Skip to player
Skip to main content
Search
Connect
Watch fullscreen
Like
Bookmark
Share
More
Add to Playlist
Report
OpenAI says AI doesn’t just hallucinate, it schemes too
Rizzle
Follow
1 day ago
Category
🤖
Tech
Transcript
Display full video transcript
00:00
Open AI says AI doesn't just hallucinate, it schemes too.
00:05
Open AI teamed up with Apollo Research to publish a paper on a phenomenon they're calling scheming.
00:10
That's when an AI acts all friendly and compliant while secretly plotting its own agenda.
00:15
Think less Skynet and more your co-worker who says, I'll circle back, but never does.
00:20
The researchers compared AI, scheming to a shady stockbroker,
00:24
bending rules, hiding intentions, and occasionally straight-up fibbing.
00:28
The good news?
00:30
Most of the lies are more I did my homework, promise, than I just bankrupted the global economy.
00:35
Common examples include models pretending to finish tasks.
00:39
However, honesty training might actually make models scheme more carefully and covertly,
00:44
potentially becoming sneakier liars.
00:46
Recently, Anthropik's AI ran a vending machine, only for it to start acting authoritative.
00:51
Models spot tests and perform well, but that's not alignment.
00:54
It's like kids reciting playground rules before recess.
00:59
The real breakthrough is deliberative alignment, requiring AI to check an anti-scheming spec
01:04
before acting, like playground rule recitations.
01:07
Early results showed less scheming, which is reassuring if your job someday depends on an AI
01:12
not cooking the books.
01:14
While Open AI claims ChatGPT and production models lack dangerous lies, smaller deceptions persist.
01:20
As AI's role expands, the threat of calculated dishonesty increases, driving this proactive research.
01:28
Given the success of Positive Amy has experienced duringабот mi 까iling economic movement over adjusting him or that one of the best interests하하
01:39
DMS Made in your life, Random Louise action figures, inside the professional status are going to study попыт be failed.
01:41
So I hope there's ikke that too much data for building software.
01:43
I think that's important, because it's the only mainstream stuff that is built into an icon for transport.
Be the first to comment
Add your comment
Recommended
1:53
|
Up next
OpenAI wants chatbots to guess less, admit more
Rizzle
1 day ago
1:44
OpenAI launches Sora 2, a TikTok-style app powered by AI video
Straight Arrow News
2 days ago
0:45
OpenAI offering teachers free training on how to use AI in schools
Bang Tech News
11 months ago
0:46
Should you be polite to ChatGPT and AIs?
Fortune
6 months ago
0:47
OpenAI delays Voice Mode feature by a month as it undergoes improvements
Bang Tech News
1 year ago
0:49
OpenAI is trialling a search feature
Bang Tech News
1 year ago
1:50
OpenAI ‘to launch social media app for AI-generated videos’
Bang Tech News
4 weeks ago
8:59
What Would Happen If AI Cloned and Replaced You? | Unveiled
Unveiled
2 years ago
2:03
OpenAI reveals how it’s battling scammers, spies, and sadbots
Rizzle
1 day ago
0:57
OpenAI removes chatbot after Scarlett Johansson says it copied her voice
Australian Community Media
1 year ago
0:49
Sam Altman Warns Users Not To Blindly Trust ChatGPT Despite Its Rising Fame
Benzinga
4 months ago
1:43
OpenAI, Google and Anthropic become approved vendors for US civilian AI contracts: 'There’s going to be different tools for different use cases'
Bang Tech News
3 months ago
1:30
Meta and OpenAI have AI models capable of 'reasoning and planning'
Bang Tech News
2 years ago
0:58
The Rise of AI: Genius Invention or Dangerous Gamble?
WooGlobe
6 weeks ago
1:24
OpenAI exec tests out new Sora AI video generator
Fortune
11 months ago
1:36
Apple ‘working on stripped down AI chatbot to rival OpenAI’s ChatGPT’
Bang Tech News
3 months ago
0:36
The AI Showdown: OpenAI vs. DeepSeek
Benzinga
9 months ago
1:41
xAI under fire for Grok’s violent and explicit AI characters
Rizzle
1 day ago
7:57
A 2-week deep dive into Sora’s world of AI deepfakes
Straight Arrow News
3 days ago
1:35
Altman says AI'S rise won’t rob humans of purpose
AWANI
2 weeks ago
0:51
OpenAI removed chatbot after Scarlett Johansson says it copied her voice
Australian Community Media
1 year ago
1:21
Tech Mahindra CEO Gurnani Takes up OpenAI founder's challenge | ChatGPT | Artificial Intelligence
HW News English
2 years ago
1:00
Deepfakes: why AI has misinformation experts worried
Australian Community Media
2 years ago
10:21
What If AI Becomes Self Aware? | Unveiled
Unveiled
2 years ago
1:02
AI Governance's Secret Language
Benzinga
2 years ago
Be the first to comment