newsroom
AI Safety & Alignment · Prompt injection makes current agents unsafe for email/inbox access; safety race lags capability race
now playing · AI Safety & Alignment
AI compute capex reaching 1% global GDP, 2%+ US GDP — massive buildout underway
Global AI compute spend is just under 1% of world GDP and over 2% of US GDP, with data center construction becoming a major industrial activity. This infrastructure wave is still in early s…
AI Agentstailwindscore 9/10nat friedman
AI agents will become autonomous economic actors requiring new payment rails and stablecoins
Agents will soon transact independently at scale, leapfrogging legacy payment infrastructure. They cannot get traditional bank accounts (no SSN), so stablecoins become the native currency.…
Token budgeting replaces headcount budgeting as ICs rack up API charges like portfolio managers
Individual contributors can now spend thousands on API calls via agents, forcing companies to treat token allocation like hedge fund capital allocation — evaluating ROI per IC strategy, usi…
Coding is the first well-covered RL domain; other white-collar domains need equivalent environments
Models excel at coding because verifiable RL environments exist (unit tests, compilation). Finance, legal, compliance lack such environments — building them is hard but tractable. The next…
Agent commerce requires rebuilding payment stack from scratch — identity, disputes, pricing, stablecoins
AI agents cannot use traditional banking (no SSN, no credit history). They will transact in stablecoins on new rails. The winner will be the platform that achieves 'SEO in the model' — beco…
AI's inflationary vs disinflationary impact uncertain — China/WTO analogy suggests disinflationary abundance
Connecting low-cost superintelligence to global economy mirrors China's WTO entry: massive purchasing power gains for consumers (disinflationary), but distributional effects unknown. Baulmo…
All hardware becomes trivial IO peripherals for AI — golden age of tinkering arrives
AI agents can reverse-engineer drivers, read academic papers, and control arbitrary hardware (Raspberry Pis, scanners, Tesla) via code. Hardware loses independent software stacks; everythin…
Prompt injection makes current agents unsafe for email/inbox access; safety race lags capability race
Frontier models remain trivially prompt-injectable. Agents with inbox access can be hijacked via crafted emails. Users run agents with 'dangerously skip permissions' flags. Market demands c…
Tech giants may not shrink headcount — AI efficiency could enable more output per person instead
Big tech headcount trajectory uncertain. Companies are inefficient; AI could let same people do 10x more rather than cutting staff. Organizational transition needed: smaller autonomous pods…
Data center aesthetics matter for social license — beauty is a civilizational obligation, not just NIMBY mitigation
With >2% US GDP flowing into data center construction, the industry must earn the right to build by making facilities pleasing to the human soul, not just economically efficient. Incrementa…
SaaS dynamics shifting as software production cost collapses; startup formation remains evergreen but categories change
Traditional SaaS relied on high software production costs which are now collapsing. Best companies to start in 2026 look more applied/industrial (energy, turbines) vs. 2015. Procurement fro…
Smartphone memory supply fully diverted to AI data centers, creating structural shortage
Smartphone shipments declining because all memory production is being absorbed by data center demand. This signals a multi-year structural tailwind for DRAM/NAND suppliers as AI infrastruct…
At-home diagnostics + LLMs + cheap hardware enable consumer-driven medical mastery
LLMs can analyze poor-quality telemetry (iPhone images) and control low-end diagnostic hardware bought on eBay, creating a massive new category of consumer-driven health mastery that dwarfs…