AI digest: money, agents, and a grade inflation scandal
OpenAI's eye-watering finances, AWS patching agent gaps, and proof that students are absolutely using AI to cheat.
Big week for AI infrastructure, big week for uncomfortable truths about how people are actually using these tools.
OpenAI tripled revenue but burned £3.7 billion doing it
In Q1 2026, OpenAI pulled in $5.7 billion and spent $3.7 billion getting there. Stock-based compensation alone was $2.3 billion, which is a remarkable number. With $73 billion in reserves they’re not going bust, but the gap between revenue and costs is still vast. Full figures here.
AWS says AI agents are broken and wants to fix them
At its New York summit, AWS launched two new services to address the two most obvious problems with enterprise AI agents: they don’t know enough about your business, and they’re not secure enough. Continuum handles code vulnerability detection and patching; Context builds a knowledge graph from corporate data so agents actually have something useful to work with. This feels like the most honest thing anyone in the cloud space has said about the current state of agents. Read more.
AI is inflating student grades and it’s not because anyone is learning more
A UC Berkeley study of over 500,000 grades found that writing and coding courses saw sharp grade increases after ChatGPT launched, concentrated in homework rather than exams. That pattern points pretty clearly to outsourcing, not improvement. Norway has already moved to ban generative AI in primary schools entirely from August, which is a blunt response but arguably the right instinct. Berkeley study coverage here. Norway ban here.
OpenAI Codex learns by watching you work once
The new Record and Replay feature in the macOS Codex app lets you demonstrate a workflow once and Codex turns it into a reusable skill it can repeat on its own. It is not available in the UK yet, which is frustrating. When it does arrive, this kind of task automation at the application layer is going to be genuinely useful for repetitive dev work. More details here.