I really believe that we’re on the cusp of creating these new companions, so meaningful, lasting friendships
– Mustafa Suleyman, the chief executive of Microsoft AI, told Joanna Stern
More Human than Human? A 14 year old boy did not just kill himself after having an addictive, all-consuming, relationship with a chatbot on Character.AI’s website (“they were in love”). Or at least, that is what one would think after reading Mustafa Suleyman’s latest comments. The rapid commercialization of AI is steering product development towards intimate, human-like relationships with users, raising ethical and societal concerns. When Joanna Stern went on a weekend getaway with several AI voice assistants, GPT addressed her with her nickname all by itself, underscoring the intentional eroding of emotional boundaries in human-machine relationships, eventually leaving users vulnerable to manipulation or emotional dependency.
Hot off the Press. The European Union has released a draft of it’s first Code of (Good) Practise for General-Purpose AI. And guess what? It’s really good and worth reading. It covers aspects including model disclosures, risk classification, technical risks, and governance measures. We think there’s further to go, though, and one of the goals of our project is to continue to understand and promulgate best practices.
Hello Common Corpus. Want to build a foundation model without violating copyright? It finally has arrived: the first fully-documented and cleaned training dataset that doesn't rely on copyrighted or web-scraped content. Common Corpus provides 2 trillion tokens of cleaned and permissibly licensed data to train LLMs with (technically the pre-training stage). This is mostly in English and French. But is this dataset large enough? GPT4 was apparently trained on 5 - 6 trillion unique tokens, while Common Corpus ‘only’ has 0.808 trillion of cleaned English language tokens? You tell us.
A deep dive into power and data center capacity in China. SemiAnalysis explores how Chinese companies are circumventing U.S. sanctions in the semiconductor industry. Dylan Patel, Jeff Koch and Sravan Kundojjala detail Huawei’s extensive fab network and the role of strategic acquisitions and partnerships in building a self-sufficient semiconductor supply chain. The report also discusses the role of wafer fabrication equipment (WFE) suppliers and their lobbying efforts, as well as the implications for future export controls. The (ultra)-detailed analysis shows just how difficult it is for the U.S. to enforce chip sanctions and the need for more new regulatory frameworks to address this.
Resilience, Opportunity and Risk in Taiwan. Chris Schroeder reflects on his recent visit to Taiwan. One striking quote: “We met with experts on how AI is changing the dynamics of fake news at a scale and inability to detect that make both Taiwan and Ukraine at the front lines of this innovation. One tech executive told us that ‘Cyber security IS national security.’” There’s much more.
AI agents as Abundant as Sheep? In New Zealand there are approximately 4.3 sheep per person (down from 22 in 1982). In the near future, the ratio of AI agents to humans may surpass New Zealand like sheep-to-human ratios. In Fortune, Gillian Hadfield of the University of Toronto discusses the potential economic disruptions posed by an army of autonomous AI agents operating in a legal and regulatory void. Hadfield calls for the “registration and identification of AI agents” — to companies or even to their own legal personalities – lest “millions of law-free AI agents casually [wreak] economic havoc.”
LLMs Look Increasingly Like General Reasoners. In a recent LessWrong blog post, the author revisits their earlier skepticism about large language models (LLMs) being able to perform general reasoning tasks. They highlight advancements in LLM capabilities, particularly on things like blocksworld, planning and scheduling, and the ARC-AGI benchmark. These developments suggest that LLMs are increasingly exhibiting general reasoning abilities, according to the author. At the same time, The Information reports that there are widespread concerns that the rate of AI improvements may be slowing. We’d love to know who you think is right.
Terms of Service, Terms of Exploitation? Just coming across this excellent piece by Kevin Klyman at The Stanford Center for Research on Foundation Models (CRFM) breaking down the acceptable use policies (AUPs) of the major foundation models, along with other private tools they use to try to restrict model usage. The article emphasizes the importance of transparency and enforcement in AUPs to ensure responsible AI usage. The mention of licensing inevitably conjures up memories of the Windows era (*shudder*), but it serves as a reminder that, at their core, foundation models are still software — even if the ways in which the software can be accessed, licensed and deployed might be more complicated than in the past. As Kevin Klyman shows us, considering AI regulation through a software lens might yield more practical oversight mechanisms. Thoughts? Comments? Let’s discuss. (And speaking of terms of service, US regulators plan to investigate Microsoft’s cloud business for punitive licensing terms.)
Ain’t No Business Like a Digital Advertising Business. Perplexity AI, an AI-powered search engine, finally has a U.S. date by which it will display query results ads – first as “recommended” follow-up questions – alongside the query responses, with a commitment to share a portion of the ad revenue with publishers whose content is cited in the answers. The company emphasizes that ads will be relevant to user queries, enhancing the overall search experience.
Regulate First, Ask Questions Later? Martin Casado, General Partner at Andreessen Horowitz (a16z), critiques current AI regulatory efforts, arguing that they often target hypothetical future scenarios rather than addressing present-day risks. We agree. Casado emphasizes the importance of first understanding the actual dangers introduced by AI technologies to develop effective regulations. Casado also highlights the need for AI regulation to be informed by existing regulatory frameworks.
Tiny Automated Focus Groups. Microsoft released a new library called TinyTroupe designed to simulate interactions between groups of people using LLM-powered multi-agent personas. It allows one to generate many tiny AI people that take on unique personas and have them interact or ignore each other in a “tiny world.” While they note it is still experimental, they propose that it could be used for anything from evaluating ads to testing search engines. We have already starting experimenting with it and will report back soon.
Using Every Trick in the Book. In the ongoing quest to replicate OpenAI’s o1 style reasoning, Nous Research announced an upcoming API to make their existing Hermes 70B model (or any model) more powerful by supplementing it with a code interpreter and advanced reasoning capabilities. This uses techniques such as 'chain of code', having multiple LLM models collaborate, and Monte Carlo Tree Search to improve output quality. While this new API shows some improvements over other models, it’s still in beta and there doesn't seem to be plans to open source it. (Let’s hope it’s not another “Claude in a trench coat”!)
For more about our project: https://www.ssrc.org/programs/ai-disclosures-project/