Embed Notice

HTML Code

<blockquote style="position: relative; padding-left: 55px;"><section><a href="https://mastodon.social/users/caseynewton/statuses/114027933122710689">Casey Newton (caseynewton@mastodon.social)'s status on Wednesday, 19-Feb-2025 10:23:32 JST</a><a href="https://mastodon.social/@caseynewton" title="caseynewton@mastodon.social"><img src="https://gnusocial.jp/avatar/61006-48-20240126011805.webp" width="48" height="48" alt="Casey Newton" style="position: absolute; left: 0; top: 0;">Casey Newton</a></section><article><p>Training Grok 3 took Elon 200,000 GPUs and untold billions, and it's ... decent at best? I wrote about AI's commodity problem</p><p><a href="https://www.platformer.news/too-many-ai-companies/" rel="nofollow">https://www.platformer.news/too-many-ai-companies/</a></p></article><footer><a rel="bookmark" href="https://gnusocial.jp/conversation/4590413#notice-8986261">In conversation</a><time datetime="2025-02-19T10:23:32+09:00" title="Wednesday, 19-Feb-2025 10:23:32 JST">about 3 days ago</time> <span>from <span><a href="https://mastodon.social/@caseynewton/114027933122710689" rel="external" title="Sent from mastodon.social via ActivityPub">mastodon.social</a></span></span><a href="https://mastodon.social/@caseynewton/114027933122710689">permalink</a><h4>Attachments</h4><ol><li><label><a rel="external" href="https://gnusocial.jp/attachment/4162385">On one hand, it’s legitimately striking that xAI could stand up a data center with 200,000 GPUs and train a decent model in less than two years. On the other hand, it would have been more striking had DeepSeek not accomplished a similar feat at a tiny fraction of the cost, and with worse technology. It’s too early to say with any certainty how good Grok’s models are. Leading researcher Andrej Karpathy finds that it passes many of the tests he throws at it, but it also fails almost as many. Even if you accept, though, that this is a state-of-the-art model, not a single person working in AI believes it will stay there for long. Leading AI labs push out new models every few days, and any innovations are almost all quickly copied and absorbed by their rivals.</a></label><br><a href="https://files.mastodon.social/media_attachments/files/114/027/932/875/874/457/original/6a747cd7d7fcf661.png" rel="external">https://files.mastodon.social/media_attachments/files/114/027/932/875/874/457/original/6a747cd7d7fcf661.png</a></li></ol></footer></blockquote>

Corresponding Notice

Embed this notice
Casey Newton (caseynewton@mastodon.social)'s status on Wednesday, 19-Feb-2025 10:23:32 JST Casey Newton
Training Grok 3 took Elon 200,000 GPUs and untold billions, and it's ... decent at best? I wrote about AI's commodity problem
https://www.platformer.news/too-many-ai-companies/
In conversationabout 3 days ago from mastodon.socialpermalink
Attachments
1. On one hand, it’s legitimately striking that xAI could stand up a data center with 200,000 GPUs and train a decent model in less than two years. On the other hand, it would have been more striking had DeepSeek not accomplished a similar feat at a tiny fraction of the cost, and with worse technology. It’s too early to say with any certainty how good Grok’s models are. Leading researcher Andrej Karpathy finds that it passes many of the tests he throws at it, but it also fails almost as many. Even if you accept, though, that this is a state-of-the-art model, not a single person working in AI believes it will stay there for long. Leading AI labs push out new models every few days, and any innovations are almost all quickly copied and absorbed by their rivals.
  https://files.mastodon.social/media_attachments/files/114/027/932/875/874/457/original/6a747cd7d7fcf661.png

Public

Embed Notice

HTML Code

Corresponding Notice