Broadening AI in Science: The report highlights a critical fact experts have been trying to get across: AI is not just for chatbots, but incredibly useful in “science, medicine, and agriculture”. The majority of the report is directed on capitalizing on this fact, by investing in AI research and adoption. 

AI as Great Power Competition: The national security section continues this positive narrative shift, focusing on clear measurement of AGI, domestic technology investments, securing energy for AI, and … space debris? 

A big question around AI and national security was whether policymakers would take a “war on terror” approach, focusing on surveilling and controlling AI, or a “great power competition” approach focused on industrial policy and competing with China. The report takes a step towards the latter.

Aside from a brief mention of CBRN in the agency coordination section, specific “doomsday scenarios” are not discussed. This indicates a further shift away from the “war on terror” narrative and a greater emphasis on competition with large adversarial states. 

Getting informed on short term concerns: Additional sections take an open-ended approach to addressing reasonable short-term concerns about elections, privacy, and copyright. This reflects a growing desire to understand AI at a more fine-grained level in many offices I meet with.

Once again, I cannot emphasize enough how much this has changed in a positive direction, in just a few months. 

This week’s podcast is with Keegan McBride, a lecturer at Oxford and an Alliance for the Future policy advisor. We discuss exactly that: how modernizing the government succeeded in Estonia and what it would take for it to do the same in the US.

Policy

We published a research summary on why AI is likely to follow historical patterns of scientific development. This refutes the idea that AI will autonomously make a discontinuous jump in ability. “Because returns diminish superexponentially, a slowdown in science occurs despite the theoretical ability for self-improvement ability to grow exponentially.”

The Abundance Institute, a technology-focused offshoot of the Center for Growth and Opportunity, launched this Monday. A completely different Inclusive Abundance Institute launched on the same day. The winds seem to be shifting in a more pro-tech direction.

Google released a massive 250-page AI Ethics paper. It is very dense with serious treatment of the broader societal and philosophical questions around AI. It is worth reading if you are curious about these philosophical questions, but may be difficult and unrelenting if you aren’t. 

Industry

Meta deployed and open-sourced 8B and 70B models of Llama 3. Try it here. They are also working on multimodal. 

In an interview with Dwarkesh Patel, Mark Zuckerberg disclosed that the yet to be released Llama 3 405B is a dense model. An oversimplified explanation of why it matters is that it indicates Meta rejects newer algorithmic approaches and focuses instead on increasing the scale of older algorithms.

Research

Chatbot Arena published a new benchmark using user interaction data. This is a good reason to introduce you to Arena, which is a platform for users to rate and compare LLMs. It uses an ELO ranking based on real user feedback, which is a better way of rating performance than fixed benchmarks in my opinion.

Alongside Llama 3, Meta also released a new cybersecurity eval. 

My interview with Dean W. Ball from George Mason University / Mercatus on competition with China and existing AI regulation. He argues that the historical approach to strategic competition, incentivizing domestic production with a light hand, continues to be the optimal strategy on AI.

Policy

The Open Philantropy funded Center for AI Policy released draft legislation calling for, among other things, unilateral emergency powers to seize AI hardware and software, as well as to arrest developers. Meanwhile, OpenPhil ally Paul Christiano was appointed to the NIST AI Safety Institute.

Matt Mittlesteadt on the new OMB AI Guidlines

“At a high level, these new restrictions represent a potential catastrophe for public sector AI diffusion. By going all in on AI safety rules the administration is failing to consider budgetary and administrative realities and, as a result, is risking slamming the breaks on federal AI all together.”

Dean Ball on the comments to the NTIA on open source/open weight AI. After analyzing all 332 comments using Google Gemini, he found that “the results are clear: the comments overwhelmingly support maintaining access to open models”.

Industry

Udio, or “ChatGPT for Music” launched. Its $10 million raise represents interest in a rapidly diversifying AI market. It joins competitor Suno in the AI-generated music market.

Stanford released a massive report on the leading players in the AI industry. Here are key charts on global competition, and competition between leading players. One example:

Research

An early paper on scaling laws, a key factor in predicting the growth of AI, has some newly identified weaknesses. Performance returns to investing in more AI hardware may be more uncertain than they seem. 

Google researchers published results on a state space model structure. This is important because it enables longer context windows, in other words the ability for their language models to process much longer documents and user requests at a time. 

Subscribe now