(2024-11-07) Zvim Ai89 Trump Card

Zvi Mowshowitz: AI #89: Trump Card.

Table of Contents

• Trump Card. What does Donald Trump’s victory mean for AI policy going forward?
• Language Models Offer Mundane Utility. Dump it all in the screen captures.
• Language Models Don’t Offer Mundane Utility. I can’t help you with that, Dave.
• Here Let Me Chatbot That For You. OpenAI offers SearchGPT.
• Deepfaketown and Botpocalypse Soon. Models persuade some Trump voters.
• Fun With Image Generation. Human image generation, that is.
• The Vulnerable World Hypothesis. Google AI finds a zero day exploit.
• They Took Our Jobs. The future of not having any real work to do.
• The Art of the Jailbreak. Having to break out of jail makes you more interesting.
• Get Involved. UK AISI seems to always be hiring.
• In Other AI News. xAI gets an API, others get various upgrades.
• Quiet Speculations. Does o1 mean the end of ‘AI equality’? For now I guess no.
• The Quest for Sane Regulations. Anthropic calls for action within 18 months.
• The Quest for Insane Regulations. Microsoft goes full a16z.
• A Model of Regulatory Competitiveness. Regulation doesn’t always hold you back.
• The Week in Audio. Eric Schmidt, Dane Vahey, Marc Andreessen.
• The Mask Comes Off. OpenAI in official talks, and Altman has thoughts.
• Open Weights Are Unsafe and Nothing Can Fix This. Chinese military using it?
• Open Weights Are Somewhat Behind Closed Weights. Will it stay at 15 months?
• Rhetorical Innovation. The Compendium lays out a dire vision of our situation.
• Aligning a Smarter Than Human Intelligence is Difficult. More resources needed.
• People Are Worried About AI Killing Everyone. Color from last week’s poll.
• The Lighter Side. Well, they could. But they won’t.

Trump Card

The ‘AI safety case for Trump’ that I’ve seen primarily seems to be that some people think we should be against it (as in, against safety), because it’s more important to stay ahead of China – a position Altman seems to be explicitly embracing, as well. If you think ‘I need the banana first before the other monkey gets it, why do you want to slow down to avoid poisoning the banana’ then that certainly is a take. It is not easy, you must do both.

He has promised on day 1 to revoke the Biden Executive Order, and presumably will also undo the associated Biden administration memo we recently analyzed

In principle he is clearly in favor of enabling American infrastructure and competitiveness here, he’s very much a ‘beat China’ guy, including strongly supporting more energy generation of various types, but he will likely lack attention to the problem and also technical state capacity.

There is also now a substantially greater risk of a fight over Taiwan, according to Metaculus

Language Models Offer Mundane Utility

Report from my AI coding experiences so far: Claude 3.5 was a huge multiplier on productivity, then Cursor (with Claude 3.5) was another huge multiplier, and I’m enjoying the benefits of several working features of my Chrome extension to assist my writing.

Roon: There is a sub culture of smart, emotionally well adjusted, but neuro atypical people who talk more to Claude than any human.

Compile data using screen capture analysis while browsing Gmail and feeding the video to Gemini? There’s something superficially bizarre and horrifying about that being the right play, but sure, why not? Simon Willison reports it works great.

Sully likes Claude Haiku 3.5 but notes that it’s in a weird spot after the price increase

This bifurcation makes sense. The cost per query is always tiny if you can buy compute, but the cost for all your queries can get out of hand quickly if you scale, and sometimes (e.g. Apple Intelligence) you can’t pay money for more compute.

But if you’re a human reading the outputs and have access to the cloud, of course you want the best.

Language Models Don’t Offer Mundane Utility