Andrew Trask challenged the claim, dismissing any significant discrepancy.
there is now a very real capability & perhaps speed discrepancy between what models ai labs are using internally & what is available to the general public.
this gap has likely never been as far as it is now & it might actually widen over time.
it’s a real advantage even if it only lasts for just ~6 months at a time.
Many users criticized AI labs for hoarding advanced internal models through hype and artificial scarcity that disadvantages indie developers and the public.
@signulll meh... not really
there is now a very real capability & perhaps speed discrepancy between what models ai labs are using internally & what is available to the general public.
this gap has likely never been as far as it is now & it might actually widen over time.
it’s a real advantage even if it only lasts for just ~6 months at a time.
@signulll Makes you wonder if "publicly available" will start meaning something different in the next couple years.
Not a product tier. An actual structural gap in who gets to build with what.
There was a chart anthropic released about their coding velocity increase with each model and there's been an exponential increase since the last two models which they kept internally for months first (and mythos which is still internal). Gives them years of advantage in a span of months.
@signulll Are you saying this because of the release cadence speed that’s slowed down a little bit if not at least definitely not sped up in the way you would expect if they’re using their own models for their next release?
@signulll Most advantages don't come from knowing something nobody knows.
They come from recognizing what's happening before everyone agrees it's happening.
By the time everyone agrees, the advantage has already changed hands.
@signulll I don’t think the gap is nearly as big as using a highly-tuned harness vs. not.
@signulll this is humbling ngl we're out here optimizing prompts and chunking strategies while researchers are using models we wont see for 6 months the gap is wild
i'm looking forward for reading balance sheets of both OpenAI and Anthropic when they'll get listed..
difference between revenue and operating cost will change the dynamics i guess..may compell them to release their most capable models as fast as they can..
if gross profits are already good..then we'll never get to see their internal models..
@signulll Would be curious to lift the veil and see what their release pipeline looks like nowadays. Outside perspective is research/eng teams iterating on fine tuning a feature (ie improved images in gpt 5.5) plus bundling of improved general model capabilities for each version.
@signulll I wrote about why a couple months ago:
@signulll The gap between AI labs and public is widening
Labs keep the most advanced models internal first and it gives them a real edge even if only for months
@signulll the gap is the point. they build the hype engine then hoard the actual fuel.
@signulll ohh really.
@signulll they know they will become soon a commodity
@signulll artificial scarcity's a powerful tool. makes it tough for indie hackers.
@signulll yeah I feel like models capture <10% of the value they provide
therefore in most cases using them internally provides better roi
@signulll how much of that gap is the model vs the harness though. internal folks get no rate limits, no cost ceiling, tools wired straight into prod. i'd kill for that setup with today's public models
@signulll The 'compute gap' is becoming the new competitive moat. For us building on top of these models, it’s a constant challenge to design for portability
@signulll I believe this will become more substantial as recursive learning increases for the state-of-the-art models
@signulll The gap reveals inefficiencies in publicly available AI models
Andrew Trask challenged the claim, dismissing any significant discrepancy.
there is now a very real capability & perhaps speed discrepancy between what models ai labs are using internally & what is available to the general public.
this gap has likely never been as far as it is now & it might actually widen over time.
it’s a real advantage even if it only lasts for just ~6 months at a time.