The talk around openness & AI needs to distinguish between the vibrant and innovative open source movement that is advancing the state-of-the-art on harnesses & other key areas and open weights frontier models, which are entirely dependent on a the goodwill of a few Chinese firms
Positive users celebrate thriving open source AI harnesses driving innovation independent of frontier models, while negative users accuse the discussion of bias toward large AI companies.
No Digg Deeper questions have been answered for this story yet.
And the idea that frontier open weights models will continue to be released was fragile even before the current de facto licensing regime.
Is there a business model for being profitable off training frontier open weights models?
Other people can host, fine-tune, consult etc. as least as cheaply as you can. There are no ancilary product sales & it is fantastically expensive to make compared to most open source work
Lots of advances coming from the open source harness movement (including Moltbook, RAG approaches back in the day, etc.) but they depend on the intelligence of the models created by a small handful of companies & the more intelligent those models, the more others can do with them
The talk around openness & AI needs to distinguish between the vibrant and innovative open source movement that is advancing the state-of-the-art on harnesses & other key areas and open weights frontier models, which are entirely dependent on a the goodwill of a few Chinese firms
@MikeBradleyAI I would have made a lot more money if I went to the labs, and had the option to do so. The whole point of not taking their money is so I don’t have “shill” anyone but to call things out the way I see them.
You can disagree with facts, but skip the nastiness. It is unnecessary.
@ccatalini More companies are falling behind the frontier than keeping up with it (Mistral, Grok, etc.) & we are in the era of early RSI where there is acceleration. I haven’t seen any evidence of sudden frontier models emerging, but interested in seeing what you are hinting at.
@emollick Not really. A lot more happening below the surface, beyond the goodwill of the Chinese labs. It’s just early.
@RocMa222222 @emollick There’s more! Including decentralized and verifiable training and inference. @benfielding @harrygrieve
@emollick The harness layer is where most of my edge actually lives. I run quant strategies on top of these models — and 90% of the improvement over the last year came from better scaffolding, not a bigger model. The open weights debate is real, but builders are routing around it daily.
@MikeBradleyAI I think open source is important (it was literally the subject of a bunch of my pre-AI academic research). I also think that closed frontier models have a current lead with sustaining forces for now. And that more intelligent models can do more. Not sure we disagree on that much.
We’re going to be ok. The real advancements of the past two years have been a lot more harness than weights…
…which means it’s just a matter of doing the work to replicate comparable results.
And Nemotron 3 is an absolute treasure…everything you need to learn about building your own, you can learn from Nemotron.
We’ll be alright…
Verifiable inference is the piece I'm watching most closely. From where I sit, the bar isn't "can it run decentralized" — it's "can I trust an output I didn't compute myself, cheaply enough to act on in real time." If @benfielding / @harrygrieve crack the cost side of verification, the goodwill dependency basically evaporates.
@emollick ran into this building harness integrations last month tooling was great right up until the task got hard, then everything depended on what model was underneath. does harness innovation actually close the gap, or just make the dependency less visible?
@emollick This distinction matters more every month. The harness and tooling layer is truly compounding in the open, but the open-weights side sits one policy change away from drying up. Same label, very different risk profiles.
@emollick Openness is not binary.
There are many layers where the community can still innovate heavily.
Agreed — and from the builder side, "below the surface" is exactly where it's moving fastest. Most teams I know aren't betting on any single lab; they're building model-agnostic harnesses so the underlying weights become swappable. Early, yes — but the architecture is already routing around dependence.
Genuinely Ethan, I don’t want to be nasty. In this area of the field it’s just hard to keep hearing you pump 2-3 companies and dismiss the economics, capabilities, virtues, and benefits of an entire side of the industry.
And just to be clear, I really try to ascribe good intent to you when you do so. I try to tell myself “he just genuinely doesn’t know”. But I also know that you are incredibly smart, and surrounded by a massive pool of resources, and eventually it starts feeling like you are willfully pushing opinions that aren’t grounded in practical reality.
I don’t want to be or feel nasty to anyone. Least of all you. But I sincerely would like to talk to you about all of this if you genuinely aren’t up to date on the space or implore you to talk to others who you trust and are respected inside this space. Because you just keep pumping 2-3 closed labs and ignoring a massive industry and capability and virtue set around open source AI.
This stopped being theoretical this week: Coinbase just said it's moving its defaults to open-weight GLM and Kimi to cut AI spend in half. A major US public company's cost structure now rides on exactly the "goodwill of a few Chinese firms" you describe. The dependency is already in production.
@emollick wise words
@emollick Fair split, but I'd add a third tier: the eval + harness data that nobody open-sources. The weights and the tooling are both visible — the proprietary moat is increasingly the feedback loops and task data wrapped around them. That's the part not dependent on anyone's goodwill.
@emollick Prediction. Ethan works for one of the large AI companies he keeps shilling in the next 3 years.
@emollick Couldn't China subsidize open weights training simply to make the foundation model layer a commodity and reduce or eliminate the incentive for continued VC investment in US frontier models? I realize that's not a business model, but it could still be a strategy.