LLenny's Podcast
← All frameworks
InnovationAsha Sharma (CVP of AI Platform at Microsoft)

Adapt the Model, Don't Build One

Post-training is the new pre-training — steer an off-the-shelf model to your outcome instead of pre-training your own

Difficulty
Advanced
Time to result
~months to results
Steps
4
Confidence
90%

A decision rule for how to invest AI budget: above roughly 30 billion parameters, the capex to pre-train your own foundation model stops making economic sense. Instead, take a strong off-the-shelf model and adapt it via post-training — fine-tuning or reinforcement learning on your own, purchased, or synthetic data — to optimize for a specific outcome.

Origin

Asha Sharma, drawing on Nathan Lambert's study of top leaderboards showing the ~30B-parameter post-training threshold, and survey data that ~50% of developers now fine-tune. She frames post-training spend eventually exceeding pre-training spend.

Core principles

  • 01Above ~30B parameters, pre-training your own model rarely makes economic sense
  • 02You get more economic and 'taste' leverage by steering an existing model than building from scratch
  • 03Adaptation data can be your own, purchased, synthetic, or a mix
  • 04Same industrial logic as ranking — never ship what's off the shelf as-is; tailor it to your use case
  • 05Prefer a model system / ensemble over 'one model to rule them all'

How to run it

  1. 1

    Default to adapting, not building

    Assume you will start from a strong off-the-shelf model rather than invest heavily in pre-training your own, given the compute, science expertise, and capex required.

    Pro tip Rising scientist costs and scarce expertise make DIY pre-training even less attractive — let someone else pay that capex.

  2. 2

    Choose the model(s) per use case

    Select from the many good domain models by what each task needs — latency, thinking time, retrieval speed, quality — favoring a model system or ensemble over a single model.

    Pro tip Different models suit different jobs (e.g. one is great where you want quick retrieval, another where thinking time is fine); match model to task.

  3. 3

    Decide your adaptation data source

    Pick the data to steer the model: your own proprietary data (strongest), purchased data, synthetic data, or a combination.

  4. 4

    Post-train toward the outcome and run the full loop

    Apply fine-tuning or reinforcement learning to optimize the model for price, performance, or quality — and go through the full loop rather than stopping at basic fine-tuning.

    Pro tip Fine-tuning alone helps, but running the complete rewards-and-iteration loop yields materially better results.

In the wild

GitHub Copilot ensemble

GitHub uses an ensemble of models fine-tuned across ~30 countries and their languages, iterating in a loop for next-edit suggestions and code completions.

Tailored, multi-language code assistance from adapted models rather than one monolithic pre-trained model.

Common mistakes

Investing to pre-train your own large model

Above ~30B parameters the capex and scarce science expertise rarely pay off versus adapting an existing model.

Shipping the off-the-shelf model as-is

Like taking raw ranking output or default React components, an un-adapted model isn't tailored to your specific users or use cases.

Is it for you?

Best for

Founders and product leaders deciding how to spend AI budget who have a specific outcome and access to relevant data

Not ideal for

Frontier labs whose entire business is pushing the pre-training frontier, or teams with no differentiating data to adapt on

From the transcript

once a model hits 30 billion parameters the capex to actually train a model and put you know billions of tokens into a a kind…

06:30

I believe we will see you know just as much money spent on post-training uh as we will on pre-training and in the future more…

45:00

more and more companies and organizations kind of start to think about how do I adapt a model rather than how do I take something…

47:30

I'm much more in the like model system uh rather than one model to rule them all

48:30

From the episode

How 80,000 companies build with AI: products as organisms, the death of org charts, and why agents will outnumber employees by 2026

Asha Sharma (CVP of AI Platform at Microsoft)