Can someone explain what are teachers in Multi-Teacher On-Policy Distillation? I can imagine math, coding and other verifiable domains, but they also have biology? Is it where distillation from bigger models come in?
They're other models yes.
Kimi-K3 Technical Report [pdf]
91–100 of 199 posts
Re: Kimi-K3 Technical Report [pdf]
#92Re: Kimi-K3 Technical Report [pdf]
#93Earlier quoted context omitted.
> Kudos to Moonshot for truly being what OpenAI should have been. Kudos to Moonshot for making these weights available for download. Lets not fool ourselves and claim these are "open source" by any understanding of the concept though, there are usage restrictions (even if you download them) and also training data isn't clearly broken down either, nor it it actually using a FOSS license.
> Kudos to Moonshot for making these weights available for download. Lets not fool ourselves and claim these are "open source" by any understanding of the concept though, there are usage restrictions (even if you download them) and also training data isn't clearly broken down either, nor it it actually using a FOSS license. They will probably never release the training data because that represents a large part of the…
>They will probably never release the training data because that represents a large part of their competitive moat
Not that what you've written isn't the case. However, in addition to what you've written, (or probably even before what you've written), there's the fact that everything they're training on is stolen IP. Same with US LLM labs.
Let's not kid ourselve's about where the training data is coming from. They are not asking artists, writers, coders, content creators, etc etc etc for permission to use their creations.
Anthropic, Moonshot et al are doing incredible things, but we shouldn't gloss over the costs. Both present and future costs are kind of enormous.
Re: Kimi-K3 Technical Report [pdf]
#94Earlier quoted context omitted.
> Kudos to Moonshot for truly being what OpenAI should have been. Kudos to Moonshot for making these weights available for download. Lets not fool ourselves and claim these are "open source" by any understanding of the concept though, there are usage restrictions (even if you download them) and also training data isn't clearly broken down either, nor it it actually using a FOSS license.
> Kudos to Moonshot for making these weights available for download. Lets not fool ourselves and claim these are "open source" by any understanding of the concept though, there are usage restrictions (even if you download them) and also training data isn't clearly broken down either, nor it it actually using a FOSS license. They will probably never release the training data because that represents a large part of the…
> They use private datasets that cost a lot to acquire, synthetic datasets and a lot copyrighted material for which they don't have licensing.
Yeaaah, and this, of course, is worth it, because it leads to you being able to download a near frontier model. Don't get me wrong, long-term humanity is probably better of with science with little regards to pesky things like ethics and provenance, but we also have a tendency to not fully realize the downstream or wider effects until way too late.
And sure, there is a lot of reasons to go with keeping your software proprietary too, I'm not trying to claim otherwise, same with training data. It's just that usually we don't call proprietary software "open source" unless it is open source, regardless of the reasons someone keep it proprietary or not, could be for whatever reason really.
Re: Kimi-K3 Technical Report [pdf]
#95Earlier quoted context omitted.
And hire 2 or 3 dev ops to keep it running ? That another 400 to 700k. It becomes your problem and not someone else’s. However, I don’t trust hosted LLMs for anything that needs to be private.
> And hire 2 or 3 dev ops to keep it running Not a devops but I'd say one full time is already too many.
Yes but zero is not enough and where do you get a fraction of a competent dev op from?
Re: Kimi-K3 Technical Report [pdf]
#96License: https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE > If the Licensee or any of its affiliates operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars (or the equivalent in other currencies) in total over any consecutive 12 months, the Licensee must enter into a separate agreement with Moonshot AI before using the Software or…
What is it a license for, though? Copyright? Are model weights even copyrightable?
Re: Kimi-K3 Technical Report [pdf]
#97Also open sourced a bunch of infra to go with it. Anyone who claims open source and open weights models are "decel" needs to get their head checked https://github.com/MoonshotAI/MoonEP https://github.com/kvcache-ai/AgentEnv https://github.com/MoonshotAI/FlashKDA
To me its clear that it is decel the only reason other labs can catch up is because the frontier labs can be distilled, and they siphon a % of the labs' revenue to reinvest into the next iteration full accel would mean nationalizing the big 2 labs and locking in manhattan project style until RSI (Edit: some great counterpoints in the replies. my view has definitely been changed!)
being able to replicate it in the open means that there's nothing special about frontier models.
Frontier models would have to do something extraordinary or unique, or unreplicatable, because clearly there is no moat, and US companies are sitting on huge nvidia valuations and get surprised when competitors beat them.
Re: Kimi-K3 Technical Report [pdf]
#98Earlier quoted context omitted.
Latest "big" release from any of the bigger American lab must have been GPT-OSS-120b I think? Released ~summer 2025, so pretty much one years ago. Doesn't seem like it'll happen by itself, so something either forcing their hand figuratively, or something forcing their hand literally. Personally I was wishing/hoping for one of the recent Gemma releases to be in the ~100B class at least, but sadly Google is keeping tha…
the longer i read this comment the wronger it gets
Don't make me add more of my thoughts to the comment, I can still edit it.
Deeply fun "care about improving people's lives" quote on your user page :p
Re: Kimi-K3 Technical Report [pdf]
#99Can someone explain what are teachers in Multi-Teacher On-Policy Distillation? I can imagine math, coding and other verifiable domains, but they also have biology? Is it where distillation from bigger models come in?
They reference https://thinkingmachines.ai/blog/on-policy-distillation/
If I understand correctly, it's distillation via having a teacher model score each of the student's tokens for a problem based on their own probabilities of generating that token at each step in the sequence. The reward/loss is then applied as RL.
The multi-teacher bit seems to imply they're distilling from multiple models. It's light on the details, but it seems like it could be part of distilling from frontier/closed models. Provided they calculate the logprobs, which OpenAI seems to allow via API but not Anthropic. Maybe they have a way of estimating the logprobs externally?
This method can be used to learn any domain from the teacher. Biology included.
Re: Kimi-K3 Technical Report [pdf]
#100Also open sourced a bunch of infra to go with it. Anyone who claims open source and open weights models are "decel" needs to get their head checked https://github.com/MoonshotAI/MoonEP https://github.com/kvcache-ai/AgentEnv https://github.com/MoonshotAI/FlashKDA
To me its clear that it is decel the only reason other labs can catch up is because the frontier labs can be distilled, and they siphon a % of the labs' revenue to reinvest into the next iteration full accel would mean nationalizing the big 2 labs and locking in manhattan project style until RSI (Edit: some great counterpoints in the replies. my view has definitely been changed!)
Distilling can be done by fast follower closed models too, so this argument against open models doesn't hold up.