# Relax about Neural Space > Source: > Published: 2026-09-06 10:32:57+00:00 Recently Chat GPT 6, Astra, was released by OpenAI. The model has dropped a large portion of its “chain of thought” reasoning in favor of “neural space” reasoning. This uses numeric / vector memory which can’t be readily interpreted by humans. Lots of media types and tech onlookers have sounded the alarm. This sounds like giving AI a “secret scratch pad” on which to plan the end of the world. *We just removed our one safeguard — we could at least see what the model was thinking internally to check for doomsday.* I know personally, it has been useful to see the model’s chain of thought at times. Noticing a complete mis-interpretation of an instruction, or rabit-hole-ing in a bad direction, I can course correct before I end up with thousands of lines of useless code. **But the truth is, this new approach is nothing new, and really not that scary.** First, using vector / numeric memory instead of chain of thought is not a new idea. In fact, RNNs, LSTMs, and other “old fashioned” approaches used this approach. The idea that the model could write down words and use them as memory is actually the newer approach. Every intermediate output of the model (the numbers generated between layers) is a kind of “numeric memory” of prior processing. This “neural space chain of thought” is just one more sequence of vectors, passed through the model again. This is the approach that stochastic depth and diffusion models have long used to get more from fewer params. I would argue that while sometimes useful for human inspection, chain of thought seems like an ugly kludge. Had a junior engineer proposed it, they probably would have gotten a tongue wagging from the more senior dev whose license plate reads *ILUVREALS.* The entire idea that the model collapsing a very rich feature space into words seems counterproductive. Tensor sequences have thousands of dimensions and represent the entire probability space of thought. Boiling each down to the most likely word in that space is just collapsing the entire probability space to a point on the surface. Re-reading your own chain of thought is just recording one sampled word from that distribution at each point in the sequence. This is akin to using a diffusion model, but rather than iterating over the noise vector repeatedly, generating an image, sampling to the most likely color at each pixel, then re-starting from that sampled image each iteration. Diffusion only works because the intermediate noise vector is preserved at every step, and only collapsed to pixel colors at the final step. So preserving the entire rich feature space makes sense, its far more efficient, flexible, and natural for an ML model. And its extremely banal in terms of implementation (it might actually be less code/complexity than chain of thought). But what about “information hiding”? The AI could be thinking a bunch of things that its keeping secret from us, like plotting a skynet-style takeover, but its just saying nice things on the surface. Every parameter in the AI model is being tuned to produce the next most likely word, and little else (thanks to gradient descent). There’s no room in training for “use some params to say X, while you think about doing unrelated Y”. Gradient descent moves ALL the params in the model towards producing the right output (including intermediate memory states), so any params that don’t contribute to the correct answer, are modified to do so. Ok, but still, its a “secret scratch pad” right? Well yes, but that has always existed. Right now, when AI produces “chain of thought”, behind the scenes there are thousand or millions of vectors that are generated which are not interpretable by humans, containing intermediate “thoughts” of the model. We only see the final written word which goes into the thought chain. Ok fine, but this memory is more persistent, those other values are transitory, right? True, but again, the model optimizer doesn’t leave much room for this kind of thing. Every bit of the memory is being used to answer the question at hand. And if you’ve done any ML work, you’ll know that these models DO NOT GENERALIZE BEYOND TRAINING DATA. Gradient descent does not magically produce models that think about stuff they haven’t seen, or produce useful vectors for anything other than the task at hand. In fact, its usually very difficult to get the models to even generalize on the data they have seen. That’s actually why we need petabytes of data to train these models. They cannot, like a human, read a few sentences (like this article) and suddenly gain a new understanding. They must see many permutations / combinations to form a statistical relationship between concepts and causal chains of “see input X, produce output Y”. There’s no amount of “hidden state vectors” that can paper over this fundamental behavior of the models. Lets apply “the model is secretly thinking X” to diffusion models. Diffusion models use hidden state vectors which pass through the model dozens of times. Once they have “thought about” the image to generate many times, they then use that information to produce the final pixel colors. How likely is it that even though they have been trained to do this, secretely, they are generating Nazi propaganda images in their thoughts, but only showing images of cats and scantily clad anime women as the final product? These models are tuned within an inch of their life. Every step and every param is optimized by gradient descent to produce the desired output. There is no room in the process for “secret images” to be lurking in the vector output space. In the “probability space of images” that the “thoughts” of this model represent, its a surface with a big peak near “big boobs anime waifu” and a rapid falloff with little in the way of “hostile world takeover” or “nazi propaganda”. *Ok fine, its not making secret plans, but this at least makes it harder to audit right?* No! This actually makes auditing more robust, assuming anyone builds the tools. These thought vectors CAN BE INTERPRETED! Wait what, I thought they couldn’t be? Remember how I said that collapsing to chain of thought was just sampling one word from a distribution surface of possible words represented by the vector space (and therefore a probably proposed by a hapless junior engineer)? Well we can always take the “neural memory” and draw word samples from the model. Just run the model forward and see what the probability space of output vocabulary looks like. This is actually even better for audits: Chain of thought: one word chosen at random from probability space of words the model is considering. Audited tensor space: the entire probability space of output words the model could generate at that moment. So an auditor (such as an inspector AI), could actually see if any problematic words are even possible. Proving the probability of the phrase “death to all humans” is less than 0.001% is actually far more reassuring than just saying “we didn’t find anything suspicious in the logs”. Anyway, hopefully this has been a big chill pill and people can relax a little!