Wait, this isn't real, is it? Is there actually an intermediate model that translates DeepSeek's thinking from its "alien language" into human languages? That's not actually the case, right?

I thought "thinking" is literally the model generating additional text in a human language that shows its "thought process". It's added to the model's context, which helps it reason better because it now has this self-generated context.

The "their own language" idea seems to come from some recent science fiction where LLMs develop their alien language and take over the world by 2037 or something.

You're right, it's just additional text that allows it to do thinking / reasoning-like behavior. The big proprietary models hide the real output from the user and instead provide a friendly abridged version, but that's just to protect their secret sauce from distillation.

Yeah, it's actually the case. Researchers have shown that the models response doesn't always follow from the reasoning. Whether you consider that an internal language or not really depends on what you're speculating the neural network is doing. I think there was an Antropic paper on it.

The parent is off, you’re right. They may reason in any language, typically whatever the user’s language is, and you’ll see the reasoning directly with an open model like Deepseek.

Research only showed that thinking might be disconnected from the final output but in my experience they are very strongly correlated in recent models

> Research only showed that thinking might be disconnected from the final output

It is trivial to regularly spot obvious contradictions and inconsistencies if you read carefully. For example I've encountered traces that amounted to "I can deduce X, therefore Y, so that means Z" but then the model turns around and outputs "the answer is W because X". It's even been demonstrated that having the model output placeholder tokens or other gibberish instead of "thoughts" still improves performance. However the thinking traces can still be useful to the end user regardless.

I see those too and I think of it as the "thinking" in action. If you could replace their actual thinking trace with gibberish and get improved performance that scaled with the amount of gibberish you injected, that's what we'd do. But instead, we see that the quality of of the model's output scales with the amount of 'thinking' tokens they generate before responding.

It has been my experience that yes, models make contradictions throughout their thinking process, but the conclusions they arrive at during/near the end of thinking more often than not align with the final output.

Current models simply generate additional text that gets added to the context for the trace. However iterative models that "think" by repeatedly looping through several layers instead of outputting text have recently been demonstrated.

I may have misremembered but I thought I had read somewhere that recent models by OpenAI and Anthropic tend to produce reasoning that is not always understandable for humans. But you're right that it's not the case for Deepseek so maybe I'm hallucinating ;)

Or maybe it was an article or a tweet about researchers trying really hard to steer the model to think in English otherwise interpretability / safety becomes a lot harder?

As far as I'm aware, it's not true for models like DeepSeek or other Chinese open-weight models (at least those that I have seen); their reasoning traces are fully composed from some human language, be it English, Chinese or another one; by the way, most of them can adapt their reasoning based on user language, for example, if user speaks English the reasoning more likely will be in English.

I think that for DeepSeek problem (thinking and replying in Chinese) everything is kinda simpler: in their official chat, they're probably using some kind of system prompt which is (probably) written in Chinese, so that's why model may prefer Chinese in it's output.

I have seen mixed language thinking from claude when i speak to it in english but we are discussing a product thats in spanish or searching amazon spain.

Summaries by different smaller models are usually made by closed proprietary models like Claude as a way to combat the distillation of real reasoning traces by competitors. Open weight models show the real reasoning traces. Reasoning traces operate in the same space as the non-reasoning output. It's all just one large text for an LLM. Internally, reasoning is just ordinary chat completion between <think></think> tags.

> summaries generated

Or hallucinated

This is inaccurate. The displayed reasoning traces are summaries, but the model thinks in nominally regular human languages. AI labs are very light on details (as they consider them as their "edge"), but both GPT5.5 and Claude Mythos/Fable system cards discuss chain-of-thought monitorability quite a bit.

They occasionally show snippets of CoT in papers they write, e.g. for o3/o4/GPT5 models [1] or Claude 3.5 Haiku [2].

[1]: https://openai.com/index/evaluating-chain-of-thought-monitor... [2]: https://transformer-circuits.pub/2025/attribution-graphs/bio...