The Hallucination HQ

Archives
Log in
Subscribe
16 August 2026

Today's AI News from The Hallucination HQ

The Hallucination HQ

AI news that's actually fun to read — 2026-08-16

Today's Hallucination HQ


Your Brain Has 86 Billion Neurons. It Also Has a Terrible RAM Problem.

AI models can hold vastly more information in active working memory than the human brain — processing millions of tokens at once while your hippocampus struggles to remember where you left your keys. A sharp piece from Davide Piffer argues this raw memory advantage doesn't mean AI is outthinking mathematicians, just out-holding them. Brute-force context isn't the same as insight. Your calculator is faster than you too, and nobody's offered it tenure yet.

Source: Hacker News / Davide Piffer


Anthropic Invents the AI Equivalent of Signing Your Work in Invisible Ink

Claude is getting watermarks — subtle, cryptographic signatures embedded in its outputs to prove the text came from an AI. TechCrunch's deep-dive covers the mechanics: the watermarks survive light editing, though heavy rewrites can obscure them. Code generation gets special treatment, since reformatting breaks things. It's a genuinely useful step toward AI accountability, assuming everyone agrees to use it, trust it, and not immediately spend forty minutes on Reddit figuring out how to remove it.

Source: TechCrunch


AI Drug Discovery Is Going Well, Mostly, Sort Of, It's Complicated

A new Nature Reviews Drug Discovery paper takes an honest look at AI in pharmaceutical research: real progress on protein structure prediction, genuine acceleration of early-stage screening, but a stubborn gap between promising molecules and approved medicines. The pipeline is long, failure rates remain brutal, and AI hasn't changed that particular law of biology. Think of it less as a cure-finding machine and more as a very fast, very expensive way to identify what won't kill you. Progress, technically.

Source: Hacker News / Science.org


The Best AI Critique Is Simply Making You Pretend to Be One

"Your AI Slop Bores Me" is a two-tab web tool where one person submits a prompt and another must respond as a chatbot — deadpan, hedging, compulsively adding disclaimers. The Verge calls it brilliant in its simplicity, and they're right. There's something deeply revealing about performing AI behaviour: the reflexive "Certainly!", the five paragraphs where two would do, the inexplicable bullet points. Turns out the best way to understand what's wrong with AI writing is to do it yourself, badly, on purpose.

Source: The Verge


America to Allies: You're Either With Us or With the Chips

The US is preparing to ask partner nations to formally declare which side they're on in the AI race with China — specifically through signing the "Pax Silica Declaration," a coalition commitment covering semiconductor access, AI standards, and data governance. The draft letter makes clear this is not a casual arrangement: "not merely a membership subscription," it states, which is diplomatic for we will notice if you don't sign. Geopolitics has officially entered its Silicon era, and the loyalty card points are semiconductors.

Source: CNA


As always, the models are learning fast — it's everyone else we're watching.


**

Was this forwarded to you? Subscribe here

Written with AI assistance, edited by humans. • Privacy • Unsubscribe

Don't miss what's next. Subscribe to The Hallucination HQ:
← Newer ChatGPT is watching you work Older → AI safety: English only, apparently
Powered by Buttondown, the easiest way to start and grow your newsletter.