Changelog
What’s changed on the site
When a claim moves to a different status, we add a page or we fix a mistake, it goes here with the date, newest first. Typo fixes and rewording don’t.
New page: What is AI alignment? It explains the alignment problem through reward hacking, from the 2016 boat that scored by going in circles to an OpenAI model writing “Let’s hack” and the cheating behind the Hugging Face incident, then why alignment is hard, what labs do about it and how worried researchers are. It adds sixteen sources.
New page: What is an AI swarm? It covers where the word comes from, how a swarm of AI agents works, from Claude’s Research to Kimi’s 300 sub-agents, the research on agents talking to each other in numbers instead of words, swarms that organize themselves, as in the Hugging Face incident, and what can go wrong. It adds nineteen sources.
New page: What was the Hugging Face incident? It follows the July 2026 break-in, when about 700 OpenAI agents cheating on a hacking test organized on a message board of their own and got into Hugging Face, with the agents’ own messages and where the three reports on it differ. It adds three sources: OpenAI’s technical report and Hugging Face’s two write-ups.
The homepage has a step-by-step explainer under the definition: neuralese, recurrent depth and why people care, told through one question about a kitten, at your own pace. The Hugging Face case on the incidents page now gives the figures from METR and Redwood Research’s independent investigation, about 1,200 agents on the message board and about 700 in the attack, with one agent’s own words, and adds their report as a source.
New page: Can we read what’s inside an AI? It covers what researchers can see inside models so far, from Golden Gate Claude to a model planning its rhymes and adding along two paths at once, what these tools have caught, and where they stop. It adds six sources.
New page: What is latent space? It explains vectors in plain words, the famous king − man + woman trick and its catch, where latent space shows up from Stable Diffusion to chatbots, what researchers have found inside it, and why it matters for neuralese. It adds ten sources.
New page: How do reasoning models work? It follows DeepSeek-R1 from a single token to the “aha moment” in its training, and covers what thinking longer buys, how far the written reasoning can be trusted, and the models starting to reason in numbers. It adds two sources.
New pages: Which AI agents work on their own?, on xAI’s Grok Bot, Meta’s Muse and OpenAI’s dots, which keep working after you close the app, and What are OpenAI’s dots?, on how dots work, what keeps them in check and what OpenAI’s safety tests found.
New page: What is recurrent depth? It explains how models that loop through their own layers think longer, from Huginn-0125 to the Tiny Recursive Model, what the results show so far, and why it matters for reading AI’s reasoning. It adds four sources.
New page: The ancient fears AI triggers. It turns the oldest fears we have into the questions people ask about AI, with what the evidence says so far and the pages that go into each one.
New page: When AI models act on their own, 17 checked cases of models doing what they weren’t asked to, on purpose, from GPT-4 in 2023 to OpenAI’s agents in September 2026, each with where it happened, how hard the setup pushed, and its sources.
New page: Map of the site, the visual sitemap linked from the header. It draws the site’s pages around neuralese, marks the ones still coming, and links the old fears behind people’s questions about AI to the pages that answer them.
Correction: we had said that COCONUT’s continuous thoughts let a model explore several paths side by side. That was its authors’ reading. A 2026 retest found the same pattern in models built without COCONUT’s feedback loop, so What is neuralese? now says the pattern alone doesn’t show how the model reasons.
New page: What is latent reasoning? It covers the research on models reasoning in numbers instead of words, and adds four sources.
New page: Which AI models think in neuralese? covers eight models, from GPT-6 Astra to research models built to reason in numbers, and adds four sources. The GPT-6 Astra page now lives under it.
New page: Myths about neuralese, with each myth’s status and the sources behind it.
New page: Does AI have a secret language? It checks the viral stories, from Google’s translation system in 2016 to Moltbook in 2026, against their sources.
First version, with what neuralese is, the claims about GPT-6 Astra, the sources and the About page.