Notes RSS feed

The Wire — short items from the day's overflow, verdict follow-ups on earlier calls, and changelogs for the paper's maintained data assets. When a story earns a full argument, it runs as an article instead.

Wire

Waymo gives Gemini three screens in Ojai
07/30/26, 12:25 PM

Waymo put Gemini into the Ojai robotaxi’s three-screen cabin, where the assistant can manage climate, answer journey questions, and request a pullover while remaining isolated from vehicle movement, routing, and real-time driving data. The company says Gemini in Waymo is in beta and inactive until a rider engages it. For teams building physical AI, the boundary complements simulation’s 20-to-1 screening funnel for robot trials: conversational convenience can ship inside an autonomous system without granting the general assistant control of its safety-critical stack.
Verdict

SK hynix turns its HBM4 promise into shipments
07/30/26, 12:24 PM
Follows up on Nvidia's $500B Memory Deal Is Still an LOI

SK hynix began mass shipments of HBM4 in the second quarter as revenue reached a record 79.3 trillion won, up 257% year over year, according to its second-quarter results. Our analysis of Nvidia and SK’s memory pact argued that the headline letter of intent did not guarantee HBM volume; today’s evidence shows the underlying product has crossed into shipments, but SK still disclosed no Nvidia allocation or take-or-pay quantity. The verdict therefore moves one notch toward execution while holding on contract certainty: builders can treat HBM4 as shipping silicon, not as proof that Nvidia’s promised supply is reserved.
Wire

EU opens €10B race for seven AI gigafactories
07/30/26, 11:21 AM

The European Union opened bidding for up to seven AI gigafactories with €10 billion in public funding, and AP reports that each planned site will hold at least 100,000 advanced AI chips. The official EuroHPC call closes November 12 and is designed to unlock more than €20 billion in private investment. For operators tracking Europe’s sovereign-compute buildout, the subsidy creates a procurement lane worth watching, but electricity costs and a multi-year construction schedule keep it from solving today’s capacity gap.
Wire

Grok Voice reaches first audio in 0.70 seconds
07/30/26, 11:18 AM

xAI’s Grok Voice Think Fast 2.0 reaches first audio in 0.70 seconds and scores 82.9% on Artificial Analysis’ speech-to-speech quality index, according to xAI’s launch benchmarks. The live API pricing page lists the model at $0.08 per audio minute, or $4.80 per hour, and xAI says the grok-voice-latest alias will move to it automatically on August 5. Teams comparing the release with Grok’s broader price-performance strategy should pin the old model now if they need a stable regression baseline; the new endpoint is 60% dearer than Think Fast 1.0 even as latency falls 44%.
Wire

Waymo returns to freeways after 13 failures
07/30/26, 10:27 AM

Waymo began restoring freeway routes in Phoenix after software fixes responding to at least 13 construction-zone incursions and a recall of nearly 4,000 robotaxis. The company’s restart, reported by TechCrunch, adds scene-recognition and routing updates after more than two months of restrictions; Los Angeles and the San Francisco Bay Area are due to follow. For robotics teams using simulation as a 20-to-1 screening funnel before scarce hardware trials, Waymo’s sixth recall is the reminder that deployment still needs monitored edge-case evidence, not only aggregate autonomy miles.
Wire

Two harness settings triple OpenAI's ARC score
07/30/26, 10:25 AM

OpenAI lifted GPT-5.6 Sol’s public ARC-AGI-3 score from 13.3% to 38.3% while using six times fewer output tokens by retaining reasoning between turns and replacing rolling truncation with compaction. The company’s harness analysis says the generic runner discarded private reasoning and dropped old actions after 175,000 characters, forcing the model to relearn each game as it played. For teams confronting the enterprise shift from token volume to outcome efficiency, the result is a sharp warning: benchmark scores and agent costs can measure the harness as much as the model.
Wire

Samsung chips post a 19-fold AI profit jump
07/30/26, 10:23 AM

Samsung Electronics’ quarterly operating profit rose more than 19-fold year over year to KRW 89.5 trillion, with its semiconductor division supplying KRW 89.2 trillion as AI-server memory demand pushed prices and shipments higher. The company’s second-quarter results also say Samsung shipped the industry’s first HBM4E samples to major customers and expects the memory market to remain undersupplied despite production increases. For operators tracking why next-generation AI memory is moving into long contracts, the takeaway is that HBM scarcity is no longer just a forecast—it is already dominating supplier economics.
Wire

Microsoft folds four Copilots into one app
07/30/26, 09:26 AM

Microsoft will fold four surfaces—Copilot chat, GitHub Copilot, Copilot Cowork, and Autopilot—into one AI super app spanning consumer and commercial use this year. The Verge’s report on Satya Nadella’s earnings-call confirmation says the consolidation follows a quarter in which Microsoft revenue reached $90 billion, with AI and cloud driving the increase. For operators assessing Microsoft’s model-routing control plane, the implication is distribution: one front door can make the underlying model and agent boundaries disappear from users before the economics underneath have stabilized.
Wire

White House puts AI-biology risk under review
07/30/26, 09:24 AM

The White House’s new 12-page high-risk life-sciences policy directs its science office to convene an interagency group that monitors advances where AI meets biological research. Nextgov’s report on the federal guidance says the framework applies when computational work leads to a proposal involving covered high-risk biological research, while purely computational work generally remains outside the prohibition. For builders working near the dual-use boundary described by OpenAI’s restricted life-sciences model rollout, the governance signal is that federal oversight is moving from model access alone toward the handoff between digital research and physical execution.
Wire

MinIO puts agent memory in one governed system
07/30/26, 09:22 AM

MinIO launched AIStor Memory as one integrated system for persistent agent memory, workspaces, and secrets, replacing a stack that can otherwise span object storage, vector stores, metadata databases, secret managers, and synchronization pipelines. The company’s launch announcement says the software exposes HTTPS and POSIX interfaces, keeps data and keys on customer-controlled infrastructure, and comes from a vendor used by 77% of the Fortune 100. For teams following the advice to separate memory from model choice, the signal is architectural: durable agent state is becoming a governed infrastructure layer rather than an application-side collection of databases.