Your brief doesn't make sense, maybe you need to expand?
They're only offering 3.5 for legacy reasons: pre-Deepseek, 3.5 did legitimately have some things that open source hadn't caught up on (like world knowledge, even as an old model), but that's done.
Now the wins come from relatively cheap post-training, and a random Chinese food delivery companies can spit out 500B parameter LLMs that beats what OpenAI released a year ago for free with an MIT license.
Also as you release models you're enabling both distillation of your own models, and more efficent creation of new models (as the capabilities of the LLM themselves are increasingly useful for building, data labeling, etc.)
I think the title is inflammatory, but the reality is if AGI is really around the corner, none of OpenAI's actions are consistent with that.
Utilizing compute that should be catapulting you towards the imminent AGI to run AI TikTok and extract $20 from people doesn't add up.
They're on a treadmill with more competent competitors than anyone probably expected grabbing at their ankles, and I don't think any model that relies on them pausing to cash in on their progress actually works out.
Longcat-flash-thinking is not super popular right now; it doesn't appear on the top 20 at open router. I haven't used it, but the market seems to like it a lot less than grok, anthropic or even oAI's open model, oss-20b. Like I said I haven't tried it.
And to your point, once models are released open, they will be used in DPO post-training / fine-tuning scenarios, guaranteed, so it's hard to tell who's ahead by looking at an older open model vs a newer one.
Where are the wins coming from? It seems to me like there's a race to get efficient good-enough stuff in traditional form factors out the door; emphasis on efficiency. For the big companies it's likely maxing inference margins and speeding up response. For last year's Chinese companies it was dealing with being compute poor - similar drivers though. If you look at DeepSeek's released stuff, there were some architectural innovations, thinking mode, and a lottt of engineering improvements, all of which moved the needle.
On treadmills: I posit the oAI team is one of the top 4 AI teams in the world, and it has the best fundraiser and lowest cost of capital. My oAI bull story is this: if capital dries up, it will dry up everywhere, or at the least it will dry up last for a great fundraiser. In that world, pausing might make sense, and if so, they will be able to increase their cash from operations faster than any other company. While a productive research race is on, I agree they shouldn't pause. So far they haven't had to make any truly hard decisions though -- each successive model has been profitable and Sam has been successful scaling up their training budget geometrically -- at some point the questions about operating cashflow being deployed back to R&D and at what pace are going to be challenging. But that day is not right now.
You're arguing a different point than the article, and in some ways even agreeing with it.
The article is not saying OpenAI must fail: it's saying OpenAI is not "The AGI Company of San Francisco". They're in the same bare knuckle brawl as other AI startups, and your bull case is essentially agreeing but saying they'll do well in the fight.
> In fact, the only real difference is the amount of money backing it.
> Otherwise, OpenAI could be literally any foundation model company, [...] we should start evaluating OpenAI as just another AI startup
Any startup would be able to raise with their numbers... they just can't ask for trillions to build god-in-a-box.
It's going to be a slog because we've seen that there are companies that don't even have to put 1/10th their resources into LLMs to compete robustly with their offerings.
OpenRouter doesn't capture 1/100th of open weight usage, but more importantly the fact that Longcat is legitimately robustly competitive to SOTA models from a year ago is the actual signal. It's a sneak peak of what happens if the AGI case doesn't pan out and OpenAI tries to get off the treadmill: within a year a lot of companies catch up.
They're only offering 3.5 for legacy reasons: pre-Deepseek, 3.5 did legitimately have some things that open source hadn't caught up on (like world knowledge, even as an old model), but that's done.
Now the wins come from relatively cheap post-training, and a random Chinese food delivery companies can spit out 500B parameter LLMs that beats what OpenAI released a year ago for free with an MIT license.
Also as you release models you're enabling both distillation of your own models, and more efficent creation of new models (as the capabilities of the LLM themselves are increasingly useful for building, data labeling, etc.)
I think the title is inflammatory, but the reality is if AGI is really around the corner, none of OpenAI's actions are consistent with that.
Utilizing compute that should be catapulting you towards the imminent AGI to run AI TikTok and extract $20 from people doesn't add up.
They're on a treadmill with more competent competitors than anyone probably expected grabbing at their ankles, and I don't think any model that relies on them pausing to cash in on their progress actually works out.