LLenny's Podcast
← All frameworks
StrategyGarrett Lord (Handshake CEO)

The Quality-Volume-Speed Priority Stack

What high-stakes data buyers actually rank: quality first, then volume, then speed

Difficulty
Moderate
Time to result
~weeks to results
Steps
3
Confidence
90%

A three-part priority order for anyone supplying training data (or any high-stakes B2B input) to frontier AI labs. Buyers rank quality first, then the ability to reach volume, then speed of turnaround. Understanding the order tells you where to invest: quality gates before scale, scale infrastructure before optimizing for turnaround.

Origin

Articulated by Garrett Lord based on Handshake AI's direct work with seven frontier labs.

Core principles

  • 01Quality is non-negotiable and first: bad data is like teaching a student the wrong thing — extremely hard to overcome once trained in
  • 02Volume is the second constraint: producing thousands of pieces of high-quality data in advanced domains is a scale problem, not just a quality problem
  • 03Speed is third but decisive: labs run 3-4 parallel experiment pipelines and double down on whichever shows a gain, so days-not-weeks turnaround wins the next batch
  • 04The supplier who can go from high quality to high volume to fast turnaround captures effectively unlimited demand

How to run it

  1. 1

    Lock quality first

    Build the capability to produce genuinely correct, expert-grade data and the internal apparatus to verify it — a post-training team, per-unit quality assessment, and instructional design to train contributors.

    Pro tip Approximate the actual model gain from each unit of data so you can prove quality, not just assert it.

    Watch out Wrong data poisons the model and is extremely challenging to overcome — never trade quality for volume or speed.

  2. 2

    Engineer for volume

    Solve the supply problem: reach enough qualified experts to produce thousands of high-quality units in the hardest domains (physics, chemistry, math).

    Pro tip Go straight to the source of density — e.g. top-GPA students at the best physics departments (Stanford, Berkeley, MIT) — rather than broad, thin sourcing.

  3. 3

    Optimize for speed of turnaround

    Build technology to assess each unit fast and turn batches around in days, so you can feed a lab's winning pipeline the moment their experiment shows a gain.

    Pro tip Sit directly with the researchers and share what you're seeing in the data so you can grow the pipeline that's working and drop the ones that aren't.

    Watch out Labs ditch two or three pipelines the instant one shows improvement — if you're slow, you miss the pipeline that's about to scale.

In the wild

Serving parallel lab experiments

A researcher runs three or four batches at once; as soon as one shows a gain, they grow that pipeline and ditch the others. Handshake built per-unit assessment tech and its own post-training team to turn data around in days at high volume and quality.

Working with seven frontier labs and facing effectively unlimited demand — 'if you can produce high quality volumes of data, you most likely will be able to sell whatever you produce.'

Common mistakes

Chasing volume before nailing quality

Scaling low-quality output just poisons more models and destroys trust with buyers who rank quality first.

Treating speed as a nice-to-have

Because labs reallocate to winning pipelines in days, slow turnaround means losing the exact batch that was about to scale — speed is a revenue lever, not just service quality.

Is it for you?

Best for

Founders and operators selling a high-stakes input (data, research, components) to sophisticated technical buyers running rapid experiments

Not ideal for

Commodity markets where buyers optimize purely on price and quality is undifferentiated

From the transcript

They care about like quality first and foremost. You have to have high quality data

20:00

the other huge problems you have is like volume like how how do you generate thousands of pieces of data in the most advanced domains

20:30

the other thing I'd say model builders care about is speed because they have all these hypotheses and they're constantly testing different pipelines

21:00

if you can produce high quality volumes of data, uh you most likely will be able to sell whatever you produce

48:00

From the episode

Inside the expert network training every frontier AI model

Garrett Lord (Handshake CEO)