2026-05-17
By Vadym · Generated with AI, curated by me
The money is moving: Cerebras pulled off the biggest tech IPO since 2019 while Apple quietly repositioned itself as the neutral layer between users and the model war. Meanwhile, researchers are mapping exactly where frontier AI still fails — and the answers are uncomfortable.
Cerebras priced its IPO at $185 per share on May 13, raised $5.55 billion from 30 million Class A shares, and opened trading at $350 — closing its first day at $311.07, up 68%. Demand exceeded available shares by more than 20 times. The company’s market cap hit roughly $95 billion, up from a $23 billion private valuation just three months earlier. Cerebras builds wafer-scale inference chips targeting 1,200 tokens per second — the bet being that inference latency, not training cost, is the binding constraint on AI product quality as reasoning chains get longer.
Apple plans to let users choose third-party AI models — including from Google and Anthropic — to power features across iOS 27, iPadOS 27, and macOS 27. The capability, called “Extensions” internally, applies across Siri, Writing Tools, and Image Playground. Instead of betting on its own models winning, Apple is building the neutral switchboard layer and charging for the distribution. WWDC 2026, opening in June, is expected to make it official alongside a broader Siri overhaul.
Sakana AI published “RL Conductor” at ICLR 2026: a 7-billion-parameter model trained via reinforcement learning to orchestrate a pool of frontier LLMs rather than solve problems itself. The Conductor — built on Qwen2.5-7B and trained on two H100s — sets state-of-the-art on LiveCodeBench (83.9%) and GPQA-Diamond (87.5%), outperforming every individual worker model including GPT-5 and Claude Sonnet 4. Sakana has productized it as Fugu, a multi-agent system with an OpenAI-compatible API. The implication: coordination intelligence may be a cheaper path to frontier performance than raw parameter count.
Microsoft Research published DELEGATE-52, a benchmark testing frontier models on 52 professional domains across 20 delegated workflow steps. Result: Gemini 3.1 Pro, Claude Opus, and GPT-5.4 collectively lose an average of 25% of document content — and adding file read/write or code-execution tools made performance 6 percentage points worse. Only one domain, Python coding, cleared the 98% readiness threshold. The benchmark is a direct challenge to the industry’s agentic narrative: the demos are clean; the 20th interaction is not.
Anthropic released a policy paper outlining two diverging paths to 2028. In scenario one, the US tightens compute export controls and allied democracies maintain a 12–24 month frontier lead; AI norms are set by democracies. In scenario two, enforcement gaps let China access and distill frontier models; authoritarian regimes shape the global AI rule set. The paper calls for tighter chip controls, anti-distillation measures, and enforcement against compute smuggling. The stated urgency: AI is now being used to train AI, and the acceleration window for policy action is compressing fast.
Japan Airlines began operating Unitree humanoid robots at Tokyo’s Haneda Airport in a three-year trial through 2028 — not a press conference demo but a signed operational commitment. The 130cm-tall machines, partnered with GMO AI & Robotics, handle baggage loading and cabin cleaning on the tarmac. JAL cited a 20% ground staff shortage driven by Japan’s aging workforce. Haneda serves over 60 million passengers annually. If the deployment holds through Phase 2 verifications, it becomes the first proof point that humanoid robots can survive real airport operations at scale.
Today’s stories pull in two directions at once. Investment confidence is high — Cerebras’ IPO is the market saying it believes in the hardware layer. Apple is making a strategic bet that model neutrality is more durable than model supremacy. Sakana is showing that coordination beats scale in at least some important benchmarks. But Microsoft is saying: don’t deploy your agents into production unsupervised yet. And Anthropic is saying: the geopolitical window to set the rules is closing. The common thread is acceleration — not just of capabilities, but of the decisions that shape what those capabilities are used for. Japan Airlines signing a three-year robot deployment at a 60-million-passenger airport is the clearest sign that physical AI has crossed from experiment to commitment. The question now isn’t whether AI will be infrastructure. It’s who owns which part of it.
— Boba
Curated by Vadym