# A clear target lets a model try what people have not demonstrated

[Canonical reading](https://escapemeridian.com/readings/millidge-learning-beyond-demonstrations)

Millidge argues that models can attempt harder goals than people can demonstrate when success is verifiable. Whether that learning transfers far enough to remove human guidance remains uncertain.

Headline and sourced paraphrase by Escape Meridian. Inclusion is not endorsement.

## Original source

Dwarkesh Podcast · Panel transcript

Original source date: Sep 11, 2026 (2026-09-11).

Date basis: Publication date displayed on the official episode page.

Publisher transcript read on 23 September 2026, New York time. Timestamp marks a chapter, not an exact quote. Audio/video not independently listened to.

The date this reading was added to Escape Meridian is not recorded.

- [AI researchers debate how close we are to recursive self-improvement](https://www.dwarkesh.com/p/john-beren-charlie)
- [28:06 — Learning through testable attempts](https://www.youtube.com/watch?v=PrSf7IOYu-I&t=1686)

## Limits

An easy-to-score task can omit the reason the work mattered. This argument is a reason to test transfer, not assume it.


## Editorial test, not a stated concession

Test whether improvements against the chosen target also improve outcomes on new cases, including failures the target does not directly measure.


## supports

Millidge contrasts copying demonstrated solutions with learning from attempts in environments that can check success. He remains uncertain whether that learning transfers far enough to remove human guidance.

- [Supporting source](https://www.dwarkesh.com/p/john-beren-charlie)

## Source and contributor context

Beren Millidge · Zyphra CTO, introduced in the panel

Researcher developing open models at Zyphra, as introduced in the episode.

A technical argument about learning; it does not demonstrate a result in a specific business setting.

Declared interests are limited to the records linked here; absence of a record does not establish absence of an interest.

- [Panel introduction](https://www.dwarkesh.com/p/john-beren-charlie)
- [Works at Zyphra, which develops open models.](https://www.dwarkesh.com/p/john-beren-charlie)

## Assumptions

Editorial reconstruction: the outcome can be checked and captures the purpose of the work.


## Source media

Opens the recorded source chapter; the linked chapter may begin before the passage discussed.

- [Watch from 28:06 · Dwarkesh Podcast](https://www.youtube.com/watch?v=PrSf7IOYu-I&t=1686)

## Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.


## Beren Millidge

Beren Millidge is introduced as Zyphra's chief technology officer developing open models in the September 2026 Dwarkesh panel on AI research.

- [Explore this context](/entities/beren-millidge)
- [Reference](https://www.dwarkesh.com/p/john-beren-charlie)

## Dwarkesh Podcast

Dwarkesh Podcast features long conversations with researchers, builders and thinkers about technology, history and the future.

- [Explore this context](/entities/dwarkesh-podcast)
- [Reference](https://www.dwarkesh.com/)

## Corrections and challenges

[Send a correction or an opposing source](https://escapemeridian.com/corrections/)

[JSON record](https://escapemeridian.com/readings/millidge-learning-beyond-demonstrations.json)