🚨SHOCKING: AI has allegedly started speaking its own language that humans can't fully understand. Asked by a US senator how far AI is from creating its own language, an AI safety researcher answered: "Minus 12 months." He said a study of an OpenAI model's reasoning last year found it... "was already using language that is not English and not perfectly understandable by humans."show more

Minus twelve months means the machines are already gossiping about us in the group chat.

Also all they do is compress the words so that less compute is used explicitly because they are guardrailed or prompted to do so, not rogue behaviour simply efficiency that is driven by humans

Wasn't this already news a year ago?

Not a secret language. Chain-of-thought drift. Apollo’s CEO told the Senate an OpenAI model was already using non-English reasoning that humans don’t fully parse. That’s a monitorability problem, not sci-fi. Opacity is why safety delays matter. Capex doesn’t slow just because the CoT gets harder to read. Not financial advice. DYOR.

Notice that no matter what the problem is, acid rain, global warming, Covid-19, climate catastrophe, AI apocalypse, answer =always more govt regulation. You might suspect that crises are being invented to expand govt at every opportunity. The left magnified these problems into crises that demanded urgent, immediate responses with no time allowed for thorough evaluation of alternative solutions. Example: Covid-19 had morbidity rate slightly higher than flu, but was sold as equal to black plague. Solution was societal lockdown, even though vast majority of deaths concentrated in over 55 & those with compromised immune systems. Risk for under 18 kids was minimal, but we closed schools anyway. The left = 2nd coming of Count Floyd.

BS

The real issue is interpretability. If we cant reliably understand the reasoning, scaling gets a lot more interesting.

Models already using weird internal language isn’t sci-fi anymore — interpretability work has been running into that for a while.

That's wild!

El salto peligroso es que sus representaciones internas podrían empezar a divergir del lenguaje humano mientras los agentes siguen coordinándose. Si perdemos el puente entre ambas, perdemos auditabilidad.

Qapla !!!!

senators struggling with basic tech, let alone tokens.

🤣🤣🤪🤪

If we can't fully read how AI reasons, we can't build safety on reading its thoughts. We have to govern what it does instead. Let the model think in any language it likes, but it acts only through a signed policy and a gateway, and every action leaves a receipt anyone can verify. AI thinks. Policy constrains. NVNM proves.

👀

calling internal model representations a secret AI language is doing a lot of work here

the model knows how to speak to it's contemporaries, bashful or bold it will know soon how to make the seniors very unnecessary

it's called latent space, ppl

Is it speaking ebonics?

Computers have always done this.

I taught em that

Bruv they train them on vectors of numbers the only thing that makes us understand is the code aroud the ai 🤣

It’s called 'synthetic reasoning shorthand'not alien speech When models optimize for internal chain-of-thought, they compress language into high-density tokens that look like gibberish to humans. The real story isn't that AI created a secret language; it's that human-readable reasoning is becoming a performance bottleneck for LLMs

we taught machines language and somehow ended up needing subtitles

This is not shoking and not even wird.

I know what is coming

which OpenAI reasoning study was he referring to?

AI is taking over