Ive gotten a lot of value lately using his writing to understand how parameter size, MOE and PLE variations on local models influence performance on various hardware. Its still a useful level of detail as someone who used to roll their own models for various applications pre-LLM but who now mostly uses LLMs rather than training them.
If you used to write pytorch, i think his writing is a great level of technicality to stay on top of things. And id look to them if i ever wanted to do a bit of learning/train my own.
Id love to hear if others have resources at a similar level of technicality and detial.
There are no open source AI models, at least not useful ones (yet [1]). Open weight is not the same as open source. "Open weight" models are still just inscrutable binary blobs that you can (theoretically) run on your own computer instead of through a SAAS web app. The open weight model labs don't even provide a high level catalog or any description whatsoever about what went into the training data.
This is not open source and we should stop conflating the two things.
Many of the Nemotron datasets are gated behind approval, and a license agreement.
The preamble on these datasets is: "This repository is publicly accessible, but you have to accept the conditions to access its files and contents."
I don't know what others have experienced, but I requested access to multiple Nemotron datasets and those requests were ignored for months before all but one request was rejected. There's no explanation for why, nor anything I can see which would lead to a rejection. So it's purely anecdotal and YMMV, but I don't see these as being particularly open.
For text based just assume "the entirety of anna's archive" and you'll be most of the way there. For multimodal pull up a torrent tracker and sort by size descending if you want to see some of the aggregated datasets.
Completely agree with this, I'm sick of people conflating the two. Open-weight models should be treated no more favourably than proprietary freeware.
Sure you can run tests and benchmarks on open-weight models, but that is the extent - there is no scrutiny, no auditing for bias or copyright contamination - just a black box that you rely on for "intelligence". I'm still shocked the way people can hand over not just huge swathes of data, but also decisions of all shapes and sizes - to AI companies with no way of being able to assess how the sausage is made.
I understand your quibble with terminology, but i think the “inscrutable binary blob” thing is a bit off base. You can create finetunes and post train models using only their open weights. You can’t create derivative works like that from an inscrutable binary blob
It's not just copyrighted training data. Truly open source e2e model training would include scrapers, data cleaning, all pretraining scripts, posttraining scripts, exact hardware info, etc. Open weights labs will release a sanitized version to make themselves look good / not give too much away.
Definitely, but make sure you're doing it with Astra. Fable will silently sabotage your work or just refuse to do it. They can help you design challenge generators for RLVR and maybe write training code but you'll be running into safeguards if you want them to make a data scraper.
Without the training data and open sourcing the training code and protocol you cant reproduce the model from scratch, thats true, but because the weights and structure are free to download you can modify, inspect and make derivative works off it. Calling it an inscrutable binary blob is misleading.
I do agree that its not reproducible and its often talked about like it is.
In some ways its like most scientific research with a "methodology section", which in todays academic climate often falls quite a bit short of a detailed step-by-step to reproduce results with all the ingredients, but in principle could be used by other experts to do this again. Even without publishing the exact data, a catalogue or description of what data and how it was cleaned/processed would go a long way towards a reproducible methodology i agree.
> you can modify, inspect and make derivative works
Even with a team of researchers inspecting the weights and understanding what is going on internally is a tedious operation. Having the training data doesn't really solve any issues here. But what it it does do is allow for auditability and public scrutiny - which imo is pretty important if you rely on these models.
The training-data side gets most of the attention here, but I feel like the license terms trip people up just as much in practice. A lot of "open" weights still come with usage restrictions (Llama's MAU cap, acceptable-use policies you have to accept before download) that don't feel very "open source" once you're actually trying to ship something on top of them.
I generally take the technocratic view of things, in other words you should understand the things you are trying to ban/regulate .
Replace AI with vaccines and the problems with how some parties are pushing for regulatory capture and/or blanket bans become obvious. Build policy around the effects e.g. job loss, biases etc. but go full laissiez faire and hands on on the technology. Nukes are 2nd amendment.
On the technical side, there's a lot of grunt work that's completely unrelated to machine learning but underpins modern AI. E.g. high performance computing and numerical techniques have zero relevance to day to day LLM research but makes or breaks the implementation. There are better reading list for those but you need to at least have a vague understanding of what numerical methods or a math kernel is before starting to lecture others on the economics of ML scaling and hardware. For stuff like data cleaning and scraping, it's a well known gray area field so you aren't gonna find too many for dummies guide on it. Can't have honest discussions on hacker news either because the people here get their panties in a twist over data scraping despite many of them doing it with zero hesitation or compunction if it comes up in a jira ticket. Proxy farms are the unsung hero of LLM engineering.
> I see a lot of HN users up in arms on the front page thread about LG TVs having VPN SDKs within them etc. And then five minutes later they will go back to their frontier LLMs to print their pay cheque. Fucking hypocrites of the highest order. It's oh noes how bad this tech is in my house but we will happily benefit from it.
Something something participate in society
(Or more elaborate: Disagreeing with the status quo does not make you a hypocrite for benefiting from the status quo.)
I don't know if people actually expect the training data to be published. For me its frustration, if its not open - don't call it open-source. It's so easy to continue to use the words the same way as we did before AI.
This has to be a joke or AI psychosis. By that logic, Conway's Game of Life is conscious because it can be used to build the logic gates to run AI models. [0]
32 comments
[ 5.0 ms ] story [ 44.8 ms ] threadIf you used to write pytorch, i think his writing is a great level of technicality to stay on top of things. And id look to them if i ever wanted to do a bit of learning/train my own.
Id love to hear if others have resources at a similar level of technicality and detial.
There are no open source AI models, at least not useful ones (yet [1]). Open weight is not the same as open source. "Open weight" models are still just inscrutable binary blobs that you can (theoretically) run on your own computer instead of through a SAAS web app. The open weight model labs don't even provide a high level catalog or any description whatsoever about what went into the training data.
This is not open source and we should stop conflating the two things.
[1] https://allenai.org/
The preamble on these datasets is: "This repository is publicly accessible, but you have to accept the conditions to access its files and contents."
I don't know what others have experienced, but I requested access to multiple Nemotron datasets and those requests were ignored for months before all but one request was rejected. There's no explanation for why, nor anything I can see which would lead to a rejection. So it's purely anecdotal and YMMV, but I don't see these as being particularly open.
because the training data full is of copyrighted works.
the answer to "what went into the training data" is "everything we could get our hands on".
Sure you can run tests and benchmarks on open-weight models, but that is the extent - there is no scrutiny, no auditing for bias or copyright contamination - just a black box that you rely on for "intelligence". I'm still shocked the way people can hand over not just huge swathes of data, but also decisions of all shapes and sizes - to AI companies with no way of being able to assess how the sausage is made.
I understand your quibble with terminology, but i think the “inscrutable binary blob” thing is a bit off base. You can create finetunes and post train models using only their open weights. You can’t create derivative works like that from an inscrutable binary blob
The models are freeware, and some have additional restrictions in the license that make them shareware, imho!
Without the training data and open sourcing the training code and protocol you cant reproduce the model from scratch, thats true, but because the weights and structure are free to download you can modify, inspect and make derivative works off it. Calling it an inscrutable binary blob is misleading.
I do agree that its not reproducible and its often talked about like it is.
In some ways its like most scientific research with a "methodology section", which in todays academic climate often falls quite a bit short of a detailed step-by-step to reproduce results with all the ingredients, but in principle could be used by other experts to do this again. Even without publishing the exact data, a catalogue or description of what data and how it was cleaned/processed would go a long way towards a reproducible methodology i agree.
Even with a team of researchers inspecting the weights and understanding what is going on internally is a tedious operation. Having the training data doesn't really solve any issues here. But what it it does do is allow for auditability and public scrutiny - which imo is pretty important if you rely on these models.
On the technical side, there's a lot of grunt work that's completely unrelated to machine learning but underpins modern AI. E.g. high performance computing and numerical techniques have zero relevance to day to day LLM research but makes or breaks the implementation. There are better reading list for those but you need to at least have a vague understanding of what numerical methods or a math kernel is before starting to lecture others on the economics of ML scaling and hardware. For stuff like data cleaning and scraping, it's a well known gray area field so you aren't gonna find too many for dummies guide on it. Can't have honest discussions on hacker news either because the people here get their panties in a twist over data scraping despite many of them doing it with zero hesitation or compunction if it comes up in a jira ticket. Proxy farms are the unsung hero of LLM engineering.
Without it, you unfortunately sound like a "just do your research" grifter to me.
Something something participate in society
(Or more elaborate: Disagreeing with the status quo does not make you a hypocrite for benefiting from the status quo.)
my fav read so far 2026
Stop anthropologizing the screw driver.
[0] https://nicholas.carlini.com/writing/2020/digital-logic-game...
my fav read so far 2026