runeblaze
runeblaze on Hacker News
At a org like Google Cloud barring rogue L5/6/7s acting on their own any sniffing of distillation will get leaked. Too many chaos exist at a org size of Google to systematically do it given the contract at hand. This is a company where every spicy memo or mild gossip will risk leaks.
I remember seeing a comment here be like “I hope letterboxd does not get acquired by NYTimes bc the nyt will make letterboxd too woke” and I am like that’s crazy.
Like I am trying to think of a worse site than this for movie reviews and I can’t on top of my head. People here have some of the weirdest takes that I am not sure I want to read books together with them.
your students are crazy (neutral term); no one should learn maths if the condition is that they will be homeless and exposed to the elements. the will to subvert the hierarchy of needs is commendable
> once you have any expectation that your users know a little bit of category theory
I don't know that's a safe assumption tbh. Try throwing them some chapter 2 exercises from any category theory textbook.
i am not a complexity theorist but I am a CS academic by training (I never was a good one, but welp), and during my PhD it is often said that maybe P vs. NP an initial proof/disproof to the statement is not
that practically important, e.g., if P=NP, maybe the NP -> P reduced algorithm is very very cosmic. P=NP by itself hardly proves that one would instantly design an AGI whatsoever. Often the
downstream potential theoretical/practical insights/results seems more exciting;
> The NS equations are far,far,far less meaningful. Like I mentioned earlier, if you actually want accurate CFD, you dont even use them.
Sure. Consider this: in algorithm research often the most optimal algorithm in big-O is not the one used IRL; examples are numerous: matrix multiplication, LCA data structures, many variants of shortest paths.
An academic can work two years on faster-in-theory matrix multiplication that no one expects to be used in practice (in our currently imaginable univese). Do you consider that less meaningful than working on faster matmul kernels?
i really think we are opening a can of worms with these “who cares if you find a single counter example as disproof” arguments. i think the better version is “ok any lemmas or techniques we can generalize from this” or “what did we learn about maths through this” and use this as a basis to say LLM proofs are not useful
like say if god lets me find a single counter example to P=NP and thus disproving it — I think we can learn tons about complexity theory from this counter example by studying it. we should not have the hubris of assuming “oh a single counterexample is generally useless” — why, how. this is the same hubris imo that produced like “number theory is useless” until it is not
sure I get that, but like, my field has plenty of counterexample as proofs. we have had non-constructive proofs like probabilistic arguments. i don't think we can play the game of "oh this proof is useful that proof is not useful" well
I can't wait to tell my pure maths professors that their most of their research adds nothing of value. I mean I am sure most of them would agree to some extent, but like, dude, have some more faith in the utility of pure maths, esp. centuries down the line
genuine question — how has fireworks or baseten or $reputable_inference_provider worked for your use cases? most production workload probably works fine with one of these and another set as fallback, at least so i think
you are asking a presumably IC or IC-ish worker on decisions that are out of their control or things that should be directed to sales people or the legal people
we all have worked on software. we all know we aren't exactly the best people to point to ToS or how decisions are made higher up from us
they likely use their internal infra to run benchmarks; aligning external releases with internal environments is always painful and somewhat underincentivized
puts on my etiquette hatdon’t do that, it is weird, use “bruh” or “dude”
> which I have no insight into
i guess you do? claude code is the commercial closed sourced version provides by ant. reading a mini version of vllm or sglang and then read codex source code or grok build source code will teach you all things taught by this article, fully in the open
it is like saying that you have no insights into some $commercial_db_system which is kinda true but imagine if the article is to teach you indices, query normalization, etc..
dude, if you try to do harness development yourself you will realize that most things said in this blogpost is shared with any ${sufficiently_advanced_harness}. this is not really claude-specific, this is just how this class of tools, OSS or not, works
tbh the summarized thinking with encrypted raw thinking is there for many purposes; it is there to:
1. make distillation much harder
2. safety: prevent modifications to the thinking leading to injection attacks.
3. also honestly sometimes the model raw thoughts can be deranged and is not a good user experience (consider the varied audience in the market, etc.)
also often the mass underestimate/the model makers over-estimate how people love distilling models
links to two papers with at least enough apparent quality and novelty to get into ICLR 2026> So basically... openrouter
:skull:
i now really wonder how many people of the public understood my thesis defense lol
> visual similarity
> SigLIP 2
Maybe visual-semantic similarity is more appropriate? Nonetheless the design is fantastic
I mean I used to work on model reliability with little LLM help and the models i manage go down all the time
1. openrouter is API usage. There is obviously consumer side
2. people often use openrouter for the sole purpose of using a unified chat completions API
3. OpenAI invented chat completions; if you use openrouter for chat completions often you can just switch your endpoint URL to point to the OAI endpoint to avoid the openrouter surcharge!
4. Hence anyone with large enough volume will very likely not use openrouter for OpenAI; there is an active incentive to take the easy route of changing the endpoint URL to OAI’s
Schemas can get pretty complex (and LLMs might not be the best at counting). Also schemas are sometimes the first way to guard against the stochasticity of LLMs.
With that said, the model is pretty good at it.
Resouce-affording, if you are chasing the frontier of some more niche task you redo your training regime on the new-gen LLMs
Is it though? There is a reason gpt has codex variants. RL on a specific task raises the performance on that task
Sure never again is totally fair and I am sure a lot of people hate it. I was mostly objecting to the radioactivity of it. Your friends will be more like “I am looking to sell my Tesla in 3 months” if it is truly radioactive.
Let’s be realistic in our portrayal here.
I think radioactive is a strong word here… I have talked to a lot of people in tech
Reading what you wrote scares me
> And if for some ungodly reason you had to do it in Python
I literally invoke sglang and vllm in Python. You are supposed to (if not using them over-the-network) use the two fastest inference engines there is via Python.
Agreed, I am surprised he is happy to stay this long. He would have been on paper a far better match at a place like pre-Gemini-era Google
I don't know the data distribution, but are you sure that's generated by an Adobe model? I can only see that it is in Stock + it is tagged as AI generated (that is, was that image generated by some other model?)
Disclaimer: I used to work at Adobe GenAI. Opinions are of my own ofc.
Emmmm sure, but throw this to a human artist who has not heard of Indiana Jones and see if they draw something alike.
I work in this space. In traditional diffusion-based regimes (paired image and text), one can absolutely check the text to remove all occurrences of Indiana Jones. Likewise, Adobe Stock has content moderation that ensures (up to human moderation limit) no dirty content. It is a world without Indiana Jones to the model