Agentic Brew Daily
Your daily shot of what's brewing in AI
Fresh Batch
Meanwhile OpenAI shipped Codex inside the ChatGPT mobile app on every plan (including Free), hired outside counsel to evaluate suing Apple over the Siri integration, and watched its own developers post degradation complaints about GPT-5.5 the same afternoon Greg Brockman called the mobile launch a "huge step forward." In Oakland, the Musk v. Altman trial wrapped closing arguments — jury deliberates Monday, while Musk himself was in Beijing with Trump's delegation. A lot of plotlines crossed each other this week.
Bold Shots
Today's biggest AI stories, no chaser
Cerebras priced its IPO at $185 (above the $150-$160 marketed range), raised $5.55B, and opened Wednesday at $350 on Nasdaq — the largest US semiconductor IPO ever. Market cap briefly cleared $100B before closing at $311 with a $66B valuation, then sold off about 10% Thursday as analysts including Jim Cramer balked at a ~187x trailing-sales multiple (Nvidia trades at 26x, AMD at 21x). Orders ran 20x oversubscribed; co-founders Andrew Feldman (~$3.2B) and Sean Lie (~$1.7B) became billionaires on paper.
Why it matters: Reopens the AI IPO window after years of drought, but the customer concentration is the story most coverage glossed over — MBZUAI was 62% of 2025 revenue, G42 was 24%, and OpenAI is on the hook for a ~$20B / 750 MW capacity deal. That's a $66B valuation resting on roughly three relationships.
OpenAI rolled Codex into the ChatGPT mobile app on iOS, iPadOS, and Android — preview, but available on every plan including Free and Go. From your phone you can start tasks, scrub terminal output, review diffs and test results, approve commands, and switch models. Pairing uses a secure relay; mobile only talks to the macOS Codex desktop for now (Windows "coming soon"). Same release: Codex Hooks GA, programmatic access tokens for CI, and a HIPAA-compliant Codex for eligible enterprise local environments. Weekly Codex users are now 4M+, double the 2M reported in March.
Why it matters: Anthropic shipped Remote Control for Claude Code in February, so OpenAI is roughly four months late — but it's now on every tier including Free, which Claude Code is not. The product thesis is fire-and-forget delegation: agents run in cloud sandboxes, you approve from your phone. The Windows-host gap plus HIPAA + CI tokens signals where the enterprise battle is actually being fought.
OpenAI has retained outside counsel to weigh a breach-of-contract notice against Apple, claiming Apple never made an "honest effort" to surface ChatGPT inside iOS — Siri, Writing Tools, and Visual Intelligence shipped in iOS 18.2, but subscription revenue is "nowhere close" to the billions projected. The kicker: iOS 27's new Extensions API will let users route Siri queries through Claude or Gemini, ending ChatGPT's de facto exclusivity. Apple separately pays Google ~$1B/year for a custom 1.2-trillion-parameter Gemini for next-gen Siri. OpenAI has been paid $0.
Why it matters: This is default-button economics at iPhone scale. Apple writes Google a billion-dollar check and gives OpenAI shelf space; OpenAI poaches 40+ Apple engineers via Jony Ive's hardware unit; now Apple opens the door to Claude and Gemini routing. The NYT read is that the lawsuit threat is a bargaining chip, not a real courtroom move — but the asymmetry is the actual story.
Closing arguments in the OpenAI breach-of-charitable-trust trial concluded Thursday in Judge Yvonne Gonzalez Rogers's Oakland courtroom. Two claims remain (down from 26): breach of charitable trust and unjust enrichment. The 9-person jury (6F/3M) starts deliberations Monday, May 18 — but the verdict is advisory only. Judge Gonzalez Rogers keeps final liability authority, and the remedies phase starts Monday concurrently. Musk skipped his own closing arguments to travel to Beijing with Trump's delegation. OpenAI's lawyer Sarah Eddy: "Mr. Musk isn't here today — my clients are here."
Why it matters: A finding for Musk could force the removal of Altman and Brockman, disgorge up to $134B back to the nonprofit, and unwind the October 2025 OpenAI Group PBC recap. Microsoft's ~27% / ~$135B stake is exposed on the same hook. Even a narrow finding gives the judge statutory tools under California charitable trust law to unwind the deal.
Trump's 36-hour Beijing state visit ended Friday with no signed AI governance framework, zero Nvidia H200 chips shipped to the ten approved Chinese buyers (Alibaba, Tencent, ByteDance, JD.com, Lenovo and others), and rare-earth exports still ~50% below pre-restriction levels. The US delegation included Jensen Huang (added last-minute in Anchorage), Tim Cook, Elon Musk, Sanjay Mehrotra, and Dina Powell McCormick. Under the January 2026 Commerce rule, Nvidia would have to remit 25% of those H200 sales to the US government — but Beijing is steering buyers to Huawei's Ascend 950PR, which Huawei claims delivers 2.8x the FP4 performance of the banned H20.
Why it matters: The binding constraint on US-China chip flows just flipped. It used to be US licensing; now it's Chinese willingness to buy. Huawei is targeting $12B in AI chip revenue this year (up from $7.5B in 2025), and if 1M H200s actually flowed it would add +250% to China's AI compute. The 25%-revenue-share framework is genuinely novel — CFR called it "strategically incoherent and unenforceable."
The Blend
Connecting the dots across sources
The agentic coding fight is now a three-vendor mobile race
- OpenAI shipped Codex inside the ChatGPT mobile app on every plan including Free, then formally reorganized with Greg Brockman owning all products and Codex builder Thibault Sottiaux running the unified ChatGPT/Codex/API platform.
- Anthropic answered the same week by leasing 220,000+ GPUs from xAI's Colossus 1 cluster to handle 80x usage growth, while xAI itself launched Grok Build CLI beta scoring 70.8% on SWE-Bench Verified.
- On GitHub the top three movers today are all skills repos — mattpocock/skills, obra/superpowers, and K-Dense-AI/scientific-agent-skills — and Skills.sh shows find-skills at 1.5M installs while Pragmatic Engineer's lead blog this week is literally titled "Did capacity shortages turn Anthropic hostile to devs?"
Anthropic is authoring the US-China narrative across every channel at once
- Dario Amodei's "2028: Two Scenarios for Global AI Leadership" paper is the most upvoted research item of the day on X, arguing for tighter export controls to preserve a 12-24 month US lead.
- The same framing shows up verbatim in the Trump-Xi summit coverage (zero H200s shipped, 25% revenue-share rule), in Wes Roth's viral thread, and in The Neuron's "AI Cold War got a protocol" blog the same week.
- The counter-evidence isn't quiet either — Stanford's Alvin Wang Graylin called the framing "irresponsible," Reddit treats it as regulatory capture, and a senior Anthropic researcher reportedly left over the adversarial framing. One company is setting the terms regardless.
The $725B capex / 102K layoffs paradox finally has a public face
- Cisco posted a record $15.8B Q3 quarter with AI orders guidance raised to ~$9B and the stock at an all-time high — while quietly cutting nearly 4,000 jobs in the same week.
- On X, a viral post pointing out that $700B was spent on AI data centers the same year entry-level software developer employment hit a five-year low landed alongside Sanders and AOC's bill to pause US AI data center construction.
- Goldman Sachs argues it's earnings-led, not a bubble, and that liquid cooling is the next leg — which is precisely what Cisco's networking supercycle looks like from the supplier side. Both bulls and bears are now arguing from the same number.
Slow Drip
Blog reads worth savoring
Walks through Anthropic's 80x usage surge, the Claude Code nerfs, paid-access revocations, and the quiet xAI/Colossus 1 GPU lease keeping it all up.
AI-generated code is now nearly half of commercial output and creating "comprehension debt" — teams feel 20% faster while shipping 19% slower.
Practical PM playbook with the CLAUDE.md router pattern and Dispatch tips for mobile multitasking.
Working code for sub-500ms speech-to-speech with function calling, barge-in, and multilingual support on Bedrock.
Quiet-day digest spotlighting the agent-first IDE wars (Conductor vs GitHub), Figure's 24-hour autonomous robot run, and LangChain's self-improving SmithDB loop.
Dense roundup covering Recursive Superintelligence's $650M raise, Nous's 2-3x faster pretraining trick, NVIDIA's elastic reasoning, and Anthropic's $30B ARR roadmap.
97M-parameter embedding model that beats every sub-100M open model on MTEB Multilingual Retrieval by 9+ points, with 200+ languages, 32K context, Apache 2.0.
Red-team test exposes a time-delayed payload attack that exploits "trusted skill" reputation to bypass on-device safeguards.
Concrete n8n + Jotform + GitHub + Vercel recipe for shipping multi-step AI website updates with human-in-the-loop preview and post-deploy monitoring.
The Grind
Research papers, decoded
Anthropic's CEO sketches two futures for 2028: one where the US and allied democracies hold a decisive lead in frontier AI, one where authoritarian states catch up. The argument: the next 18 months of compute access, export controls, and energy build-out will decide which it is. Whether you buy the framing or not, this is the document everyone in the policy conversation is now responding to.
A diffusion language model that stays in continuous embedding space until the very last step, when a single shared-weight network maps it to discrete tokens — a sharp contrast to today's leading diffusion LMs that operate over tokens throughout. Because it lives in continuous space, ELF can borrow proven image-diffusion tricks like classifier-free guidance and hits stronger quality with far fewer sampling steps and roughly 10x less training data (~45B tokens vs 500B+). Early but credible signal that continuous diffusion may become a real text-generation paradigm where inference cost matters.
LC-MAPF is a pre-trained, decentralized model that lets large fleets of agents (warehouse robots, search-and-rescue swarms) plan collision-free paths by exchanging short messages with their immediate neighbors over multiple rounds. Unlike prior learning-based pathfinders that either ignore communication or fail to scale once agents talk, LC-MAPF holds its coordination gains as fleet size grows and generalizes to unseen maps. A drop-in policy for multi-robot logistics, last-mile delivery, or game-AI swarms that improves on imitation- and RL-based MAPF baselines without sacrificing scalability.
On Tap
What's trending in the builder community
"Skills for Real Engineers. Straight from my .claude directory." The top mover on GitHub today.
Turns commodity WiFi signals into real-time spatial intelligence, vital sign monitoring, and presence detection — no cameras involved.
Agentic skills framework and software-development methodology that's now the #1 skills repo overall at 192K stars.
Ready-to-use Agent Skills for research, science, engineering, analysis, finance and writing.
AI meeting companion with cross-meeting memory — the productivity hit of the day on Product Hunt.
AI sleep companion designed to help you fall asleep without struggle. Hardware + AI combo.
AI platform that hands any task off to a vetted human expert when the agent gets stuck.
Build on Notion, not just inside it — also the subject of a SF hackathon this weekend.
Working session on trace analysis, failure categorization, and code-based + LLM-as-judge evals for agent systems.
Deep dive on activation functions from Heaviside through ReLU to GELU and SwiGLU.
The robotics data bottleneck explained through a "data pyramid" framework — why embodied AI hits a wall before it scales.
Rigorous benchmark of speculative decoding versus MoE offloading on consumer GPUs.
"Codex team is aware of reports of GPT-5.5 performing worse for some users and investigating. We don't have anything conclusive yet and systems are healthy but we will share updates as we go." Same day as the mobile launch.
Launching personal finance in ChatGPT for Pro users with bank-account connections via Plaid: "Help me save money is one of the core benefits people hope to get out of AI."
"Reporter: Did the Nvidia H200 advanced AI computer chips come up with China? President Trump: 'It did come up... They have a much higher level than H200... China needs it and so yeah it came up.'"
"What happens when the smartest AI models become too expensive for most people? Do we end up with consumer AI for everyone else and $2000/mo frontier models only for power users and companies"
Discover and install skills from the open agent skills ecosystem — the directory that launched a thousand directories.
Distinctive, production-grade frontends that reject generic AI aesthetics. Second on Skills.sh by installs.
70 rules across 8 categories for React/Next.js performance optimization.
Captures learnings, errors, and corrections across sessions — currently #1 on ClawHub.
Roast Calendar
Upcoming events & gatherings
Two-day hackathon shipping on Notion's new Developer Platform primitives (data sync, agent tools, workflow triggers) with just a CLI — perfect if you're wiring agents to tools.
Premier AI builders hackathon (314+ interested) spanning SF Bay, Singapore, and Tokyo — one of the largest agent-building gatherings of the weekend.
Full-stack AI hackathon hosted by AI Collaborate at Santa Clara University; good fit for students and early-career devs to ship something end-to-end.
Casual data + dim sum meetup focused on self-improving analytics agents (JetBrains Databao); high signal for data + agent practitioners.
East Bay session reframing AI compliance and governance as a moat — relevant for founders and execs navigating the new regulatory landscape.
Pickleball-meets-pipeline mixer hosted by databar.ai for GTM and data folks who need to stand up and stop staring at dashboards.
Knowit Owlz Web3 + AI social with 76+ interested; good casual networking at the intersection of crypto, DeFi, and AI.
Last Sip
Parting thoughts & a teaser for tomorrow
The through-line of the week is that the surface area of AI keeps expanding — phones, courtrooms, charitable trust law, Beijing state dinners, Wall Street comps — but the actual choke points keep shrinking. Cerebras is a $66B company resting on three customers. Anthropic's growth depends on a GPU lease from a guy publicly feuding with one of its largest investors. OpenAI's mobile launch hit the same day as a 503K-view post saying GPT-5.5 got worse. Monday brings the Musk v. Altman jury, which is technically advisory but practically anything-but. We'll see you back here with whatever they decide.