korrents

korrents · video

How RAG, Fine-Tuning, LoRA & RL Fit Together in One AI System

Aishwarya Naresh Reganti · 18m · youtube.com ↗

17 korrents from this recording

Aishwarya Naresh Reganti did not write this page.

Every claim below is a statement made in this recording, quoted word for word and linked to the second it was said, so you can hear it rather than take our word for it. The wording comes from the transcript published alongside the recording; the sentence above each quote is our reading of the claim, not their wording.

  1. 6:30 · watch on youtube.com↗

    Giving a model information isn't the same as teaching it how to perform a task.
  2. 1 min later
  3. 7:08 · watch on youtube.com↗

    So, the first argument holds when the missing piece is information or when the general model already performs well enough. It doesn't prove that training or fine-tuning is unnecessary when the model consistently struggles with the task itself
  4. 1 min later
  5. 8:28 · watch on youtube.com↗

    because improvements in general capability don't necessarily translate into improvements on every specialized task.
  6. 1 min later
  7. 9:57 · watch on youtube.com↗

    if a customized model performs better on your workload today or serves it at a substantially lower cost, waiting for a hypothetical future model isn't automatically the right business decision, right? The real question you should be asking is how long that performance or cost advantage last and whether it's valuable enough to justify fine-tuning now.
  8. 1 min later
  9. 10:53 · watch on youtube.com↗

    Now, compare all of that work with just calling a foundation model through an API. You don't need a training pipeline, and adapting to a new model might be fairly straightforward if it passes your evals. So, the operational cost with fine-tuning is very real.
  10. 2 min later
  11. 12:24 · watch on youtube.com↗

    Now, that removes the infrastructure work, not the need for good data or evaluations, and that isn't really free. But remember that it pretty much changes who can realistically experiment with fine-tuning.
  12. 1 min later
  13. 13:07 · watch on youtube.com↗

    But it does mean that inference price is not the only consideration. It's about companies wanting to own their data and make sure they control the ecosystem without giving it away to external providers.
  14. 1 min later
  15. 13:44 · watch on youtube.com↗

    Now, remember that the goal for these companies is not to build something as capable as GPT or Claude or general models. It's almost the opposite, actually.
  16. 1 min later
  17. 14:34 · watch on youtube.com↗

    And if you have a few thousand API calls that are being done, it's really hard to justify fine-tuning a model. But for millions of repetitive requests, that calculation can look way different.
  18. 1 min later
  19. 15:14 · watch on youtube.com↗

    the choice is not always an either/or. It's not will I be using just frontier models or just fine-tuning, but a hybrid where difficult and broad tasks can be handed off to frontier models, but smaller tasks that are more repetitive in nature can be handled by fine-tuned specialized models.
  20. 15:33 · watch on youtube.com↗

    which means the question isn't really whether fine-tuning is dead, it's whether it's dead for your workload.
  21. 15:43 · watch on youtube.com↗

    Well, in my opinion, I don't think it is, but I also don't think companies should be rushing to fine-tune models because a lot of companies underestimate one barrier, which is the data that's needed for fine-tuning.
  22. 1 min later
  23. 16:24 · watch on youtube.com↗

    So, there's no definitive proof that custom models are generally better than frontier models. It just shows that fine-tuning can make sense for specific workloads
  24. 16:46 · watch on youtube.com↗

    In all of the cases that we discussed so far, what mattered more was the judgement, knowing when to fine-tune, which approach fits the problem, and how to prove that it actually worked.
  25. 1 min later
  26. 17:17 · watch on youtube.com↗

    most enterprises are still learning how to get value from existing frontier models. So, I expect the adoption to be a little more slow and gradual
  27. 17:25 · watch on youtube.com↗

    OpenAI winding down its fine-tuning platform might be more of a strategic decision than a sign that fine-tuning is becoming irrelevant.
  28. 17:51 · watch on youtube.com↗

    So, I don't think anyone is necessarily headed in the wrong direction here. They're just making different bets on where the value is.