Hacker News new | past | comments | ask | show | jobs | submit login
Crickets on the fact that US-based AI has been a shambolic enterprise
1 point by fl4tul4 18 hours ago | hide | past | favorite | 8 comments
Why nobody is recognising the feat by the Chinese? 5% of computing power to reach the same outcomes as OpenAI's ChatGPT. All I hear is crickets on an open field.





This got submitted to HN just a few minutes before your post [1].

When ChatGPT-3 came out I said two things: (1) this is not going to develop godlike intelligence even if you scale it up considerably [2], but (2) we had no idea how these work so that particular model was terribly inefficient and there will be future models that are more efficient.

And here we are. First-mover advantage turned out to be a disadvantage. If you are OpenAI or Facebook and have a good model you're going to be inclined to take what you've got and try to improve its accuracy. The right thing to do is throw it out and start all over using everything we know about how to increase efficiency, but there's the fear that it won't be as good as the original model, customers won't want to rebuild all their RAG vectors, etc. See [4]

[1] https://garymarcus.substack.com/p/five-things-most-people-do...

[2] they get the right answer by the wrong method, hallucinations won't go away, past a certain point trying to improve performance will be like pushing a bubble around under a rug, see [3] Structurally inadequate for many tasks such as sorting (that N log N thing) and chess (look at a few million moves with alpha-beta and you can beat the worst player at the chess club, look at a few billion you beat grandmasters)

[3] https://en.wikipedia.org/wiki/Asymptote

[4] https://en.wikipedia.org/wiki/Disruptive_innovation


I feel this has been just a way to inflate costs and infrastructure expenses (NVIDIA, and so on).

They never cared about tackling performance, just making money. They deserve everything coming their way (market loss, lack of trust).

Let's see how things unfold.


Perhaps crickets because nobody wants to wake the beast (market sell-off). But I believe it's a tempest in a teapot. Technology is changing so rapidly now (almost vertical rate of change) that by this time next year this "revelation" is likely to be moot. Just stick to what you trust in your work, and keep moving forward. The problems will be figured out and fixed by the market very quickly--or else.

Where exactly do you get your news that you have heard nothing about this?

almost nothing here .

Except the ten different articles that have made the frontpage about DeepSeek R1 in the last week. It’s probably almost 20 by now, I’ve lost count. There’s three on the frontpage now (there were five+ yesterday) and they’re full of the exact discussion you’re looking for. It’s by far the most popular topic of conversation on HN right now.

You’re suffering from an acute case of confirmation bias.



Nobody? I have nothing else in the British Press.



Consider applying for YC's Spring batch! Applications are open till Feb 11.

Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: