Pretty torn on this. One the one hand he's not wrong. On the other, it probably means proton are going to leak everyone's email to the world and cease trading shortly afterwards.
Since I'm a customer and don't care _at all_ about any "feature work" that might be being done on email at this point, this isn't great. Leave the damn thing totally alone would seem much better.
This construction splits the space into 'small' models with a quadratic lookup cost and 'large' with sub-exponential cost. The small one is by enumerating all SAT problems with their answers, then looking up the answer. The large one is something in polynomials I don't understand. An answer from openai about where the cutoff is was:
the constructed lookup machine has vastly more states than the estimated total number of physically distinguishable configurations of the observable universe.
I'm not inclined to consider "wrote down all the answers ahead of time, so that's free" as a quadratic time solution to SAT. Nor very convinced by the other half.
So on balance, if this is considered proof that the exponential time hypothesis is false, and openai model reading of the lean was vaguely on point, then I reject the axioms of this system. Not a proof, even if there is lean. I'm OK with being downvoted for that position.
They do that to make it easier for readers to verify the statement (which is entirely contained in that file) separately from the proof (which is spread over thousands of files and github can't even show them all).
Weird paper. Models have had tokens for each byte for ages now. They can read and write individual bytes just fine, in addition to also having multibyte tokens.
The point is not to have byte tokens: it's to have only byte tokens, so that the usual failures of tokenisation (e.g. the number of Rs in strawberry) can be avoided.
Would byte tokens really solve how many r’s? It seems like it would require a level of introspective awareness of the tokens being processed that I don’t think transformers have. But I’m not expert enough to go beyond that intuition, so could easily be wrong.
The tokenizer would have to know to switch over to lexing as individual bytes rather than whatever else it might have been able to tokenize into but that seems like a question of labeling more than anything else.
A frustration with inference hosting companies is that 'GLM 5.3' is wildly underspecified. What quantization? What numerics? Do you match the official API's probability distribution? Deepseek flash is _all over the shop_ in terms of what people are selling under the name.
We will release our data soon. It's not just quantization, we are using NVFP4 from RadixLabs for GLM 5.3 but with our configuration and the OpenRouter harness for benchmark we are actually #4 in accuracy out of about 30+ providers. We run strictly on Blackwells (B200) on SGLang.
So its not just quantization, its the software, its version, its configuration, the exact hardware etc. They all matter.
Also, since DeepSeek 4.1 Flash is already quite quantized, the same OpenRouter benchmark harness puts us at #9, but the first 10 places (out of about 30) are within less than 1% of each other - so that is much harder to set.
But here, again, we run on B200, the official weights with specific configurations that we have on vllm.
reply