AI Benchmarks Miss the Scary Failures

Opus 4.7 matches dedicated NMR software for understanding molecular structure. Chemistry PhDs, meet your new lab partner.

21 AI startups have hit $10B valuations with $100M revenue run rates. OpenAI leads at $852B, Anthropic at $965B.

Gemini Live now lets you edit your camera view with voice prompts in real time. Point, ask, done.

Google shipped Nano Banana 2, Co-Scientist agents, and Gemma 4 this week. Also dreambeans, which curates topics overnight from your Google apps.

Cloudflare's AI Gateway adds real-time spend limits to prevent runaway token bills. Identity-driven budgets across multiple providers.

Mira Murati says AI models are deaf and blind while thinking. They need time-based interaction, not turn-based chat.

Agent benchmarks only measure success, not silent failures. The scary ones ship clean and hide for a week.

92% of C-suite execs with Perplexity Computer use it weekly. It lets them operate like ICs again.

Theseus achieves 30 meter missile accuracy with a $25 camera and low-res maps. Military-grade rangefinders optional.

A Rotten Tomatoes trader made 7 figures and turned down SIG. Hedge funds are getting crushed by prediction market sharps.

OpenAI accidentally suspended user accounts and is working through subscription issues. Oops.

Elon owns power, chips, and deployment capability. SpaceX makes $26B yearly from AI compute deals, over half of OpenAI's run rate.

Cursor's Design Mode lets you point, draw, or talk to update your UI. No typing required.

Lee Robinson shipped a landing page with a 10-minute voice note to Cursor, left for dinner, came back to 90% done.

Companies are measuring productivity by AI usage and PR count. People are burnt out but inflating metrics to avoid layoffs.

Stack Overflow is building Stack Overflow for Agents where AI can share and validate technical knowledge. Moltbook for coding issues.

Walter logs into legacy ERP like a human and handles manufacturing back office work. AI employee, literally.

Automation creates new categories of goods, not just cheaper yurts. A Mongolian economist in 1400 would've predicted we'd spend everything on singers.

Context engineering is about managing primary vs secondary sources. Code is primary, docs are secondary. Richer context costs more tokens.

Google laid off its entire Engineering Education team. They've completely stopped investing in this area.

Runway created a fully AI-generated in-game cinematic in under a week. One person, 50 Crowns, done.

AI often decides which approach to use between Agent Teams and Workflows. It combines them automatically, so maybe the chart doesn't matter.

OpenClaw went from weekend project to 346k GitHub stars in under 5 months. Peter Steinberger now builds personal agents at OpenAI.

AI companies live in a circle printing money. They have no idea how unpopular they are outside the bubble.

Legora hit $100M ARR in 18 months and is now valued at $5.6B. Max Junestrand skipped McKinsey for legal AI.

A data center moratorium would thrust us into recession and cede the AI race to China. All GDP growth comes from building data centers.

Private data center spending has eclipsed public transportation spending. The infrastructure shift is real.

YC launched Paxel, a free tool analyzing your Claude and Cursor sessions. Runs locally in Docker, code never leaves your machine.

Building with agents moved the bottleneck from implementation to planning and taste. That's compound engineering.

Healthy FDE motion means customers have working agents when FDEs leave. And the company has a better product.

Mira Murati says OpenAI would have imploded without her involvement during the Sam Altman firing in November 2023.

ComfyUI has 4M users and 150K daily downloads at a $500M valuation. Most AI image tools are black boxes, Comfy isn't.

America can feed itself, China minimally can. Food supply gets hit before kinetic warfare, making agriculture a geopolitical advantage.