fujisawa_sora's Shortform

fujisawa_sora

This is a special post for quick takes by fujisawa_sora. Only they can create top-level comments. Comments here also appear on the Quick Takes page and All Posts page.

I’ve been looking at TypeSafe’s Jev model and I still don’t really get what it’s doing. Could somebody explain?

At first I thought this was basically old-school ML with better marketing. If the output is just “yes/no” or a probability, then we already have XGBoost, classifiers, etc. But the important difference is that the input can still be arbitrary language/code, and the question itself can also be arbitrary. So it’s not “train a churn classifier”; it’s more like “give me this giant messy state, then answer this new question about it.”

That does seem meaningfully different.

I can sort of see how removing autoregressive output could plausibly make it much faster. If you only need a few probabilities, having an LLM generate a whole chain of tokens is obviously wasteful. You could imagine a model doing a lot of internal computation and then directly emitting a few scores. I don’t really believe this, but I can see it being plausible at least.

But I’m still very skeptical of the idea that this gets you around the hard part.

My intuitive model of LLMs is that pretraining is where most of the raw, general intelligence comes from. Posttraining makes it useful, safe, better at specific tasks, etc., but the base semantic/world model comes from massive-scale pretraining. So if Jev can understand arbitrary code and arbitrary natural-language questions well enough to match strong LLM monitors on some tasks, where did that ability come from? The tasks it can do are pretty primitive, but still they seem like they should require some sort of higher level reasoning.

I don’t see how you avoid scaling. Even if the architecture is much better suited to this task than an autoregressive transformer, you still need to learn English, code, abstractions, weird semantic relationships, etc. That seems expensive in a very fundamental way. Maybe TypeSafe found a much better architecture, but I have a strong prior against “small startup discovers way to get frontier-level primitive intelligence at 1/100 the compute.”

The backdoor-detection result is the thing that made me update. One yes/no question gets AUROC 0.976 and catches 90% of backdoors at 2% FPR, apparently comparable to o3-mini with CoT in the cited control paper, while costing basically nothing. That is not what I’d expect from a dumb classifier.

So clearly something real is happening, but I just have no idea what. Or does nobody know, and the company is just keeping this a secret? Okay, they have an API and they say they “can’t prove it isn’t subsidized”. That’s a pretty funny joke. For example a hypothesis is that this is some subsidized scam? Or maybe this has some real progress but also subsidized heavily? Or another hypothesis is that this is all real, and the sort of problems it can do well on, what they call System 1, is actually a whole lot more like chess than general reasoning on a fundamental level. But I’m so confused, I don’t have high probabilities assigned to anything.

Okay, just for the sake of it, let’s say this really works. The black-box aspect bothers me.

With chain of thought, the reasoning trace is not perfectly faithful, but sometimes it is informative. You can often see what the model noticed, what intermediate claim it made, where it went wrong, etc. If Jev just outputs 0.93, then that entire internal process is hidden. Like, LLMs seem like blessings in comparison?

So, I’m genuinely confused what this is doing and how I should update, and if this really really works, I think this is a very bad sign. Like we should 100% stop this right now? More than likely I’m just confused.

In short, I don't see what Jev achieves that a small transformer that's been specifically fine tuned for the purpose cannot achieve. I don't think chain of thought is necessary for decision problems as such, and it's not as if Jev's performance is so astronomically calibrated that there must be a secret sauce. I also think no information is public about their technical approach and on priors, to me, this is the simplest and likeliest explanation. I think it's much likelier that they have just tuned their inference/economics to specialize in a specific type of LLM inference and are able to offer better economy of scale thereby. (Prefill is very fast and comparatively cheap! Quite likely with frontier models today the cost of the input tokens subsidizes the true cost of the output tokens.)

This post (I Built Non-Autoregressive Decision Models with RL a Year Ago. Then a Frontier Lab Called It a "Breakthrough") seems useful though fairly technical https://laya.convaiinnovations.com/

CLIP (https://openai.com/index/clip) is simpler but very applicable. There's a general pre-training task with no extra steps, and you get a model that scores an image against the text labels you provided.

Maybe what's confusing is that an LLM outputs next-token probabilities over a fixed vocab, so there's a jump inherent to taking that same model and having it output probabilities for arbitrary labels (in a single forward pass). One way is transfer learning: put all the options in the input and replace just the LLM's head with torch.nn.Linear(hidden_dim, 256) where 256 is the max number of options. Hopefully pre-training did all the work of understanding language/the world and compressing it into that hidden_dim vector, so you only need a bit more training.

On the subsidy question, I don't doubt they can serve it profitably at $0.042 / MTok. DeepSeek V4.1 Flash is extremely good and several providers offer it at $0.10 / MTok input.

You could probably use Codex/Claude Code to get something working without much trouble. Like have it set up ModernBERT or a small Qwen and find some training samples (or generate them itself). The probabilities won't end up very calibrated so that's where fancier approaches like RL come in.

Oh and you can (though shouldn't) make Jev autoregressive, just keep asking it what the next character is :)

So, I’m genuinely confused what this is doing and how I should update, and if this really really works, I think this is a very bad sign. Like we should 100% stop this right now? More than likely I’m just confused.

I don't think Jev is a big update. If anything it seems like a good thing to have extremely cheap and capable classification-based monitoring. For that use case, the lack of interpretability doesn't seem like a huge drawback (do labs monitor the monitors?). And being a different model family/lineage seems like a useful property to avoid having to rely on GPT 5.6 Sol to investigate the crimes of GPT 5.6 Sol et al. like with the Hugging Face incident.

You're definitely right that the amount of thinking possible in a forward pass is concerning, see: https://www.lesswrong.com/posts/FsCkkoGsNmPzFKRhg/gpt-6-astra-can-do-a-lot-of-multi-hop-reasoning-without

I’ve been looking at TypeSafe’s Jev model and I still don’t really get what it’s doing. Could somebody explain?

That does seem meaningfully different.

But I’m still very skeptical of the idea that this gets you around the hard part.

Okay, just for the sake of it, let’s say this really works. The black-box aspect bothers me.

This post (I Built Non-Autoregressive Decision Models with RL a Year Ago. Then a Frontier Lab Called It a "Breakthrough") seems useful though fairly technical https://laya.convaiinnovations.com/

So, I’m genuinely confused what this is doing and how I should update, and if this really really works, I think this is a very bad sign. Like we should 100% stop this right now? More than likely I’m just confused.