EP39: White House AI Executive Order, The Bletchley Declaration & Adversarial AI Attacks

Nov 2, 2023

The podcast discusses the White House's executive order on regulating AI, the Bletchley Declaration, adversarial AI attacks, future of AI computing, and leaked prompts powering ChatGPT's multi-tool mode. They also introduce hilarious new merch and discuss key insights on the future of language models.

Ask episode

AI Snips

Chapters

Transcript

Episode notes

INSIGHT

Risk Framing Often Overreaches

The executive order sometimes frames risks in exaggerated, low-probability terms (e.g., easy biological threats).
Some risks described are already possible without advanced AI and may be framed as AI-enabled for emphasis.

ADVICE

Adopt Risk-Based AI Use In Government

Embrace AI in government with risk-aware frameworks instead of blanket bans.
Infuse AI talent across agencies and provide safe data access to foster innovation.

ADVICE

Use AI To Harden Cybersecurity

Include AI-specific checks into cybersecurity frameworks and penetration testing.
Use AI to proactively find and patch vulnerabilities rather than only enabling attackers.

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

Get the app

Join our Discord: https://discord.gg/TRrgAyeM
Buy the merch: https://www.thisdayinaimerch.com/

This week the AI guys unpack the White House's sweeping executive order on regulating AI - will this lead to the death of open-source models? They also discuss the vague and fluffy Bletchley Declaration signed by world leaders, why Geoffrey Hinton just won't stop fearmongering, and introduce some hilarious new merch including a life-size shower curtain! Tune in for hot takes on the AI ethics debate, prompt engineering tricks, and key insights on the future of language models.

CHAPTERS:
=====
00:00 - King Charles on AI (Cold Open)
00:20 - Thoughts on White House AI Executive Order
23:09 - The Bletchley Declaration & AI Safety Summit
38:04 - LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2 & They Killed Tay!
48:34 - Adversarial Attacks and Defenses in Large Language Models: Old and New Threats Paper
51:51 - Mike proposes What The Future of AI Computing Might Look Like
55:00 - Leaked: The Secret Prompt Powering ChatGPT's New Multi-Tool Mode (and How to Hack It)

1:01:39 - Anthropic Have Raised More Billions & Our Merch Store!

SOURCES:
======
https://www.whitehouse.gov/briefing-room/presidential-actions/2023/10/30/executive-order-on-the-safe-secure-and-trustworthy-development-and-use-of-artificial-intelligence/
https://www.aisnakeoil.com/p/what-the-executive-order-means-for
https://www.gov.uk/government/publications/ai-safety-summit-2023-the-bletchley-declaration/the-bletchley-declaration-by-countries-attending-the-ai-safety-summit-1-2-november-2023
https://www.businessinsider.com/sam-altman-and-demis-hassabis-just-want-to-control-ai-2023-10
https://twitter.com/ylecun/status/1718263147591573949?s=20
https://twitter.com/ldjconfirmed/status/1718456393026490523
Leaked Prompt: https://raw.githubusercontent.com/spdustin/ChatGPT-AutoExpert/main/_system-prompts/all_tools.md
https://www.cnbc.com/2023/10/27/google-commits-to-invest-2-billion-in-openai-competitor-anthropic.html

PAPERS:
======
https://arxiv.org/pdf/2310.17688.pdf
https://arxiv.org/pdf/2310.20624.pdf
https://arxiv.org/pdf/2310.19737v1.pdf