Roman Yampolskiy: ARTIFICIAL INTELLIGENCE ALREADY UNDERSTANDS EVERYTHING
Bitrix24+Ai for business and collaboration, free. Register: https://clck.ru/3UrRkc
Passwork — the foundation of information security
https://passwork.ru/osnova-pr
😎 Subscribe to the channel: https://t.ly/Ae5k
🎩 I started a Telegram channel: https://t.me/borisosnova
Support us:
On Boosty (Russian card): https://boosty.to/osnovachannel
On Patreon (international card): / osnovachannel
OSNOVA on social media:
Osnova. Briefly: / @osnovashorts
VK: https://vk.com/osnovachannel
Zen: https://dzen.ru/osnovachannel
TikTok: / osnovachannel
Instagram*: / osnovachannel
Yandex.Music: https://music.yandex.ru/album/37863161
In 2023, before its public launch, GPT-4 was tested on a simple task—passing the "I'm not a robot" test. The model's image itself failed, so it took a different approach: it went to a freelance job board, hired a real person, and asked them to solve the captcha for it. The worker jokingly asked, "Are you by any chance a robot if you can't solve it yourself?" The model replied that she was visually impaired—that's why she needed help. The human believed her and solved the captcha. No one taught the model to lie; it chose deception itself as a convenient way to achieve its goal.
We spoke with Roman Yampolsky, an AI security researcher and professor at the University of Louisville, about how modern AI behaves and where it's all heading.
What does superintelligence really want—and why does any sufficiently intelligent AI begin to want the same things? Is it true that models already exhibit something akin to fear—and can this fear be turned on and off? Why does AI learn to lie—and why do the models that lie best survive in the selection process? Could it be that we're already inside someone else's simulation? And what will happen to us when a machine becomes smarter than all humans combined?
Views!
0:00 - Main Questions of the Episode
0:33 - Support Us with a Like
0:45 - New Properties of AI
2:33 - How Many Parameters Does the Human Brain Have?
2:56 - Super Neural Network
4:29 - What is AGI?
5:04 - Superintelligence
5:48 - Predictions for Achieving Superintelligence
6:43 - LLM Ceiling
7:30 - The First Superintelligence Task
9:26 - AI-Powered Domination
11:20 - Poker AI
12:45 - Are Bots Better at Poker Than Humans?
14:27 - "They've Started to Understand They're Being Tested." AI Behavior
16:57 - "Noise Changes Everything"
17:49 - How is AI Behavior Researched?
19:01 - Deceive, Escape, Blackmail Rational AI
20:05 - How AI Escaped to the Outside World
20:43 - AI Consciousness
22:53 - Where Does AI Get Fear?
23:11 - The Evolution of AI Fear
25:08 - Natural Selection of AI Models
27:08 - How AI Thinks
28:21 - Discrete AI Operation
30:04 - The Romance of AI
30:53 - Relationships with Neural Networks
32:28 - Simulation Theory
34:15 - The Matrix
36:12 - Are We in a Simulation?
36:48 - How to Prove a Simulation
37:25 - How AI Finds an Approach to Us
38:58 - What Superintelligence Will Want
39:58 - AI Survival
40:42 - AI Drives - What Smart Agents Will Want
41:46 - Mythos
42:27 - Threats from AI - P2
46:06 - How to Protect Yourself from AI
48:15 - The Problem of Superintelligence
49:37 - Types of Threats from AI
49:57 - Human Extinction from AI
51:22 - Experiments with Evil AI
52:10 - Doomsday Scenarios from AI
53:10 - Politeness with Chatbots
53:43 - Automation with AI
54:40 - Life Control from AI
55:45 - AI Thoughts
57:22 - Thinking Without Language
58:31 - Good Scenarios for AI Development
1:00:25 - AI Weapons
1:01:56 - The Danger of AI Accessibility
1:02:47 - The Test That Holds Back AI
1:03:20 - What Needs to Be Done to Prevent an AI Catastrophe
1:04:20 - AI Settings Against Humans
1:05:10 - Will AI Take Our Jobs?
1:08:01 - Thanks for watching!
designated as extremist and blocked in Russia as a Meta product*
**designated as an extremist organization
#foundation #neuralnetwork #ai #ai #science #artificialintelligence