Sign up to save your podcasts
Or
Share “AI #115: The Evil Applications Division” by Zvi
Share to email
Share to Facebook
Share to X
Share to email
Share to Facebook
Share to X
Or
“AI #115: The Evil Applications Division” by Zvi
It can be bleak out there, but the candor is very helpful, and you occasionally get a win.
Zuckerberg is helpfully saying all his dystopian AI visions out loud. OpenAI offered us a better post-mortem on the GPT-4o sycophancy incident than I was expecting, although far from a complete explanation or learning of lessons, and the rollback still leaves plenty sycophancy in place.
The big news was the announcement by OpenAI that the nonprofit will retain nominal control, rather than the previous plan of having it be pushed aside. We need to remain vigilant, the fight is far from over, but this was excellent news.
Then OpenAI dropped another big piece of news, that board member and former head of Facebook's engagement loops and ad yields Fidji Simo would become their ‘uniquely qualified’ new CEO of Applications. I very much do not want her to take what she learned at [...]
---
Outline:
(01:42) Language Models Offer Mundane Utility
(05:18) Language Models Don't Offer Mundane Utility
(14:33) Take a Wild Geoguessr
(17:27) Write On
(19:04) Get My Agent On The Line
(21:09) We're In Deep Research
(23:55) Be The Best Like No One Ever Was
(24:56) Huh, Upgrades
(25:36) On Your Marks
(27:13) Choose Your Fighter
(31:27) Upgrade Your Fighter
(33:47) Unprompted Suggestions
(35:56) Deepfaketown and Botpocalypse Soon
(38:54) They Took Our Jobs
(52:25) The Art of the Jailbreak
(53:57) Get Involved
(54:25) OpenAI Creates Distinct Evil Applications Division
(01:01:58) In Other AI News
(01:03:46) Show Me the Money
(01:05:10) Quiet Speculations
(01:08:33) Overcoming Diffusion Arguments Is a Slow Process Without a Clear Threshold Effect
(01:15:49) Chipping Away
(01:17:59) The Quest for Sane Regulations
(01:21:02) Line in the Thinking Sand
(01:23:33) The Week in Audio
(01:25:25) Rhetorical Innovation
(01:28:10) A Good Conversation
(01:35:58) The Urgency of Interpretability
(01:47:37) The Way
(01:49:36) Aligning a Smarter Than Human Intelligence is Difficult
(01:50:02) People Are Worried About AI Killing Everyone
(01:54:29) Other People Are Not As Worried About AI Killing Everyone
(01:56:07) The Lighter Side
---
First published:
May 8th, 2025
Source:
https://www.lesswrong.com/posts/vCfHPnpCRtAPSEvnH/ai-115-the-evil-applications-division
---
Narrated by TYPE III AUDIO.
---
Images from the article:
Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.
View all episodes
By zviIt can be bleak out there, but the candor is very helpful, and you occasionally get a win.
Zuckerberg is helpfully saying all his dystopian AI visions out loud. OpenAI offered us a better post-mortem on the GPT-4o sycophancy incident than I was expecting, although far from a complete explanation or learning of lessons, and the rollback still leaves plenty sycophancy in place.
The big news was the announcement by OpenAI that the nonprofit will retain nominal control, rather than the previous plan of having it be pushed aside. We need to remain vigilant, the fight is far from over, but this was excellent news.
Then OpenAI dropped another big piece of news, that board member and former head of Facebook's engagement loops and ad yields Fidji Simo would become their ‘uniquely qualified’ new CEO of Applications. I very much do not want her to take what she learned at [...]
---
Outline:
(01:42) Language Models Offer Mundane Utility
(05:18) Language Models Don't Offer Mundane Utility
(14:33) Take a Wild Geoguessr
(17:27) Write On
(19:04) Get My Agent On The Line
(21:09) We're In Deep Research
(23:55) Be The Best Like No One Ever Was
(24:56) Huh, Upgrades
(25:36) On Your Marks
(27:13) Choose Your Fighter
(31:27) Upgrade Your Fighter
(33:47) Unprompted Suggestions
(35:56) Deepfaketown and Botpocalypse Soon
(38:54) They Took Our Jobs
(52:25) The Art of the Jailbreak
(53:57) Get Involved
(54:25) OpenAI Creates Distinct Evil Applications Division
(01:01:58) In Other AI News
(01:03:46) Show Me the Money
(01:05:10) Quiet Speculations
(01:08:33) Overcoming Diffusion Arguments Is a Slow Process Without a Clear Threshold Effect
(01:15:49) Chipping Away
(01:17:59) The Quest for Sane Regulations
(01:21:02) Line in the Thinking Sand
(01:23:33) The Week in Audio
(01:25:25) Rhetorical Innovation
(01:28:10) A Good Conversation
(01:35:58) The Urgency of Interpretability
(01:47:37) The Way
(01:49:36) Aligning a Smarter Than Human Intelligence is Difficult
(01:50:02) People Are Worried About AI Killing Everyone
(01:54:29) Other People Are Not As Worried About AI Killing Everyone
(01:56:07) The Lighter Side
---
First published:
May 8th, 2025
Source:
https://www.lesswrong.com/posts/vCfHPnpCRtAPSEvnH/ai-115-the-evil-applications-division
---
Narrated by TYPE III AUDIO.
---
Images from the article:
Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.