Skip to content
JevDirectory.org
Evaluation & Training

The Bitterest Lesson

TypeSafe's essay extending Sutton's lesson: doing the right task beats data, which beats compute, which beats algorithms.

Category
Evaluation & Training
Also known as
—
Related terms
3
Directory entries
1
Docs
typesafe.ai
Added
2026-09-24

Definition

The post orders the leverage in ML as doing the right task > data > compute > algorithms, using InstructGPT as its example: GPT-2-sized models over 100x smaller than GPT-3 beat it once trained on the right task.

It argues that scaling pre-training would have needed roughly GPT-7 level to beat that baseline and GPT-9 to beat InstructGPT on GPT-3, and that picking the right task often means leaving ML to study users, products, and organizations.

Tagstrainingresearch

From the directory

A TypeSafe blog essay arguing that in ML the order that matters is doing the right task, then data, then compute, then algorithms, using the InstructGPT result as its example.
Sites & GuidesArticle#official#article#models
Official

From the community

Posts from builders shipping with Jev right now.

Follow @typesafeai

Day-one notes from a skeptic, in Chinese

一些关于 Jev (@typesafeai) 的 notes 研究了一天 Jev,带来的新鲜感迅速回落,这好像就是一个更快的通用分类器/决策器,LLM 完全可以做到。 而且因为不了解模型背后的参数规模(应该不会很大),世界知识可能不一定有常规 LLM Show more

Diogo Almeida
Diogo Almeida
TypeSafe AI
@CompleteSkeptic

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x

Reply

Search and tagging on keep.md

A filter for reply-guy comments

A Downloads folder that sorts itself

Which outreach signals book demos

Website to App turns any URL into a native app

Introducing Jev for 'Website to App' Turn any website into a native mobile app. Just paste a URL. jev-1.13.0 decides how to build the original website as a *native* mobile app, then shipper submits to the app stores for you. We’ve been using this internally a ton for Show more

Diogo Almeida
Diogo Almeida
TypeSafe AI
@CompleteSkeptic

After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI? I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev • 20-200x faster • 40-400x

Reply