like the point on control. In actual production, the question quickly becomes more granular than open vs. closed: which parts of the workflow deserve which intelligence, at what cost and latency?
We ran into this at scale and ended up with a hybrid architecture for reasons that were as much economic as technical.
I wrote about the operator side of that argument in "The AI Model Is Not the Product."
Interesting read. One nuance I’d add: open weights are not the same thing as open AI. If the training data, filtering, reward models and post-training pipeline remain closed, we can inspect the artifact, but not fully audit how it came to exist. Is that really “open” AI (vs closed AI)?
>recent collaboration between Thinking Machines and investment firm Bridgewater. For a hedge fund, the alpha is the business, so we would expect Bridgewater to be particularly cautious in selecting an AI architecture. In this case, the two entities collaborated to train a custom model as an extension of Thinking Machines’ base one.
I believe they used Qwen3-235B as the base (rather than a Thinking Machines model) and post-trained it. Interestingly, since the post-training happened on Tinker, Bridgewater must have been Ok with their training data "leaving the building" and for the final post-trained weights (i.e., the final alpha) to be visible to Thinking Machines (Or, perhaps, Thinking Machines set up a private Tinker for Bridgewater).
I would assume that if regulatory compliance can be made a complementary asset it also means that countries have bought in to the convex returns that the frontier lab provides. Now, whether that remains to be the case in competition to an open source model is not yet decided.
Diffusion to build these models already did rely on a fair use argument to take from human written books, articles, art etc. Regulatory compliance couldn’t come in because of the fact that all of this tacit knowledge was not protected.
Now, open source models, as you make the valid point of they should be deployed differently: could be great for cyber security, not necessarily for gene research or physical security, make a lot of sense from a company’s investment standpoint.
Why not pay up big now and prevent disintermediation 5-10 years down the road?
Netflix depended heavily on AWS in the 2000s but did well to build its own cloud infrastructure before Prime Video became a thing. Imagine it had to make the move now.
I do think we are going to see heavy, heavy investment in the next few years for firms buying engineering capability on top of open source models to protect not only their knowledge but their business.
Not that OpenAI or Anthropic will see less growth but the argument of buy or build will be much more competitive than partner.
Reads like a handful of ( wealthy) companies, who acquired the base knowledge essentially for free, repurposed and mined it, and don’t want to share common knowledge without getting paid. Sounds like a glorified consultant!
This piece evokes a familiar line of reasoning: relaxed gun access benefits the firearms sector, civilians gain self-protection, widespread gun ownership creates community defense, hazards of gun violence receive little attention, and broad gun regulation ought to be removed. Unlike tangible weapons, AI weights may be copied endlessly and cannot be revoked remotely, creating safety challenges that are far more difficult to address.
like the point on control. In actual production, the question quickly becomes more granular than open vs. closed: which parts of the workflow deserve which intelligence, at what cost and latency?
We ran into this at scale and ended up with a hybrid architecture for reasons that were as much economic as technical.
I wrote about the operator side of that argument in "The AI Model Is Not the Product."
https://rmeerasahib.substack.com/p/the-ai-model-is-not-the-product
Interesting read. One nuance I’d add: open weights are not the same thing as open AI. If the training data, filtering, reward models and post-training pipeline remain closed, we can inspect the artifact, but not fully audit how it came to exist. Is that really “open” AI (vs closed AI)?
Here is my take on it:
https://bhaveshsenedhun.substack.com/p/open-weights-are-not-open-source
Yeah, we need a better shared vocabulary for the open weights vs fully open models (weights, training datasets, evaluation harness, etc).
Piece is about open weights. Fully open source is much harder to deliver for a number of reasons (copyright, etc), but also very important!
Here's so software. It's great. We tested it.
There's no official entity to validate or set standards (FAA, FDA etc).
You can't see how we tested it. Or what data we used.
But we spent a lot of money. So it's great.
Thought-provoking read, Christian!
>recent collaboration between Thinking Machines and investment firm Bridgewater. For a hedge fund, the alpha is the business, so we would expect Bridgewater to be particularly cautious in selecting an AI architecture. In this case, the two entities collaborated to train a custom model as an extension of Thinking Machines’ base one.
I believe they used Qwen3-235B as the base (rather than a Thinking Machines model) and post-trained it. Interestingly, since the post-training happened on Tinker, Bridgewater must have been Ok with their training data "leaving the building" and for the final post-trained weights (i.e., the final alpha) to be visible to Thinking Machines (Or, perhaps, Thinking Machines set up a private Tinker for Bridgewater).
I would assume that if regulatory compliance can be made a complementary asset it also means that countries have bought in to the convex returns that the frontier lab provides. Now, whether that remains to be the case in competition to an open source model is not yet decided.
Diffusion to build these models already did rely on a fair use argument to take from human written books, articles, art etc. Regulatory compliance couldn’t come in because of the fact that all of this tacit knowledge was not protected.
Now, open source models, as you make the valid point of they should be deployed differently: could be great for cyber security, not necessarily for gene research or physical security, make a lot of sense from a company’s investment standpoint.
Why not pay up big now and prevent disintermediation 5-10 years down the road?
Netflix depended heavily on AWS in the 2000s but did well to build its own cloud infrastructure before Prime Video became a thing. Imagine it had to make the move now.
I do think we are going to see heavy, heavy investment in the next few years for firms buying engineering capability on top of open source models to protect not only their knowledge but their business.
Not that OpenAI or Anthropic will see less growth but the argument of buy or build will be much more competitive than partner.
Reads like a handful of ( wealthy) companies, who acquired the base knowledge essentially for free, repurposed and mined it, and don’t want to share common knowledge without getting paid. Sounds like a glorified consultant!
But that’s exactly why they’ll be challenged by the companies with the actual knowledge, see the section on owning the weights of production.
It’s our knowledge. Just like our forefathers did exploiting land, minerals, trees. Took it.
Quite a game we humans have created.
This piece evokes a familiar line of reasoning: relaxed gun access benefits the firearms sector, civilians gain self-protection, widespread gun ownership creates community defense, hazards of gun violence receive little attention, and broad gun regulation ought to be removed. Unlike tangible weapons, AI weights may be copied endlessly and cannot be revoked remotely, creating safety challenges that are far more difficult to address.
That’s not the point of the piece. On safety, the answer is it depends on the marginal benefits vs marginal costs.