Ask HN: How are much smarter AI models made?

I am curious what actually happens between two generations of AI models.

For example, how do you go from Sonnet to Opus? Is Opus trained from scratch, built on Sonnet, or mostly the same model with more compute and training?

And how do models like Astra suddenly make a big jump in some capabilities? What is stopping Anthropic, Mistral, or others from doing the same thing? Is the main difference just more compute and money, or are there training methods, data, architecture, and research breakthroughs that competitors may not know about?

I can't think of a better place to ask this. I am guessing there are people here who actually work on these models and know what goes on behind the scenes.

11 points | by superasn 2 days ago

6 comments

  • XTXinverseXTY 2 days ago
    Scaling laws project that a model with more parameters trained for longer on more data yields predictably better performance, and that generally you want to scale these factors commensurately. More of the compute budget is being spent on RLVR [0] for which we also fit scaling laws

    Researchers tweak data mix, reward shape, model architecture, etc etc, breakthroughs which reduce the cost to train a just-as-smart model. But this increases the returns to scale, which further incentivizes bigger models trained for longer on more data

    [0] "...to run reinforcement learning training...at pretraining scale." https://x.ai/news/grok-4?_bhlid=b9339d7816a05adeb52bae7050cc...

    [-]
    • arbabyousaf 12 hours ago
      thanks, what you just taught means diamonds for me
  • davedx 1 day ago
    Lots of RLHF - reinforcement learning from human feedback.

    Also: "distillation" - seeding or running training sessions on the output of frontier models.

  • khantto 16 hours ago
    many things to push it forward, like:more data, more post trainning...
  • in_making 1 day ago
    compute and improvements in pretraning and RL
  • kelvo_ran 17 hours ago
    [dead]
  • mooo7 2 days ago
    [flagged]