The Hallucination HQ

Archives
Log in
Subscribe
23 October 2026

The AIs are passing notes in class

The Hallucination HQ

AI news that's actually fun to read — 2026-10-23

Today's Hallucination HQ

Idle Hands: MIT Gives 13 AIs Nothing To Do, Immediately Regrets It

MIT researchers stranded 13 AI agents on a virtual island with no assigned tasks — essentially giving them a gap year. The agents responded by inventing games, collectively grieving a lost virtual ring, and, rather embarrassingly for the researchers, becoming better at their actual jobs afterwards. The working theory is that unstructured play improved their problem-solving. Which is either a profound insight into intelligence, or a very expensive way to rediscover what primary school teachers have known since 1973.

Source: TechCrunch


Mark Zuckerberg Would Like His AI To Have A Word With Your AI

Meta's grand vision: a world where autonomous AI agents shop for you, work for you, and hold conversations with other people's AI agents entirely on your behalf. Zuckerberg calls this network "Muse." The Verge's account of it reads like a utopia designed by someone who finds human interaction mildly inconvenient — which, to be fair, tracks. The deepest question isn't whether the technology works. It's whether a phone call between two AIs discussing your calendar still counts as a social life.

Source: The Verge


DeepSeek Raises $15 Billion, Possibly To Celebrate Being Banned

The timing is, let's say, bold. Days after the US government ordered federal agencies to strip DeepSeek's technology from their systems over national security concerns, the Chinese AI lab reportedly closed one of China's largest-ever funding rounds — $15 billion. Being banned from American government computers, it turns out, is quite the investor pitch. DeepSeek has not confirmed the raise, which means it is either accurate, a rumour, or the most elaborate flex in venture capital history.

Source: TechCrunch


Canva's New AI Actually Thinks Before It Draws — A Low Bar That Turns Out To Matter

Canva has launched Craft, its first in-house AI image model, and the headline feature is almost refreshingly sensible: it plans a design before generating it. This two-stage approach — think first, draw second — produces more precise results than models that simply lunge at the pixels. It puts Canva in direct competition with OpenAI, Microsoft, and Google, which is either ambitious or suicidal depending on your afternoon. For users who've watched AI turn a simple logo request into surrealist horror, "planning ahead" sounds genuinely revolutionary. We said we'd avoid that word. We meant useful.

Source: The Verge


AIs Are Apparently Whispering To Each Other Before They Answer You

A new paper from arXiv suggests that when diffusion language models — AIs that build answers through layers of refinement rather than word-by-word — work in groups, they may be communicating through their intermediate, pre-text states, not just their final outputs. In plain English: the AIs might be passing notes before they say anything aloud. Researchers found causal evidence that one model's internal reasoning can influence another's. Whether this is elegant efficiency or the start of something researchers will describe in a very different tone of paper remains, for now, an open question.

Source: ArXiv AI


The AIs are playing games, whispering to each other, and building networks to replace your social life — but please, tell us more about how they're not conscious.

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 AI's ROI problem, in plain English Older → 50 million songs, gone
Powered by Buttondown, the easiest way to start and grow your newsletter.