The Week in Examples #3 [2 September]
Calls for compute governance, international oversight of civilian AI, and research focused on AI and deception
Hello and welcome to the third edition of The Week In Examples. Read on for an end of week roundup of different news from industry, government and civil society as well as opinion and analysis of all things AI. Once again, we have my thoughts on the most interesting three pieces I’ve seen this week, links to other AI resources that I’ve stumbled on, and something fun to finish.
Regular public service announcement: I want The Week In Examples to be a useful resource, so I’ll need your help to overcome the cold start problem and prevent it from freezing to death. Please don’t hesitate to write to me at hp464@cam.ac.uk with feedback to help me make it better.
Three things
1. Inflection AI boss calls for restrictions on chip use
What happened? Inflection AI co-founder Mustafa Suleyman called for the US to mandate that customers who buy NVIDIA chips sign up to the White House AI commitments. In an interview with the Financial Times, Suleyman said: “The US should mandate that any consumer of Nvidia chips signs up to at least the voluntary commitments — and more likely, more than that.” According to the article, Suleyman made the case that action was needed because the “exponential trajectory” of AI meant that in two years large models would be 100 times more powerful than OpenAI’s GPT-4.
What’s interesting? While Suleyman didn’t go into detail about other possible measures, the idea that the US government ought to mandate that certain conditions be attached to the sale of NVIDIA’s chips is an interesting one. Compute governance is one of the most effective levers out there for preventing the proliferation of powerful models, but what such moves would look like as well as how (and by whom) such a regime would be enforced remain open questions. In this framing, the recently announced White House commitments help to clarify part of the question (namely, the ‘what’) but enforcement would likely have to be tied to a compute monitoring regime (more on that here).
What else? Suleyman also claimed that “Too much of the conversation is fixated on superintelligence, which is a huge distraction…we should be focused on the practical near-term capabilities which are going to arise in the next 10 years [and] which I believe are reasonably predictable.” I have a harder time understanding this idea. Personally, I don’t really think much attention is given to superintelligence (as distinct from the broader discourse around AGI or frontier models). I’m also not sure how to square the notion that 100x more powerful systems are coming very soon with the idea that superintelligence is so far away as to be irrelevant. My other question here is about what exactly discussions of superintelligence might distract from, when it seems to me that speculated increases in future capability is one the primary driving forces behind the development of today’s governance measures.
2. A standards-based approach for international oversight of AI
What happened? Researchers from the University of Oxford released a paper proposing an governance body to certify compliance with international standards on civilian AI. The group examined a “jurisdictional certification” approach modelled on other international bodies such as the International Civilian Aviation Organization (ICAO) and International Maritime Organization (IMO) to propose the introduction of an International AI Organization (IAIO). Such a body would certify regulatory regimes (and not labs) for compliance with international standards. Divergences from said standards would, in this model, result in consequences such as exclusion from valuable trade relationships.
What’s interesting? In practice, the goal of the IAIO would be to set standards and work with national bodies to monitor states for effective AI regulation. When oversight is deemed to be deficient, states could prohibit the import of AI products from those jurisdictions and limit the export of materials and technologies needed to build AI. The model represents a middle ground between centralised international monitoring and decentralised oversight, charting a path in which standards are set internationally but enforcement is carried out by local regulators.
What else? The paper proposes standards for the development and deployment of frontier models (primarily based on a combination of evals and transparency measures); AI labs (a licensing regime focused on system security, access, and documenting past compliance with development and deployment regulations); and data centres (mandating accountings of chip purchases, cybersecurity measures, and tracking and reporting when customers are training frontier models). In general, I suspect that standards are underpriced as tools for the global governance of AI—so I like the central role they play in this paper. Standards-based approaches would, however, be contingent on highly competent, muscular domestic regulators to enforce them.
3. Understanding deceptive AI proves to be a tricky business
What happened? The Center for AI Safety released a paper considering the potential for AI systems to deceive humans. The research defines deception as “the systematic production of false beliefs in others as a means to accomplish some outcome other than the truth.” Importantly, according to the authors, this definition doesn't necessarily imply that AIs have beliefs or intentions (an idea the authors tested by reviewing the behaviour of Meta’s Diplomacy-playing CICERO system). The group analyses a dozen AI systems that have successfully learned how to deceive other agents, and introduces a taxonomy of deceptive practices for large models.
What’s interesting? What I like about this paper is the acknowledgement that deception can occur when a person directly interacts with a system, and when a system is used by one party to influence another. Based on this idea, the authors propose that individuals with malicious intent can exploit AI systems with deceptive capabilities to engage in fraud, interfere with elections, or produce propaganda. Such manipulative AI, they argue, can disseminate falsehoods across society or create a misleading impression that the AI is functioning properly. The group also makes the case that sophisticated AI models may employ deception to elude human oversight (possibly by tricking developers) and that when a business or regulatory body assesses an AI's actions, the system could intentionally display appropriate behaviour just to clear the evaluation.
What else? Defining deception is hard, but so is relating it to other important concepts like persuasion and manipulation. I would argue that persuasion the process by which one party influences another through a process of rational deliberation; manipulation is the process by which one party influences another by bypassing a person’s capacity for rational deliberation; and that deception is an instance when one who does not believe something communicates to another with the intention that someone else shall be led to believe it. Manipulation and deception are both deemed to be ethically troublesome because they bypass a person’s ability for rational deliberation (manipulation) or because they can be used to encourage a person to act against their own interests (deception). The two are not, however, always mutually exclusive: you can manipulate without engaging in deception (e.g. by changing the context in which someone makes a decision), but you can also manipulate using deception (e.g. by deliberately sharing inaccurate information).
Best of the rest
Friday 1 September
AI Policy Perspectives August 2023: Google DeepMind (Substack)
OpenAI Is the Undisputed King of AI. These Five Forces Could Take It Down Anyway: The Algorithmic Bridge (Substack)
Pelosi Says AI ‘Double-Edged Sword’ Needs Regulation: Bloomberg
Ads for AI sex workers are flooding Instagram and TikTok: NBC News
AI Hallucinations could usher in the next era of AI: cognitive AI: Fast Company (Opinion)
How the AI Revolution Will Reshape the World: Time Magazine (Opinion)
Thursday 31 August
US curbs AI chip exports from Nvidia and AMD to some Middle East countries: Reuters
China has no place at the AI Safety Summit: Air Street Capital (Blog)
How worried should you be about AI disrupting elections? The Economist
UK publishers urge Sunak to protect works ingested by AI model: The Guardian
The Inventor Behind a Rush of AI Copyright Suits Is Trying to Show His Bot Is Sentient: WIRED
OpenAI’s Moonshot: Solving the AI Alignment Problem: IEEE Spectrum
Teaching with AI: OpenAI (Blog)
Wednesday 30 August
ChatGPT Would Vote Democrat, New Study Finds—But It's Full of Flaws: The Algorithmic Bridge (Subdstack)
Large language models aren’t people. Let’s stop testing them as if they were: MIT Tech Review
Meet Aleph Alpha, Europe's Answer to OpenAI: WIRED
Chinese ChatGPT alternatives just got approved for the general public: MIT Tech Review
ChatGPT-maker OpenAI accused of string of data protection breaches in GDPR complaint filed by privacy researcher: TechCrunch
Britain must become a leader in AI regulation, say MPs: The Guardian
Schumer's AI meeting will include top labor and civil rights advocates: Washington Post
Tuesday 29 August
AI Safety #20: AI Safety Newsletter (Substack)
Language models surprised us: Planned Obsolescence
Google DeepMind has launched a watermarking tool for AI-generated images: MIT Tech Review
It Costs Just $400 to Build an AI Disinformation Machine: WIRED
Google to Add AI Models from Meta, Anthropic to Its Cloud Platform: Bloomberg
UK cybersecurity agency warns of chatbot 'prompt injection' attacks: The Guardian
OpenAI Nears $1 Billion of Annual Sales as ChatGPT Takes Off: Bloomberg
US Copyright Office wants to hear what people think about AI and copyright: The Verge
Monday 28 August
OpenAI launches a ChatGPT plan for enterprise customers – TechCrunch
Behind the AI boom, an army of overseas workers in ‘digital sweatshops’ – The Washington Post
A.I. Brings the Robot Wingman to Aerial Combat – The New York Times
Opinion: We don’t have to reinvent the wheel to regulate AI responsibly – TechCrunch
AI is biased. The White House is working with hackers to try to fix that – NPR
Vignette of the week
Incredible work! Keep it going!