targeting $2B valuation, per Techmeme x
Mio — $2.2M seed for AI hiring assistant x
Yann LeCun / AMI Labs — releases 4-hour interview elaborating on the $1.03B post-LLM thesis x
Industry & policy
OpenAI ChatGPT + Codex merger — apps and MCP marketplaces unified into a single desktop app; explains part of Codex's 8M-user jump manufact.com
Anthropic vs the consultancies — Ode's forward-deployed engineers positioned to bypass Deloitte/McKinsey/Accenture; none appear as partners x
Chinese AI labs — Deedy's tally: massive funding and capabilities across the Chinese frontier x
OpenAI on Apple — denies stealing trade secrets but declines a detailed investigation x
Continuing threads
Bonsai 27B (cont. from 07-15 tail) — coverage continues on the 93%-size-reduction 27B multimodal running on RTX 5090 and iPhone x
Karpathy + Obsidian second brain (cont. from 07-13 & 07-14) — MyWestLord's Obsidian + Claude Code combo continues to circulate x
Anthropic rate-limits concession (cont. from 07-12 Top dev) — the trading of usage caps for developer share continues; today's move undoes the 5-hour and weekly limits entirely x
Recursive self-improvement debate (cont. from 07-14 & 07-15) — Weco's agent rewriting its own harness in 8 days amplifies yesterday's Zhengyao Jiang result, drawing Domingos' diminishing-returns pushback (see Sharp takes) x