Today's Hallucination HQThe Pentagon Called Anthropic a Risk. The White House Didn't Get the Memo.
The US government contains multitudes. Trump administration officials are reportedly nudging banks to pilot Anthropic's new Mythos model — a perfectly normal suggestion, except the Department of Defense recently labelled Anthropic a supply-chain risk. So one arm of the same government is saying "try this" while another arm is saying "be careful with that." Washington: where the left hand doesn't just not know what the right hand is doing, it actively contradicts it.
Source: TechCrunch
Everyone Wants to Write Your Code. No One Agrees Who's Winning.
The race to become the AI that developers actually use has turned properly fierce. OpenAI, Google, and Anthropic are all throwing considerable resource at coding tools, as "vibe coding" — writing software by describing what you want in plain English, then hoping for the best — goes from curiosity to genuine workflow. The product differences are real but narrow, which means the winner will likely be whoever gets the distribution right. That, historically, is Google's game to lose. They are trying very hard to lose it.
Source: The Verge
GPUs in Orbit. Because Apparently the Earth Wasn't Big Enough.
Kepler Communications has put 40 GPUs into low Earth orbit, making it the largest compute cluster currently floating above your head, which was presumably not a sentence anyone expected to read in 2026. Its first customer, Sophia Space, will use the cluster to process satellite data directly in orbit rather than beaming it all down to Earth first. It's faster, cheaper on bandwidth, and only mildly terrifying as a concept. The cloud, it turns out, was not sufficiently cloud-like.
Source: TechCrunch
Slow and Steady, If You Own the Hardware
Apple has spent two years being gently mocked for arriving late to the AI party, but the argument for Apple winning anyway is gaining traction. The case: Apple owns the device, the chip, the operating system, and crucially, the user's trust with their personal data. Every other AI company is a guest in that ecosystem. Apple is the landlord. Landlords, as a rule, do not need to be technically brilliant. They just need to own the building — and Apple very much owns the building.
Source: Hacker News
Teaching AI to Remember Why Its Last Drug Idea Was Terrible
Researchers have proposed a new memory system for AI drug-discovery agents called Constraint-Aware Corrective Memory, which does roughly what it sounds like: it helps AI models remember past failures, constraints, and dead ends when searching for viable drug candidates. The problem isn't that these agents can't find interesting molecules — it's that the final set of candidates must simultaneously satisfy multiple biological and chemical requirements, which current models handle with all the grace of someone packing a suitcase at the last minute. This is the suitcase organiser.
Source: ArXiv AI
As ever, the machines are learning. Whether we are is a separate question.
|