The editorial argues the manifesto treats 'open source AI' as a switch you can flip today, but no major frontier-class model meets OSI's October 2024 OSAID v1.0 bar requiring training data detail, training code, weights, and OSI-approved terms. Llama 3.1, DeepSeek V3, Qwen 3, and Mistral Large all fail — weights are downloadable but the training corpus, training code, and RLHF pipeline are not.
The one-page manifesto argues that if you can't run, audit, or fork the model behind your product, you're a tenant rather than an owner. It frames the choice as urgent and binary — switch to open source AI or accept permanent dependence on proprietary API providers.
Commenters point out that Llama 3.1's license explicitly restricts use above 700M monthly active users and prohibits using outputs to train competing models — neither clause is OSI-compatible. Calling these models 'open source' muddies the term and lets vendors claim openness credit without meeting the bar.
Some commenters argue DeepSeek V3 publishing weights plus a technical report still enables fine-tuning, local inference, and auditing of behavior — which delivers most of the practical benefits the manifesto cares about. Demanding full OSAID compliance sets a bar so high that no useful model qualifies, which risks making 'open source AI' a purity test rather than a movement.
A bare one-page manifesto at opensourceaimustwin.com climbed from 477 HN points to 1,262 in under a day. No product, no team, no founder bio — just an argument: if you can't run, audit, or fork the model behind your product, you're a tenant. The site has already been covered here for that framing. What's worth a second look is the comment section, which spent the back half of the day pulling at a thread the manifesto never touches.
The thread: 'open source AI' is not a category that meaningfully exists yet. The manifesto treats it as a choice you can make tomorrow — switch from a proprietary API to 'an open source model' the way you'd switch from MySQL to Postgres. But every model the audience is actually thinking of when they read it — Llama 3.1, DeepSeek V3, Qwen 3, Mistral Large — fails any rigorous definition of open source. The weights are downloadable. The license is permissive-ish. The training data, training code, and RLHF pipeline are not.
The Open Source Initiative settled this argument 14 months ago. In October 2024, OSI shipped v1.0 of the Open Source AI Definition (OSAID). It requires four things to call a model 'open source': (1) sufficient detail about training data so a skilled person could recreate a substantially equivalent system, (2) complete source code used to train and run the system, (3) parameters including weights, and (4) all four under OSI-approved terms. By that bar, the count of major frontier-class models that qualify is approximately zero.
The gap between 'open weights' and 'open source' is the entire game, and the manifesto skips it.
Llama 3.1's license explicitly restricts use above 700M monthly active users and prohibits using outputs to train competing models — neither clause is OSI-compatible. DeepSeek V3 publishes weights and a technical report, but neither the training corpus nor the full recipe; you can fine-tune from the released checkpoint, but you can't reproduce it from scratch, which means you can't audit what's in it. Qwen 3 is similar. Mistral's open releases (Mixtral, Nemo) come with research-license carve-outs. AI2's OLMo and OLMo 2 are the closest thing to actually-open — full data, full code, full weights — and they're trained at scales where nobody's tempted to host them as a frontier substitute in production.
This isn't pedantry. The manifesto's three operational claims — *run*, *audit*, *fork* — collapse without reproducibility. You can run Llama 3 locally; congratulations, you have the equivalent of a free binary. You cannot audit it for training-data contamination, copyrighted-code memorization, or benchmark-set leakage, because you don't have the training set. You cannot meaningfully fork the lineage: any fine-tune you publish inherits Meta's licensing posture, not your own. The 'tenant vs. owner' framing the manifesto trades on was about runtime control. Real ownership of an AI system extends backward through training, and almost no consumer of 'open' models has it.
The HN comments started catching this around the 600-point mark. The top sub-thread reframes the manifesto's argument: the fight worth fighting isn't *open* — it's *reproducible*. A reproducible model is one where, given the training corpus, the training code, and roughly equivalent compute, an independent party could rebuild it. That's a far smaller coalition than the manifesto implies. Reproducibility advocates are the OLMo team at AI2, the BigScience consortium that built BLOOM, EleutherAI, and a handful of European labs working under EU AI Act traceability requirements. They are not the audience signing opensourceaimustwin.com.
Meta, DeepSeek, Alibaba, and Mistral all have business reasons to publish weights — and none of them are 'we want the community to fork us.' Meta needs commoditized model layers to protect its ad business from OpenAI-as-platform. DeepSeek and Qwen function partly as state-tied prestige projects and partly as accelerants for domestic AI ecosystems insulated from US export controls. Mistral's open releases are a European-sovereignty play with paid API economics underneath. Calling any of them 'open source' inflates a marketing term into a property right.
If you're a senior engineer being asked to 'switch to open source AI' in the next planning cycle, the honest answer is: you can switch to open-weight AI, and that's a real and useful thing, but it solves a narrower problem than the manifesto implies.
The lock-in problem the manifesto identifies is real, but the actual fix is multi-vendor abstraction, not weight provenance. Wire your inference calls through LiteLLM, OpenRouter, a custom gateway, or anything with the same shape. The cost-to-switch from Anthropic to OpenAI to a self-hosted Llama 3.1 70B endpoint becomes a config flag instead of a refactor. You inherit none of the auditability the manifesto promises, but you get the runtime fungibility, which is most of what teams actually need. If your compliance posture requires reproducibility — regulated industries, model-card-as-evidence-in-court use cases, EU AI Act high-risk classifications — none of the 'open' frontier models qualify and you should plan accordingly: smaller OSAID-conformant models like OLMo, or in-house pretrains. There is no third option yet.
The cost of the manifesto's category confusion is that procurement and security teams will treat 'we run Llama' as equivalent to 'we use open source databases.' It is not. Tell them now, before someone has to walk that back during a vendor risk review.
The manifesto that should follow opensourceaimustwin.com is 'Reproducible AI Must Win,' and it would be signed by a different, smaller, more technical coalition. It would ask for training corpora published under terms compatible with downstream redistribution, full training and RL code, and benchmark sets disclosed prior to scoring. It would also be deeply unfun, because the patrons of 'open weights' mostly cannot or will not meet that bar. Until then, calling a downloaded checkpoint 'open source AI' is the same category error as calling a leaked database dump 'open data.' The manifesto is right that the tenant problem is real. It's wrong that the off-ramp exists yet.
This, and distributed LLM inference. We are at a point where no single person can setup a rig to run a SOTA model, it is just too expensive.So we must build and adopt frameworks that allow individuals to share resources to run SOTA models in a distributed manner. That way they will also be non-censo
It won in my house/my business right from the start. (Well, open weights, at least — which is an uncomfortable nuance.)I have never understood the willingness to make the functioning of or development of a product so completely dependent on the secret sauce of one of two big unprofitable, inscr
When "open source" means freeware, it's like saying "we want free copies".What we should be saying is: We want a public, community-ran project that does pretraining and training collectively. This means working on a training corpus in public and somehow coordinating the trai
I would be totally willing to pay $50 per month to support an open source AI lab, rather to get open source models as byproducts of corporations.
Top 10 dev stories every morning at 8am UTC. AI-curated. Retro terminal HTML email.
I've been contemplating a decentralized model training system for some time using volunteer machines that we all contribute. But, it is astronomically difficult. The communication speeds are untenable.And, there is the issue of data poisoning from untrusted nodes. I've almost cracked that