Thoughts
On tech, AI, business, and what actually works.
July 13, 2026
AI's binding constraint is no longer chips — it's electricity. A map of the companies selling the electrons: the dominant incumbents with decades of contracted hyperscaler revenue, and the wild cards that could pay off like lottery tickets over the next ten years.
July 5, 2026
OpenAI's GPT-5.6 didn't ship one model — it shipped three (Sol, Terra, Luna) priced in a clean 2x ladder. The real story isn't the 91.9% benchmark crown; it's that 'which model' is now a budget decision. A clear-eyed read of the tiers, the honest benchmark picture (Terra actually scores below GPT-5.5), and which tier a business should actually buy.
June 28, 2026
AI agents save real hours, yet only 41% of rollouts hit positive ROI within a year and 19% never pay back. The 2026 data on why most agent deployments stall — and the four moves that separate the winners. The dividing line isn't the model; it's measurement.
June 27, 2026
How Google turned itself from a search company into an agentic-AI company across two keynotes — I/O 2025 and I/O 2026 — and what its agents, open protocols, and quadrillion-token scale mean for business decision-makers planning their own AI bets.
June 26, 2026
In late June 2026 the most capable AI models stopped being something you could simply buy. OpenAI began releasing GPT-5.6 only to a government-vetted list of partners, days after Washington suspended foreign access to Anthropic's top models. It isn't a kill switch — it's a permission layer, and it changes who controls the frontier.
June 24, 2026
SpaceX's $60 billion all-stock purchase of Cursor isn't a coding-tools story — it's Elon Musk buying the surface where software gets written, turning millions of developers' keystrokes into training fuel for his AI, and opening a second front against GitHub with a platform called Origin. What the largest startup acquisition in history means for the tools your business runs on.
June 22, 2026
The protein-structure revolution that won AlphaFold a Nobel Prize has quietly split into three races — prediction, design, and medicine. A field guide to the top contenders (AlphaFold 3, Boltz, Chai, RFdiffusion, ESM3, Isomorphic Labs) and the one tension none of them has resolved: the benchmarks have never looked better, and not a single AI-designed drug has yet been approved.
June 20, 2026
In one week of June 2026, Anthropic quietly throttled its record-breaking Claude Fable 5 for AI researchers, then the U.S. Commerce Department forced Fable 5 and Mythos 5 offline worldwide via an export-control letter. The episode proved a frontier AI model is a strategic asset that can be pulled away overnight — by its maker or by the state.
June 20, 2026
Vercel deleted roughly 80% of its internal AI agent's tools and it got faster, cheaper, and more accurate. The counterintuitive result explains why most business AI fails — and hands small and medium businesses a concrete playbook for landing in the 5% that actually works.