undefined | Better HN

0 pointsgruez6h ago0 comments

>when the competition is.. near-equivalent-capability DeepSeek/Qwen/Llama on a $1k GPU with a break-even of 5 months of subscription costs

Lol no. Chinese AIs are definitely not "near-equivalent-capability". The empirical proof is pretty obvious: how many people have you heard talking about using their codex/claude code subscription vs their z.ai or qwen subscription? Moreover even the Chinese models require epic amounts of GPUs to run the full version, eg. https://apxml.com/models/glm-51 needs 1515 GB to run, and that's with a measly 1024 token context. To get it to run on your "$1k GPU" you'd need to quantize it, making it even dumber.

0 comments

parliament326h ago

Today, sure. But we already see diminishing returns with Claude releases, and we know the open models are closing the gap (~6 months behind according to the benchmarks). And when the pitch is "our models are 5% better but cost $200/mo.. also here's a mountain of restrictions" it just won't make sense anymore. Give it a year or two.

I could see the "avoid the hardship of running a local model for $20/mo" angle but Anthropic has shown they have little interest in those customers.

gruezOP5h ago

>and we know the open models are closing the gap (~6 months behind according to the benchmarks).

Looking at just the benchmarks might be misleading: https://x.com/scaling01/status/2050616057191072161

parliament324h ago

Good article. But it concludes with "Open models may be only 4–5 months behind on coding-heavy, benchmark-visible tasks... the gap is likely much larger and closer to 8 months."

So, fine. In 2024, being 8 months behind was massive. In 2025, pretty big. This year.. I guess CC has improved a bit between October and now? How much do you think it'll matter a few more years down the line?

Even now.. I bet a non-trivial number of people would happily be 8 months behind just to avoid another rent-seeker. And this will only get worse over time, which makes it an unwinnable situation for Anthropic. Hence all the panicked flailing about with restricting tooling and trying to get something even resembling a moat.

j / k navigate · click thread line to collapse