Will AI Make Humans Extinct? Did GPT-4 Get Worse? & OpenAI's GPT-4 Roadmap Revealed | E17

26 snips

Jun 2, 2023

Ask episode

AI Snips

Chapters

Transcript

Episode notes

INSIGHT

Extinction Framing Moves AI Into Politics

The Center for AI Safety's extinction warning pushes AI risk into mainstream politics and media.
Hosts debate whether the rhetoric raises helpful awareness or serves self-interested control by big AI labs.

INSIGHT

Motivations Behind Calls For Regulation

Some signatories calling for regulation may be self-preserving rather than purely altruistic.
Hosts question why companies wouldn't slow or stop development themselves if they truly feared extinction.

INSIGHT

Hardware Bottlenecks Buy Time

Short-term AI benefits (healthcare, productivity) contrast with 5–10 year worries about AGI reducing human relevance.
Hardware limits and efficiency gaps give time to address risks before explosive self-improvement.

Get the Snipd Podcast app to discover more snips from this episode

Get the app

This week we cover the Center for AI Safety's Statement and ask, "Will AI Really Make Humans Extinct?" We cover the discussion on GPT-4 Deteriorating: is it really getting worse? And discuss some likely causes.

We also use the Internet Archive to discuss OpenAI's GPT-4 roadmap including a stateful API, giant context sizes, plugins having no product market fit and why OpenAI is constrained by GPUs.

We also cover the lawyer who got caught using ChatGPT and the problems with hallucinations. Will step-by-step rewards help solve these problems? Do we need better warnings for stupid lawyers?

If you like this episode please consider subscribing, comment and liking to help others find our podcast. We appreciate your support.

----
Chapters:
----
00:00 - Is GPT-4 Getting Worse?
00:20 - Center for AI Safety Statement: Will AI Make Humans Extinct?
12:09 - Is GPT-4 Getting Worse? Is Alignment the Problem?
29:27 - GPT-4 Roadmap: 1M Context? Multi-modal?
36:11 - Vision of ChatGPT: Have Plugins Failed?
38:02 - Will OpenAI API's Have Too Much Competition?
40:21 - Giant AI Models Aren't Over: Is GPT-4 Much More Capable?
44:39 - ChatGPT for Lawyers: What Could Possibly go Wrong?
48:17 - The Hallucination Problem
53:22 - Process Supervision: Solution to Hallucination?
55:59 - Japan Removes Copyright Laws for AI Progress
57:53 - UAE's Falcon 40B Open Source & Royalty Free!
1:01:04 - Will We Value Written Language?

----
Sources:
----
https://www.safe.ai/statement-on-ai-risk
https://time.com/6283609/artificial-intelligence-race-existential-threat/
https://www.nytimes.com/2023/05/30/technology/ai-threat-warning.html
https://twitter.com/ylecun/status/1663616081582252032?s=46&t=uXHUN4Glah4CaV-g2czc6Q
https://twitter.com/chrmanning/status/1663920862439915524?s=20
https://news.ycombinator.com/item?id=36134249
https://twitter.com/OfficialLoganK/status/1663934947931897857
https://www.reddit.com/r/MachineLearning/comments/13tqvdn/uncensored_models_finetuned_without_artificial/?utm_source=share&utm_medium=ios_app&utm_name=iossmf
https://web.archive.org/web/20230531203946/https://humanloop.com/blog/openai-plans
https://www.nytimes.com/2023/05/27/nyregion/avianca-airline-lawsuit-chatgpt.html
https://simonwillison.net/2023/May/27/lawyer-chatgpt/
https://twitter.com/drjimfan/status/1663972818160332800?s=46&t=uXHUN4Glah4CaV-g2czc6Q
https://technomancers.ai/japan-goes-all-in-copyright-doesnt-apply-to-ai-training/
https://twitter.com/emollick/status/1663960017697898497/photo/2
https://www.tii.ae/news/uaes-falcon-40b-now-royalty-free