Charlie O'Neill
speaker
259 appearances
2 recordings
1 series
first heard Jul 2026
last heard 15 Sep
Charlie O'Neill’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsRecordings per month over the last 12 months — 2 in all, peaking in Sep 2026 with 1.
Appearances
But at the moment, it doesn't seem like open source versus closed source.
The intelligence ceiling that we've gotten to hasn't led to any increased concerns around, you know,
can I use this model through open source or closed source?
The risk of developing a bioweapon, for instance, is about the same in either case.
I think there's developers who have a very inelastic demand for the frontier intelligence and will always want to use the most intelligent models.
And then there's the ecosystem and the economy in general.
The way I like to think of it is that for all the economically valuable tasks that we could plausibly use an LLM for, there is some intelligence threshold.
at which below that it's very difficult to do the task, and above that you're getting very diminishing returns to having more and more intelligent models, and usually intelligence is correlated with cost.
So the obvious argument here is that there is margin pressure on all these startups, all these companies, even Enterprise now who are doing this particular task with LLMs.
They've hit the threshold of intelligence probably even a while ago with open source,
Open source has been accelerating rapidly and you just don't need a Fable or Mythos level model in order to do some of these things and you get exactly the same performance if you use a model that's a tenth of the size or even a fiftieth of the size.
Post-training is also really important here because it means you can teach a much smaller model to do one thing really, really well as opposed to taking an off-the-shelf open source or closed source model and trying to prompt engineer your way to doing that task.
So post-training really changes the economics here.
And of course, you can only post-train on open source models because you can actually touch the weights as opposed to closed source.
And so I think margin pressure is a big one.
Another one is like Anthropic and OpenAI, I think, are realizing that the recipe is the same amongst all these companies.
Like there is no secret sauce.
Yes, there's probably a long tail of optimization, small optimizations that Anthropic and OpenAI have that the rest of the ecosystem doesn't have.
But their moat is no longer in there being
them being the only ones who can train these very, very large multi-trillion parameter models, their moat now is starting to shift towards, okay, well, if we have a little bit of a head start, what if we try and hit particular verticals?
Showing 181–200 of 259 · page 10 of 13
← Previous
Next →