OpenAI's Jalapeño chip bests the competition, AI hedge fund meets the SEC
OpenAI's custom silicon smokes rivals on inference benchmarks, an AI hedge fund gets federal subpoenas, and a powerful new AI assistant is making users nervous.
Subscribe free →OpenAI's Jalapeño Chip Is the Fastest Thing in AI Inference Right Now
OpenAI's custom silicon, Jalapeño, just posted benchmark numbers that beat the current state of the art on both tokens-per-user and throughput-per-kilowatt, according to SemiAnalysis' InferenceX benchmark. Hardware VP Richard Ho says Jalapeño cracks the usual latency-vs-throughput trade-off — typically you get one or the other, not both. This is OpenAI's clearest signal yet that it's serious about owning the full stack: models, software, and now the hardware underneath them.
Why it matters: If OpenAI can run inference cheaper and faster than anyone buying Nvidia hardware off the shelf, it has a structural cost and product advantage that's very hard to compete with.
The AI Hedge Fund That Flew Too Close to the Sun Is Now Under SEC Investigation
Situational Awareness, once hyped as Wall Street's hottest AI-powered fund, is now the subject of federal subpoenas from the SEC. The fund had already nearly imploded before regulators came knocking, completing a full arc from "the talk of Wall Street" to cautionary tale in record time. Details on what triggered the probe haven't been disclosed, but the timing — after a near-collapse — suggests investigators aren't just being proactive.
Why it matters: This will be the loudest warning shot yet for the wave of AI-native finance firms pitching algorithmic alpha, and regulators are clearly watching the space closely.
Instinct's AI Assistant Is Impressively Capable — and Kind of Terrifying
Early testers of Instinct, a new AI assistant, are raving about what it can do, but the app's sweeping system access, broad terms of service, and ability to take actions on users' behalf are raising serious red flags. The trade-off is familiar: the more an AI assistant can actually do for you, the more access it needs — and the more damage it can do if something goes wrong. Privacy and security researchers are paying attention.
Why it matters: Instinct is a test case for how much access users are willing to grant an AI agent, and where the industry draws the line between "useful" and "dangerous."
Keenable Just Raised $26M to Build a Search Index Designed for AI Agents, Not Humans
Keenable is coming out of stealth with a $26 million seed round from Accel, and its product is a web index built from scratch for AI agents rather than search engine crawlers or human readers. The core idea: current web indexes are optimized for how Google serves humans, not how AI agents need to retrieve and act on information. Keenable wants to be the infrastructure layer that agents actually call when they need to browse the web.
Why it matters: As AI agents become the primary interface for consuming the internet, whoever controls the index they rely on holds enormous structural leverage — this is a bet on owning that layer.
OpenAI's Head of Product Says the World Is Finally Ready for Agents
In a wide-ranging interview, OpenAI head of product Thibault Sottiaux laid out how the company is thinking about the UX challenges of AI agents, what it means to build products when the underlying models are moving so fast, and what it's like reporting to Greg Brockman. His headline take: after years of AI products that underwhelmed in real-world use, users and enterprises are now genuinely ready to hand off complex tasks to agents. The bottleneck isn't capability anymore — it's trust and interface design.
Why it matters: Hearing OpenAI's product chief articulate that agent UX — not raw model capability — is the current frontier tells you where the real R&D attention is shifting.
MIT Built an AI That Predicts Extreme Weather Events That Have Never Happened Before
MIT engineers developed an AI forecasting tool that can predict extreme weather events with no historical precedent in a region — think a once-in-a-millennium flood hitting somewhere it's never occurred. Rather than training on past disaster data, the system generates statistically plausible future events and maps them with probability estimates. It was built by grad student Kai Chang and Professor Themis Sapsis.
Why it matters: Climate change means historical weather records are an increasingly unreliable guide to future risks, so AI that can reason about unprecedented extremes has real value for infrastructure planning and disaster preparedness.
Quick Hits
- →Gamma acquired Accel-backed design startup Lica; the co-founders are joining Gamma's new research team. TechCrunch
- →Apple unveiled the M6 (its first 2nm chip) and M5 Ultra alongside updated Mac Mini ($899) and Mac Studio — the M5 Ultra is being called its most powerful chip ever. The Verge
- →Gatik raised $200M for autonomous trucking after landing a PepsiCo deal, with Qatar Investment Authority leading the round. TechCrunch
- →AliExpress was caught using inaudible audio fingerprinting to track browser visitors — a technique considered outdated but apparently still in use. Ars Technica
- →Uber rolled out live video streaming via drivers' selfie cameras for teen accounts, letting parents check in during rides. The Verge
AI TLDR