Shumer says local models are useless, Theo writes off most labs
Two blunt calls on model quality, and not a benchmark or a dollar figure between them.
Matt Shumer posted four words, "Local models are useless", and when the replies came in he answered with "So many people in denial, it's crazy". That is the whole argument. No eval, no cost per token, no task he tried and gave up on.
So many people in denial, it's crazy
This is the second time in a week he has taken the same swing. He called local models a nerd trap earlier and builders pushed back then too. The position has not moved and neither has the evidence behind it.
Theo goes after the hosted labs
Theo took a wider shot in the same stretch, and his target was the frontier labs rather than the people running weights on their own hardware. His post says Anthropic has no small models worth using right now, OpenAI has no large models worth using right now, and Google has no models worth using right now. Three sentences, three labs, one of them written off entirely.
Google has no models that are worth using right now.
Read them together and you get a strange picture. One founder says the local option is not real, the other says most of the hosted menu is not real either. If you believed both at once you would be left with a very short list of things to build on.
The business read
Neither post comes with a number. Not a benchmark score, not a latency figure, not a monthly bill. These are two people with strong priors stating them flatly, which is fine as a signal about where experienced builders are landing and useless as an input to your stack decision.
The useful move is the boring one. Whatever you are running today, the question is whether it passes your own evals on your own workload at a price you can live with. Shumer's conviction and Theo's list of write-offs tell you what they have concluded, not what your product needs. Nobody in these posts showed their work, so treat it as taste, not data.
Steal the numbers, not the vibes
One email a day. Under three minutes. Real MRR, real pricing, real launches from founders building with AI.

