Intel / podcast

#452 – Dario Amodei: Anthropic CEO on Claude, AGI & the Future of AI & Humanity

Source: lexfridman.com
Author: Lex Fridman
Date: Published Nov 11, 2024

Context

Making an AI less verbose can have unwanted effects on its coding answers. In an interview with Lex Fridman, Anthropic CEO Dario Amodei explains that crudely penalizing long responses can encourage a model to omit code and leave the user to finish it. A change intended to improve conversation can therefore damage a different use.

Amodei presents this as a broader difficulty in steering model behavior: changing training to reduce one annoyance can alter other responses in ways developers did not intend. His example explains why an update needs across different tasks. Brevity is useful only when the answer still contains the work the user requested.

Terms in this piece · Glossary

eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.

Key quotes

“We are rapidly running out of truly convincing blockers, truly compelling reasons why this will not happen in the next few years.”
Dario Amodei

“I've seen the movie enough times, I've seen the story happen for enough times to really believe that probably the scaling is going to continue”
Dario Amodei

“The thing you need to understand about these models is they just want to learn. The models just want to learn.”
Ilya Sutskever

“I think containing bad models is a much worse solution than having good models.”
Dario Amodei

“And so I think the worst enemy of those who want real accountability is badly designed regulation. We need to actually get it right.”
Dario Amodei

Chapters

00:00Introduction
10:19Scaling laws
19:25Limits of LLM scaling
27:51Competition with OpenAI, Google, xAI, Meta
33:14Claude
36:50Opus 3.5
41:36Sonnet 3.5
44:56Claude 4.0
49:07Criticism of Claude
1:01:54AI Safety Levels
1:12:42ASL-3 and ASL-4
1:16:46Computer use
1:26:41Government regulation of AI
1:45:30Hiring a great team
1:54:19Post-training
1:59:45Constitutional AI
2:05:11Machines of Loving Grace
2:24:17AGI timeline
2:36:52Programming
2:43:52Meaning of life
2:49:58Amanda Askell - Philosophy
2:52:26Programming advice for non-technical people
2:56:15Talking to Claude
3:12:47Prompt engineering
3:21:21Post-training
3:26:00Constitutional AI
3:30:53System prompts
3:37:00Is Claude getting dumber?
3:49:02Character training
3:50:01Nature of truth
3:54:38Optimal rate of failure
4:01:49AI consciousness
4:16:20AGI
4:24:58Chris Olah - Mechanistic Interpretability
4:29:49Features, Circuits, Universality
4:47:23Superposition
4:58:22Monosemanticity
5:05:14Scaling Monosemanticity
5:14:02Macroscopic behavior of neural networks
5:18:56Beauty of neural networks

Read the source lexfridman.com