142,996 views
3,026 likes
Thank you Blue Dot for sponsoring this video: https://bluedot.org/hoog
This is an interview with the AI whistleblower: Jacob Coxon who worked for Anthropic and OpenAI.
Chapters:
0:00 The AI whistleblower
1:29 METR: AI progress has roughly doubled
2:04 Getting the interview
3:02 Jesse, Timaeus and Resolution
7:21 Why 2x means researchers move 10x faster
9:03 What's happening inside the AI labs
10:59 Math: from toddler to Fields medalist in five years
13:25 Beyond math: science and the real-world bottleneck
16:16 The capabilities to worry about: bio and strategy
18:19 Hacking: every piece of software is vulnerable
24:39 Attacks on critical infrastructure
26:36 How Anthropic caught up
28:38 Recursive self-improvement: AI doing AI research
31:00 Scaling and general intelligence
32:14 The METR report: 2x for the company, 5–10x per researcher
35:36 The bottleneck: a few dozen researchers
37:42 Scaling laws and why the trend could break
41:13 How fast could it get? Six years in six months
42:23 Why Jacob left Anthropic
44:16 "It's all marketing hype"
46:03 Is it real intelligence or a stochastic parrot?
48:31 Could AI hit a ceiling?
50:03 Fuzzy tasks, taste and creativity
53:30 The elicitation overhang
56:44 Why forecasters keep being surprised
58:26 "This is just sci-fi"
1:00:15 "Nothing ever happens"
1:02:12 Inside vs outside the labs
1:04:19 How would we even know? Redactions and whistleblowing
1:07:05 Could it already be too late?
1:08:03 Losing control: from tools to employees
1:11:46 Humans with AI vs rogue AI
1:13:04 Why models are becoming agents
1:14:17 Gradual disempowerment
1:15:56 Should we just hand over the reins?
1:16:48 Models know when they're being tested
1:18:24 What is the alignment problem?
1:18:52 Aligned to what?
1:20:22 Why alignment is hard
1:22:46 How would AI actually kill us?
1:25:15 Anthropic's safety plan
1:28:08 "Make the mistake, then fix it"
1:30:27 Safety vs capabilities: the trade-off
1:31:34 Why the labs keep racing
1:33:11 Does the "race to the top" work?
1:34:29 Racing to the brink without breaking rank
1:35:56 Why don't more people leave?
1:37:18 Why the labs don't trust each other
1:37:58 The IPO and the money
1:40:49 Dario's call for a slowdown
1:42:12 Why Dario stopped talking about extinction
1:44:26 Why Jacob joined Anthropic
1:44:59 Why he left when others stayed
1:48:20 Breaking the race: a collective action problem
1:50:22 The ideal scenario
1:52:08 Science vs advocacy
1:54:09 What you can do
1:55:15 Closing thoughts
1:58:28 Sponsor: BlueDot Impact
Made by:
https://x.com/jesse_hoogland
https://x.com/hoogyoutube