Yeah exactly. I like to joke that I wrote this as a sublimated rant, but it's as much a meditation on why it was impossible to expect it to have been there.
I don't get to break out my academic voice that often these days, so I appreciate the indulgence.
Distributed post training over the internet is much easier than distributed pretraining. What we need is just agent traces, from everyone's daily use. Hugging Face already supports uploading agent traces, see https://huggingface.co/docs/hub/en/agent-traces
The raw CoT is encrypted for closed source models, but there are plenty of ways to decrypt them. Even without CoT content, the tool calls and the length of CoT are already good signals for post training.
In China we use 'transition stations' to access GPT and Claude. They collect GBs of plain text agent traces every day, and the data eventually belong to the open source community. My estimation is that there are already 10-100 TB of agent traces freely on the internet for everyone to download.
A posttraining data commons and the infrastructure surrounding it is the next great open source project IMO, akin to Postgres or Linux.
The Marin project from Stanford (and I believe now ran by Openathena, the authors of this blog post) is really promising and cool: https://marin.community/
The intent is good and I get it. But OpenAI started the same way and look at them now. This project is already sponsored by Google and they will want a return.
So thanks, no thanks, there should not be another public common knowledge of this size so that we can cling onto jobs we still have.
I understand the skepticism, and I don’t really expect to convince you otherwise.
But thinking over the outcomes we might want over the next decade and how we might get there, it seems to me the alternative to doing something here isn’t “AI goes away,” but rather “a handful of very well capitalized companies buy what they need and squeeze the labor market.”
I get it. And I trust that you have only the best intentions at the time.
The issue is that once sharks realize how profitable this project can be they will get into power by any means necessary and then abuse it. We've seen this pattern way too many times now - a grass roots initiative to help humanity that suddenly turns into maximizing profits. See also open source.
I saw how they started mounting cameras on workers in factories. How is this approach any different?
And AI might never truly go away, sure. That doesn't mean it'll be sustainable long term. Until questions of consent, ownership and disruption are solved, we should push back on reckless adoption.
So no, I don't see how this is in anyone's interest to participate, except for the likes of Google.
I think a post-training common would have to be somewhat protected, since these things would be created with great effort and would potentially impact people's ability to earn money in their jobs.
I could also imagining them happening as state projects for the sake of competitiveness.
One of the things I'm/we're thinking through is how to do this in a way that’s participatory rather than extractive. It's not necessary that what scientists would choose to automate with agents is the same as what a company trying to automate scientists would choose.
If we can help people automate drudgery while giving them a real say in what gets built and control over how their work is used, I think there’s a possibility of a much better outcome than the current trajectory. I’m not sure we’re in that world, but I think we should try.
I'm not sure I'll do it justice, but one of the general ideas behind open-model initiatives is that the economic impacts of specialized post-training are a fait accompli. OpenAI and Anthropic have spent $many-millions procuring closed posttraining data.
The relevant question here is whether you want open-weight models — projects like Open Athena — to remain or become competitive in the same range of specialties.
> I also wonder when the NSA starts sharing their wiretaps/etc with labs for 'national security'.
What would this benefit the NSA? If they share wiretap records with labs such as Anthropic, then suddenly anyone who can jailbreak a model gets access to this data. You don't need to train a model on data to understand or interpret it, as far as I know - providing others (esp. AI labs) with sensitive data is a breach of security, or I am not understanding.
I should’ve been more specific, I mean more industrial espionage and getting access to internal data that would be useful to frontier labs for training.
See the well-documented case of NSA conducting espionage against EU’s Airbus to help Boeing.
NVIDIA has done great work with the open datasets and they're much appreciated, but it's worth mentioning that a large fraction of their post-training data is not released. cf. https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-1... But at leased they disclosed what wasn't released.
Wrong direction. NSA says “give us the weights to Fable and loan us some engineers who can get a TS-SCI clearance so we can train our own model in house, if you don’t well just take them, in fact maybe we already did, it’s for national security”.
No, NSA programs are not SCI, haven't been for a few years. They are SAP.
NSA doesn't have to post train Fable. Maven works well enough and that same infra. can be used for Fable. I imagine that's what's going to happen in 2 years once the current admin. moves out.
What gets me excited about this project is its legibility for philanthropy, like how the Internet Archive is the destination for random recordings found in Grandma's attic.
Having a trusted steward to handle the anonymization and eval design process could kickstart a world where retiring tax accountants, or even active and generous ones, are willing to donate their data to the posttraining commons at a scale that's actually much larger than the "data industry".
post training is both the future and the secret sauce.
The next iteration of AI assisted devs will not have all the devs use the same harness and model like most do now.
Instead, devs will build their own post trained models that are experts of their own domain, that have their secret sauce..
I imagine any dev would want most of those models to be something they can self host because of multiple reasons - one of them being that their employers or customers might not agree to use a cloud hosted or managed model. Plus, costs. And control.
Post training is in the future because we haven't yet reached a shared agreement on what that first "good enough" model will be like - and also because the current AI scape is in a state of flux where post training will have a lot of churn.
A very academic way of saying We have all the data in the world, except the data we actually need. You ask a company how something works and they say:
"Oh, theres a process."
Great. Where is it documented?
"in a spreadsheet."
Which spreadsheet?
"Well, theres the old spreadsheet, new spreadsheet, and the spreadsheet that Karen maintains."
Where's Karen?
"She left in 2021."
Yeah exactly. I like to joke that I wrote this as a sublimated rant, but it's as much a meditation on why it was impossible to expect it to have been there.
I don't get to break out my academic voice that often these days, so I appreciate the indulgence.
Distributed post training over the internet is much easier than distributed pretraining. What we need is just agent traces, from everyone's daily use. Hugging Face already supports uploading agent traces, see https://huggingface.co/docs/hub/en/agent-traces
We've been developing DataClaw to help everyone collect agent traces, see https://github.com/peteromallet/dataclaw . There are other similar efforts, such as in https://old.reddit.com/r/LocalLLaMA/comments/1u795pb/donate_...
The raw CoT is encrypted for closed source models, but there are plenty of ways to decrypt them. Even without CoT content, the tool calls and the length of CoT are already good signals for post training.
In China we use 'transition stations' to access GPT and Claude. They collect GBs of plain text agent traces every day, and the data eventually belong to the open source community. My estimation is that there are already 10-100 TB of agent traces freely on the internet for everyone to download.
Hi, I wrote this a few weeks ago. Didn't expect to see this here, etc., etc.
Happy to answer any questions. We have a discord https://marin.community/
A posttraining data commons and the infrastructure surrounding it is the next great open source project IMO, akin to Postgres or Linux.
The Marin project from Stanford (and I believe now ran by Openathena, the authors of this blog post) is really promising and cool: https://marin.community/
I was impressed by Marin's published evals policies: https://github.com/marin-community/marin/issues/9409
Precise and very clearly written.
The intent is good and I get it. But OpenAI started the same way and look at them now. This project is already sponsored by Google and they will want a return.
So thanks, no thanks, there should not be another public common knowledge of this size so that we can cling onto jobs we still have.
I understand the skepticism, and I don’t really expect to convince you otherwise.
But thinking over the outcomes we might want over the next decade and how we might get there, it seems to me the alternative to doing something here isn’t “AI goes away,” but rather “a handful of very well capitalized companies buy what they need and squeeze the labor market.”
I get it. And I trust that you have only the best intentions at the time.
The issue is that once sharks realize how profitable this project can be they will get into power by any means necessary and then abuse it. We've seen this pattern way too many times now - a grass roots initiative to help humanity that suddenly turns into maximizing profits. See also open source.
I saw how they started mounting cameras on workers in factories. How is this approach any different?
And AI might never truly go away, sure. That doesn't mean it'll be sustainable long term. Until questions of consent, ownership and disruption are solved, we should push back on reckless adoption.
So no, I don't see how this is in anyone's interest to participate, except for the likes of Google.
I think a post-training common would have to be somewhat protected, since these things would be created with great effort and would potentially impact people's ability to earn money in their jobs.
I could also imagining them happening as state projects for the sake of competitiveness.
One of the things I'm/we're thinking through is how to do this in a way that’s participatory rather than extractive. It's not necessary that what scientists would choose to automate with agents is the same as what a company trying to automate scientists would choose.
If we can help people automate drudgery while giving them a real say in what gets built and control over how their work is used, I think there’s a possibility of a much better outcome than the current trajectory. I’m not sure we’re in that world, but I think we should try.
The road to hell is paved with good intentions.
If any form of participation will give benefits after they've contributed, it will be abused into extraction.
I honestly think you shouldn't try.
I'm not sure I'll do it justice, but one of the general ideas behind open-model initiatives is that the economic impacts of specialized post-training are a fait accompli. OpenAI and Anthropic have spent $many-millions procuring closed posttraining data.
The relevant question here is whether you want open-weight models — projects like Open Athena — to remain or become competitive in the same range of specialties.
They're already kinda protected, open weight releases rarely come with datasets, especially post-training datasets. NVIDIA is probably the exception.
I also wonder when the NSA starts sharing their wiretaps/etc with labs for 'national security'. Probably already happening, to be honest.
> I also wonder when the NSA starts sharing their wiretaps/etc with labs for 'national security'.
What would this benefit the NSA? If they share wiretap records with labs such as Anthropic, then suddenly anyone who can jailbreak a model gets access to this data. You don't need to train a model on data to understand or interpret it, as far as I know - providing others (esp. AI labs) with sensitive data is a breach of security, or I am not understanding.
I should’ve been more specific, I mean more industrial espionage and getting access to internal data that would be useful to frontier labs for training.
See the well-documented case of NSA conducting espionage against EU’s Airbus to help Boeing.
NVIDIA has done great work with the open datasets and they're much appreciated, but it's worth mentioning that a large fraction of their post-training data is not released. cf. https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-1... But at leased they disclosed what wasn't released.
Wrong direction. NSA says “give us the weights to Fable and loan us some engineers who can get a TS-SCI clearance so we can train our own model in house, if you don’t well just take them, in fact maybe we already did, it’s for national security”.
No, NSA programs are not SCI, haven't been for a few years. They are SAP.
NSA doesn't have to post train Fable. Maven works well enough and that same infra. can be used for Fable. I imagine that's what's going to happen in 2 years once the current admin. moves out.
What gets me excited about this project is its legibility for philanthropy, like how the Internet Archive is the destination for random recordings found in Grandma's attic.
Having a trusted steward to handle the anonymization and eval design process could kickstart a world where retiring tax accountants, or even active and generous ones, are willing to donate their data to the posttraining commons at a scale that's actually much larger than the "data industry".
post training is both the future and the secret sauce.
The next iteration of AI assisted devs will not have all the devs use the same harness and model like most do now.
Instead, devs will build their own post trained models that are experts of their own domain, that have their secret sauce..
I imagine any dev would want most of those models to be something they can self host because of multiple reasons - one of them being that their employers or customers might not agree to use a cloud hosted or managed model. Plus, costs. And control.
Post training is in the future because we haven't yet reached a shared agreement on what that first "good enough" model will be like - and also because the current AI scape is in a state of flux where post training will have a lot of churn.