---
schema: 1
type: video
title: 'Bringing Continual Learning into Enterprises — Samuel Denton, Applied Compute'
description: >-
  A Qwen thinking model was taking up to 80 turns to submit on SWE bench.
  Applied Compute wanted it wrapping up by turn 40 and got the submit tool call
  rate from 22% to 60% with test pass rate flat. The interesting part is the
  mechanism: because the rollout was conditioned on an old production trace that
  never called the tool, the teacher never touched the tool call tokens at all.
  It moved the reasoning path toward the call instead, and the call followed.
  Sam Denton's frame is a grid. One axis is how online the traces are, from a
  single dump of production traces to a unified engine where serving and
  training are the same loop. The other is where the hint comes from, either
  static priors, such as knowing a support agent is too quick to refund, or a
  hint built dynamically from what the on policy model just did. Applied Compute
  works two corners of that grid. Offline hints on offline traces need no
  replayable environment and can improve an enterprise agent from a data dump on
  day one. Online hints on online traces have the far higher ceiling, and that
  is what fixed a customer whose harness required unusual hyperlink formatting:
  rewarding the format directly and finetuning on correct examples both degraded
  coding ability, while a hint written against each rollout took correct
  formatting from 15% to 80%. Two things he says make it work in practice. Let a
  judge pick where in the rollout the hint goes and distill only the next few
  steps, since the learning signal decays with distance from the hint. And mask
  which tokens you learn from, because the teacher has strong opinions about
  connector words that have nothing to do with the lesson. Throughout, the
  constraint he keeps is doing all of this without a golden answer to distill
  toward. Speaker info: - https://x.com/samueldenton -
  https://www.linkedin.com/in/sam-denton-161b50126/ Timestamps: 0:00 - The
  distillation spectrum, offline to online 2:46 - The holy grail: serving and
  training as one loop 4:00 - Where the hint comes from 4:42 - Online hints
  built from the rollout 5:19 - Four quadrants of distillation 7:50 - The two
  corners they actually work in 9:44 - Improve for free today, raise ceilings
  tomorrow 10:22 - Doing it without a golden answer 11:00 - SWE bench: wrapping
  up by turn 40 11:38 - The three metrics that matter 12:17 - What the hint
  actually says 12:55 - Moving the reasoning path, not the tool call 13:36 -
  Adding a single on policy step 14:17 - The hyperlink formatting problem 14:56
  - Why rewards and finetuning both failed 15:34 - From 15% to 80% with online
  hints 16:13 - Per step hinting 16:50 - Why the signal decays with distance
  17:27 - Relevance masked self distillation 18:07 - What it adds up to
canonical: >-
  https://liquidgold.cc/channel/aidotengineer/explore/video/bringing-continual-learning-into-enterprises-samuel-3f2536976ad8
markdown: >-
  https://liquidgold.cc/channel/aidotengineer/video/bringing-continual-learning-into-enterprises-samuel-3f2536976ad8.md
llms: null
parent: 'https://liquidgold.cc/channel/aidotengineer.md'
channel: aidotengineer
updated: '2026-10-24T05:35:14.307Z'
language: null
children: []
videoId: ZTA0GwpAUak
durationSeconds: 1143
publishedAt: '2026-08-12T17:30:06.000Z'
searchChunks:
  - ordinal: 0
    text: null
    startSeconds: 23
    endSeconds: 166
    watch: 'https://liquidgold.cc/watch/ZTA0GwpAUak?t=23'
    heldReason: rights
  - ordinal: 1
    text: null
    startSeconds: 177
    endSeconds: 244
    watch: 'https://liquidgold.cc/watch/ZTA0GwpAUak?t=177'
    heldReason: rights
  - ordinal: 2
    text: null
    startSeconds: 244
    endSeconds: 256
    watch: 'https://liquidgold.cc/watch/ZTA0GwpAUak?t=244'
    heldReason: rights
  - ordinal: 3
    text: null
    startSeconds: 256
    endSeconds: 290
    watch: 'https://liquidgold.cc/watch/ZTA0GwpAUak?t=256'
    heldReason: rights
  - ordinal: 4
    text: null
    startSeconds: 302
    endSeconds: 324
    watch: 'https://liquidgold.cc/watch/ZTA0GwpAUak?t=302'
    heldReason: rights
  - ordinal: 5
    text: null
    startSeconds: 335
    endSeconds: 480
    watch: 'https://liquidgold.cc/watch/ZTA0GwpAUak?t=335'
    heldReason: rights
  - ordinal: 6
    text: null
    startSeconds: 593
    endSeconds: 604
    watch: 'https://liquidgold.cc/watch/ZTA0GwpAUak?t=593'
    heldReason: rights
  - ordinal: 7
    text: null
    startSeconds: 640
    endSeconds: 661
    watch: 'https://liquidgold.cc/watch/ZTA0GwpAUak?t=640'
    heldReason: rights
  - ordinal: 8
    text: null
    startSeconds: 672
    endSeconds: 704
    watch: 'https://liquidgold.cc/watch/ZTA0GwpAUak?t=672'
    heldReason: rights
  - ordinal: 9
    text: null
    startSeconds: 792
    endSeconds: 824
    watch: 'https://liquidgold.cc/watch/ZTA0GwpAUak?t=792'
    heldReason: rights
  - ordinal: 10
    text: null
    startSeconds: 857
    endSeconds: 869
    watch: 'https://liquidgold.cc/watch/ZTA0GwpAUak?t=857'
    heldReason: rights
  - ordinal: 11
    text: null
    startSeconds: 901
    endSeconds: 913
    watch: 'https://liquidgold.cc/watch/ZTA0GwpAUak?t=901'
    heldReason: rights
  - ordinal: 12
    text: null
    startSeconds: 934
    endSeconds: 945
    watch: 'https://liquidgold.cc/watch/ZTA0GwpAUak?t=934'
    heldReason: rights
  - ordinal: 13
    text: null
    startSeconds: 979
    endSeconds: 990
    watch: 'https://liquidgold.cc/watch/ZTA0GwpAUak?t=979'
    heldReason: rights
  - ordinal: 14
    text: null
    startSeconds: 1012
    endSeconds: 1023
    watch: 'https://liquidgold.cc/watch/ZTA0GwpAUak?t=1012'
    heldReason: rights
  - ordinal: 15
    text: null
    startSeconds: 1023
    endSeconds: 1060
    watch: 'https://liquidgold.cc/watch/ZTA0GwpAUak?t=1023'
    heldReason: rights
---


# Bringing Continual Learning into Enterprises — Samuel Denton, Applied Compute

A Qwen thinking model was taking up to 80 turns to submit on SWE bench. Applied Compute wanted it wrapping up by turn 40 and got the submit tool call rate from 22% to 60% with test pass rate flat. The interesting part is the mechanism: because the rollout was conditioned on an old production trace that never called the tool, the teacher never touched the tool call tokens at all. It moved the reasoning path toward the call instead, and the call followed. Sam Denton's frame is a grid. One axis is how online the traces are, from a single dump of production traces to a unified engine where serving and training are the same loop. The other is where the hint comes from, either static priors, such as knowing a support agent is too quick to refund, or a hint built dynamically from what the on policy model just did. Applied Compute works two corners of that grid. Offline hints on offline traces need no replayable environment and can improve an enterprise agent from a data dump on day one. Online hints on online traces have the far higher ceiling, and that is what fixed a customer whose harness required unusual hyperlink formatting: rewarding the format directly and finetuning on correct examples both degraded coding ability, while a hint written against each rollout took correct formatting from 15% to 80%. Two things he says make it work in practice. Let a judge pick where in the rollout the hint goes and distill only the next few steps, since the learning signal decays with distance from the hint. And mask which tokens you learn from, because the teacher has strong opinions about connector words that have nothing to do with the lesson. Throughout, the constraint he keeps is doing all of this without a golden answer to distill toward. Speaker info: - https://x.com/samueldenton - https://www.linkedin.com/in/sam-denton-161b50126/ Timestamps: 0:00 - The distillation spectrum, offline to online 2:46 - The holy grail: serving and training as one loop 4:00 - Where the hint comes from 4:42 - Online hints built from the rollout 5:19 - Four quadrants of distillation 7:50 - The two corners they actually work in 9:44 - Improve for free today, raise ceilings tomorrow 10:22 - Doing it without a golden answer 11:00 - SWE bench: wrapping up by turn 40 11:38 - The three metrics that matter 12:17 - What the hint actually says 12:55 - Moving the reasoning path, not the tool call 13:36 - Adding a single on policy step 14:17 - The hyperlink formatting problem 14:56 - Why rewards and finetuning both failed 15:34 - From 15% to 80% with online hints 16:13 - Per step hinting 16:50 - Why the signal decays with distance 17:27 - Relevance masked self distillation 18:07 - What it adds up to

Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/bringing-continual-learning-into-enterprises-samuel-3f2536976ad8 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md)

> Search chunk text is held until public-source provenance is verified.

## Document chunks

> Missing: no stored document chunks are available.