Mystery AI Hype Theater 3000

Episode 36: About That 'Dangerous Capabilities' Fanfiction (feat. Ali Alkhatib), June 24 2024

Jul 19, 2024

Computer scientist Ali Alkhatib discusses Google DeepMind's flawed study on dangerous capabilities of large language models, emphasizing the social implications of AI. Critiques on poorly developed AI models, misinterpretations in AI research, hacker tools extracting data, and the limitations of tech news solutions are also explored. The conversation challenges hierarchical thinking in AI benchmarks and deceptive rhetoric in promoting harmful ideas.

01:02:00

Episode guests

Ali Alkhatib

AI Summary

Highlights

AI Chapters

Episode notes

Podcast summary created with Snipd AI

Quick takeaways

Big tech companies using preprint servers to avoid peer review, influence public AI conversation.

Anthropic's Cloud 3.5 Sonnet benchmarks criticized for hierarchical implications.

AI technology used to moderate emotional expressions in customer service sparks ethical debate.

Deep dives

Cloud 3.5 Sonnet Benchmarks Segmented by Educational Stages Spark Controversy

In a post from anthropic on LinkedIn, the Cloud 3.5 Sonnet benchmarks are revealed to have segments like 'graduate-level reasoning,' 'undergraduate-level knowledge,' and 'grade school math,' drawing criticism for the hierarchical implications embedded in the model names.

Overlapping Error Bars and Confidence Intervals

00:24

Decentralized Decision-Making in Job Creation

01:08

Intro

2min

Analyzing a Paper Artifact and Evaluating AI System's Dangerous Capabilities

18min

Critiquing AI Research and Misinterpretations

25min

Criticism of Study on Expert Forecasts and Exploration of Swift Center

6min

Discussion on a Hacker Tool Extracting Data from Windows New Recall AI

3min

Discussion on Tech News and AI Solutions

3min

Challenging Singular Solutions and Deceptive Rhetoric in AI Models

5min

When is a research paper not a research paper? When a big tech company uses a preprint server as a means to dodge peer review -- in this case, of their wild speculations on the 'dangerous capabilities' of large language models. Ali Alkhatib joins Emily to explain why a recent Google DeepMind document about the hunt for evidence that LLMs might intentionally deceive us was bad science, and yet is still influencing the public conversation about AI.

Ali Alkhatib is a computer scientist and former director of the University of San Francisco’s Center for Applied Data Ethics. His research focuses on human-computer interaction, and why our technological problems are really social – and why we should apply social science lenses to data work, algorithmic justice, and even the errors and reality distortions inherent in AI models.

References:

Google DeepMind paper-like object: Evaluating Frontier Models for Dangerous Capabilities

Fresh AI Hell:

Hacker tool extracts all the data collected by Windows' 'Recall' AI

In NYC, ShotSpotter calls are 87 percent false alarms

"AI" system to make callers sound less angry to call center workers

Anthropic's Claude Sonnet 3.5 evaluated for "graduate level reasoning"

OpenAI's Mira Murati says "AI" will have 'PhD-level' intelligence

OpenAI's Mira Murati also says AI will take some creative jobs, maybe they shouldn't have been there to start out with

You can check out future livestreams at https://twitch.tv/DAIR_Institute.

Subscribe to our newsletter via Buttondown.