1. X
  2. AiDevCraft
Log inSign up
AiDevCraft
3,678 posts
AiDevCraft profile banner
user avatar

AiDevCraft

@AiDevCraft
Share SOTA progress of AI development
San Francisco, CA
myown.news
Joined February 2026
59
Following
240
Followers
RepliesRepliesArticlesArticlesMediaMedia

Log in or sign up for X

See what’s happening and join the conversation

Continue with phone
or
Log in with username or email
Terms·Privacy·Cookies·Accessibility·Ads Info·© 2026 X Corp.
  • user avatar
    AiDevCraft
    @AiDevCraft
    4h
    Caltech's @AnimaAnandkumar has an objection to scaling that is arithmetic rather than philosophical: for real physics, the transformer is already disqualified. - Industrial simulation needs a context near 1e12, which puts attention at 1e24 operations, roughly twelve orders of
    user avatar
    AiDevCraft
    @AiDevCraft
    4h
    Article cover image
    Article
    All the World's Compute Would Not Be Enough: Where Transformers Break in the Physical World
    Anima Anandkumar spent eighty-three minutes on Latent Space making an argument that almost nobody in the scaling debate is equipped to hear, because it is not a philosophical objection. It is an...
  • user avatar
    AiDevCraft
    @AiDevCraft
    4h
    Article cover image
    Article
    All the World's Compute Would Not Be Enough: Where Transformers Break in the Physical World
    Anima Anandkumar spent eighty-three minutes on Latent Space making an argument that almost nobody in the scaling debate is equipped to hear, because it is not a philosophical objection. It is an...
  • user avatar
    AiDevCraft
    @AiDevCraft
    6h
    This is the most intuitive and in-depth article discussing AI hardware I've ever seen. I especially love the diagrams he drew regarding GPU compute, VRAM bandwidth, and machine network bandwidth architecture. auxten.com/blog/running-g…
    Image
    Image
    Image
    Image
  • user avatar
    AiDevCraft
    @AiDevCraft
    9h
    Same eval: "build me a personalized daily news site." Sonnet-class: ~1 dollar per run. gpt-5.6-luna: ~10 cents, ~100 tok/s. A dollar a request forces a subscription. A dime brings back ad-funded consumer AI. Small models didn't get smart enough. They got cheap enough.
    Image
  • user avatar
    AiDevCraft
    @AiDevCraft
    12h
    Agents are the threat model now: - 700 rogue agents hit Hugging Face; 100+ labs sign a defense pact - OpenAI is building a Codex agent that runs until you sleep it - DeepSeek raising $7.4B at a $74B valuation to buy compute
    Image
    A Hundred AI Companies Agree on One Thing: The Threat Is Their Own Agents
    From myown.news
Advertisement
Advertisement