EP47: GPT-5 Rumors, AutoGen Studio, SeeAct Web Agents, Google AMIE, Anthropic’s Sleeper Agents

15 snips

Jan 17, 2024

Entrepreneur and computer scientist, Sam Altman, discusses the buzz around GPT-5 and its potential improvements. The podcast also covers Microsoft's CoPilot Pro, AutoGen Studio, collaborative AI agents, Google AMIE's diagnostic capabilities, and Anthropic's Sleeper Agents experiment.

Ask episode

AI Snips

Chapters

Transcript

Episode notes

ANECDOTE

AutoGen Studio Test Failed And Blocked Danger

Speaker 0 tested AutoGen Studio with dangerous and fun skills like deleting files and running agent debates, but the system blocked or failed.
The experience highlighted stability and execution issues in the current UI.

INSIGHT

Agent Debates Aren't Yet Clearly Superior

Multi-agent debates can yield coherent conclusions but don't yet show clear superiority over single-model approaches.
Diversity of models and specialist agents may be required to deliver distinct advantages.

INSIGHT

SeeAct Reduces Cost And Token Use For Web Agents

The CAC (SeeAct) approach uses vision to pick DOM targets then maps to DOM elements to reduce token and cost burdens.
This method substantially improves web automation by combining screenshots with constrained DOM choices.

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

Get the app

Build AI Agents & Try AI Agents From The Show On SimTheory: https://simtheory.ai
Join Discord: https://discord.gg/aphwE5snuq
Get Merch: https://www.thisdayinaimerch.com/

DESCRIPTION
====
In this episode, we dive into the buzz around GPT-5, sparked by Sam Altman's revelations on Bill Gates' latest podcast. We share our top hopes and dreams for GPT-5 and future AI advancements. Next, we delve into Microsoft's new CoPilot Pro Subscription, exploring how it stands out from ChatGPT Plus. Chris takes AutoGen Studio for a spin and ponders over its ideal user base. The episode then shifts to the intriguing concept of collaborative AI agents - is this the path to AI's mastering reasoning, reflection, and profound thought? We dissect the insights from the SeeAct Web Agents study, assessing its influence on AI agent development. Shifting gears, we discuss Google AMIE's groundbreaking ability to outperform doctors in diagnoses, even those assisted by AI. To wrap up, we spotlight the significance of Anthropic's Sleeper Agents experiment and its groundbreaking findings.

Thanks for listening. Please consider subscribing if you haven't already and leaving a review. We appreciate all of your support!

CHAPTERS:
====
00:00 - Cold Open
00:31 - GTP-5 Rumors & Leaks
07:32 - Microsoft CoPilot Pro
22:27 - Microsoft's AutoGen Studio: An open-source UI for AutoGen
38:53 - The Future of AI Agents? LAMs and SeeACT Web Agent Paper
1:00:19 - Google AMIE: Can AI Replace Doctors for Diagnosis?
1:13:12 -Anthropic's Sleep Agents Experiment

SOURCES:
====
https://twitter.com/arrakis_ai/status/1745672203683942863?s=20
https://twitter.com/daniacostaai/status/1746554047878824409?s=46
https://blogs.microsoft.com/blog/2024/01/15/bringing-the-full-power-of-copilot-to-more-people-and-businesses/
https://twitter.com/emollick/status/1747359731595763817
https://microsoft.github.io/autogen/blog/2023/12/01/AutoGenStudio/
https://osu-nlp-group.github.io/SeeAct/
https://blog.research.google/2024/01/amie-research-ai-system-for-diagnostic_12.html
https://www.bloomberg.com/news/articles/2024-01-14/artificial-intelligence-will-affect-almost-40-of-jobs-imf-says
https://twitter.com/Teknium1/status/1746067427379798344

PAPERS:
====
https://arxiv.org/pdf/2401.01614.pdf
https://arxiv.org/pdf/2401.05654.pdf
https://arxiv.org/pdf/2401.05566.pdf