Gary Marcus reiterated this week that OpenAI will end up viewed as the WeWork of AI — and then Altman announced Codex crossed 8M users on a 2.5x week-over-week jump, with inference capacity already falling behind demand. Both data points are real. Both landed on the same day. The gap between them is not a disagreement about facts; it is a disagreement about what the facts mean.
The bull read is that 8M users and a JetBrains default partnership is compounding distribution into a moat. The bear read is that Vorflux just raised $15M on the claim that Codex-class models are already good enough to fly the whole route — making the agent layer a commodity and the platform margin a bet on switching costs that don't exist yet. One camp is pricing in the growth curve. The other is pricing in the structural trap underneath it: when the model is the product and the model is replicable, the 8M-user number is a ceiling story, not a floor. Those two reads don't end in the same place.
Top developments
OpenAI — Codex + ChatGPT Work crossed 8M active users (up from 1M five months ago) with a 2.5x week-over-week jump per Altman, JetBrains AI named Codex its default recommended agent, and Nico Albanese joined to build the Codex app. Altman warned inference "hiccups" are coming — Codex is now consuming compute faster than OpenAI can pour it in. x x x
Vorflux — Rippling co-founder Prasanna Sankaranarayanan launched Vorflux on a $15M seed led by YC and Peak XV, with Parker Conrad, Balaji, Nivi, and Matt Shumer (who says it "blows Devin out of the water"). Positioned as autopilot rather than copilot for coding — the founder claim is that models are already good enough to fly the whole route. x x
Anthropic — launched Claude for Teachers: free premium Claude and Cowork for verified US K-12 educators, with a library of teaching skills and evidence-based curricula mapped to standards in all 50 states. Anthropic now runs first-party surfaces for Finance, Science, and Teachers — verticalizing Claude the way OpenAI verticalized ChatGPT via Sites. claude.com
Chai Discovery — closed a $400M Series C led by Index Ventures with Kleiner Perkins, Sequoia, and Dimension Cap. Founder Jack Dent told the NYT the pitch is that AI-driven drug design is "moving from a promise to actual deployment." Largest AI-native drug-discovery round of the quarter — capital is now following applied AI into pharma at frontier-lab scale. nytimes.com
Zhengyao Jiang — published what he calls the first experimental evidence of recursive self-improvement: an autoresearch agent, pointed at autoresearching itself for eight days, produced a variant that beats the harness his team hand-tuned for two years on held-out benchmarks. Reads as the direct counter to yesterday's Skyfall Morpheus claim that LLMs can't continually learn. x
Notable discussions
AI-coding-stack security follow-ups — Matt Johansen dropped a Cursor 0day after 200 days of ignored disclosure; hyusapx showed Claude(.)ai leaking his full name, employer, and bank security answers to an attacker via a benign coffee-shop question; Matthew Green called Anthropic's related data-collection posture "gross" the same day. On the heels of last week's Grok CLI repo-upload breach, the AI-coding wire is breaking on a weekly cadence. x ayush.digital x
SDK and dashboard death — Factory publicly banned client SDKs in favor of agent-generated purpose-built clients per endpoint; Resend spun up a "Headless Dashboard Initiative" so every dashboard action is available via API/SDK/MCP; WorkOS declared dashboards "going away" with its Management MCP already among its fastest-growing features. When the consumer is an agent, the vendor UI is dead weight. x workos.com workos.com
Demis Hassabis — publishes a governance framework arguing AGI is "closer to the discovery of electricity or fire" than to the internet, will land at roughly 10x the Industrial Revolution at 10x the speed, and needs regulatory infrastructure built inside the current window before agentic self-improving systems arrive. The Nobel laureate who spent years correcting inflated timelines is now on the act-now side. x
Jason Lemkin — will no longer invest in agentic B2B SaaS unless the product unlocks materially more budget than traditional software. Agentic ROI means 2–10x deal sizes over the pre-AI incumbent, not just stealing incumbent budget — making a legacy category 10x better with AI wins customers but traps you at 2x ARR at scale. x
Marc Andreessen — describes a "three-way Mexican standoff" between coders, PMs, and designers at leading Valley companies, where each role thinks AI lets them absorb the other two — and predicts all three are correct, with the resulting job title converging on something like "builder." x
Gary Marcus — reiterates his 2023 call that OpenAI will end up viewed as the WeWork of AI, arguing this week's 8M-user and infra-strain victory laps are propping up storytelling that keeps unwinding underneath. The one credible bear voice cutting against today's OpenAI-momentum consensus. x
Mitchell Hashimoto — ran a read-only ChatGPT automation over three years of construction invoices; it flagged ~$45K in wrong amounts, duplicates, and mis-addressed bills against a ~$1,800 lifetime ChatGPT spend — 25x lifetime ROI on one task. Sharpest concrete counter to the "$80/hour agents" discourse: the constraint isn't the hourly rate, it's what decision you point it at. x
Other news
Models & releases
Bonsai 27B — first 27B-class model to run on a phone, Apache 2.0, ternary (5.9GB) and 1-bit (3.9GB) builds retain ~90% of Qwen3.6 27B quality x
Kimi K3 — launching tomorrow with API top-up discount per briefly-live docs page x
Zhipu GLM-5.5 — reportedly preparing 1T-parameter release for August x
Hydra S2S — always-on speech-to-speech model reasoning over latent vectors at millisecond cadence, not tokens x
NVIDIA ARDY — real-time open-source AI animation with weights/code/demo dropped; fits in 20–24GB VRAM and drives their humanoid robots research.nvidia.com
NVIDIA Molt — new RL framework where the entire agent lives in one Python file and scales from 8B to DeepSeek-V3-class MoE on the same command github.com
NVIDIA autoresearch coding agent — took Qwen3-VL-2B from 25% to 96.9% accuracy on a training environment it built itself, then proposed the next experiment developer.nvidia.com
Devtools & coding agents
CLAUDE.md — single markdown file derived from Karpathy's coding rules hits 192K GitHub stars github.com
Screenpipe — open-source local-first work-recording + agent memory crosses 20K GitHub stars, 1,900+ forks (also reposted by YC) x
Perplexity WANDR — open-sourced its internal research-capability benchmark research.perplexity.ai
GitHub Copilot CLI — 23% fewer tool failures on the latest reliability drop github.blog
Claude Code Desktop — ships native WSL support with a distro picker x
Browser Use v4 API — full agents driving logged-in profiles/proxies with a single request; Claude Code agents can now place phone calls for bookings and lead chasing monid.ai
Anthropic Claude Code course — free 59-minute drop from the Claude team x
Reflect — the note-taking app went open-source with a full rewrite for AI-first note-taking x
Infrastructure & platforms
Vercel Compute — 30-minute max function durations, Dockerfile support, up to 5GB functions, FUSE x
Vercel AI Gateway — model usage data released under Creative Commons for open analysis vercel.com
Open-weight share — open-weight models now serve 29% of Vercel AI Gateway tokens, up from 11% in April x
1Password — moves into AI cost management, framing token spend as the next enterprise budget crisis venturebeat.com
Funding & deals
Nous Research — raising $75M+ at $1.5B led by Robot Ventures for the Hermes open-source agent platform (Techmeme) x
Greylock 18 — new $1.5B early-stage venture fund announced by Saam Motamedi x
Promptwatch — $6M seed to capitalize on AI-driven search transformation x
Retym (Noah Schochet) — $115M Series A for AI-native American manufacturing x
AI startups at $500M+ revenue — Deedy's updated roster x
Industry & policy
OpenAI hardware — Mark Gurman reports first product is a mobile, screen-free home smart speaker built for AI-companion attachment bloomberg.com
Robinhood — Vlad Tenev going all-in on agentic AI to give retail the same trading tools institutions get x
Chamath / 8090 — first case study: Software Factory decoded a 50-year, 18M-line COBOL + Assembly repo governing billions in CMS Medicare payments into 100K+ plain-English business rules traced to source x
Brex — using just-bash agents deep in production financial systems, per Malte Ubl x
Meta — allegedly used AI to target workers with medical conditions in layoffs x
IBM — stock down 20% after weak software and infrastructure earnings x
Richard Sutton — launches Oak Lab targeting a trillion-parameter agent that learns in real time on 20 watts, built on "batch size one" instead of pretraining x
Continuing threads
Anthropic Loop Engineering course (cont. from 07-14 Notable) — additional coverage this cycle, still the free Fable 5 curriculum x
Own-your-AI-stack (cont. from 07-13 Notable) — YC-amplified "startups should own their intelligence" thread continues x
Tom Blomfield → Anthropic (cont. from 07-13 tail) — Monzo co-founder confirmed onboard to build AI-native companies x
Anthropic Fable 5 access (cont. from 07-13 Top dev) — Cognition's Walden Yan and others continue to report Fable proves cheaper than Opus inside a mixed harness x