The Week in Examples #5 [23 September]
Responsible scaling, economic impacts, and safety groups in UK parliament
Welcome back to The Week In Examples, a roundup of the week’s AI policy, ethics, governance, and industry news. There were a lot of big stories I wanted to cover this week, but, as always, I’m only sharing thoughts on the three most interesting pieces I’ve seen. The rest of the good stuff is in the links below.
As always, please tell me what works and what doesn’t, or just drop me a line to say hello at hp464@cam.ac.uk. A few of you gave some great feedback on the last roundup, so thanks very much to everyone who shared some thoughts with me.
Three things
1. Anthropic shares responsible scaling policy
What happened? Anthropic announced an AI Safety Levels framework designed to help mitigate the possibility of catastrophic risks. The approach defines different types of AI development using incremental levels and corresponding security procedures and restrictions. The framework, which is “loosely” based on the US government’s biosafety level (BSL) standards for handling of dangerous biological materials, introduces four different tranches. These include ASL-1 (systems with no risk like a 2018 LLM), ASL-2 (systems that show “early signs” of dangerous capabilities like Claude 2), ASL-3 (systems that substantially increase the risk of catastrophic misuse), and ASL-4 (though not defined, is “likely” to involve systems providing significant escalations in catastrophic misuse potential and autonomy).
What’s interesting? The company said it had designed the ASL system to “strike a balance” between targeting catastrophic risk and incentivising beneficial applications and progress on AI safety. The underlying idea is that while the ASL system requires Anthropic to pause training of powerful models if scaling outstrips its ability to comply with safety procedures, it does so in a way that “directly incentivises” the group to solve the necessary safety issues to unlock further scaling. This seems to me to represent a well-known idea about the need to work on powerful models in order to develop appropriate safeguards for complex systems. It’s slightly counterintuitive, but its essence is that building safe models is dependent on creating models with the potential to be dangerous. The trick here is to make sure that the developers’ ability to create more powerful systems does not outrun lessons learned from the safe development of their predecessors.
What else? This is important work and I’m glad that Anthropic is undertaking it. It’s worth saying, though, that because the group has yet to define what ASL-4 or at ASL-5 will look like, it will be difficult to determine how effective such a system will be in practice over the long run. At ASL-3, containment measures include hardened security, evaluating for ASL-4 warnings, and compartmentalising training techniques. Given the speculative nature of the capabilities associated with ASL-4 systems, however, no commensurate measures are provided. Whether or not such systems prove to be safe will ultimately be determined by the calibre of those safeguards, rather than the strength of the broader framework in which they exist.
2. Oxford researchers weigh in on economic impact of AI
What happened? Researchers from the University of Oxford argue that while ‘generative AI’ (a term I dislike) has widened the scope of automation, it will also make many jobs easier to do for people with lower skills. Carl-Benedikt Frey and Michael A Osborne argue that (1) remote jobs are more likely to be automated, while AI will increase the value of in-person communication skills; (2) hallucinations will continue to be a problem, which they anticipate means that firms will mostly keep a human in the loop; (3) AI is less likely to be deployed in what they describe as “high-stakes” contexts; and (4) creative jobs are less prone to automation, but creative professionals might face more competition and lower wages.
What’s interesting? Ten years ago, Frey and Osborne famously predicted that 47% of US jobs were at risk of automation. According to the pair’s estimates, almost half of total American employment was “at risk of computerisation.” The researchers went on to argue that wages and educational attainment exhibit a strong negative relationship with an occupation's probability of computerisation. In the 2023 study, however, they said that while generative AI has increased the scope of automation further, it will also make many jobs easier to do for people with lower skills.
What else? This dynamic was at play in recent research by Ethan Mollick of the University of Pennsylvania. It found that consultants using AI finished 12% more tasks on average, completed tasks 25% faster, and produced 40% higher quality results than those that did not use AI. A study from Brynjolfsson, Li, and Raymond (2023) found that “AI assistance disproportionately increases the performance of less skilled and less experienced workers across all productivity measures” while Peng et al. (2023) found that “results suggest that less experienced programmers benefit more from Copilot.” As the writer Noah Smith explains, “AI gives a much bigger boost to low performers than to high performers.”
3. AI safety groups pitch policymakers
What happened? Conjecture, the London-based AI safety group, spoke to the UK’s House of Lords about extinction risk from AGI. CEO Connor Leahy said “every time we do the next bigger AI run, we don’t know what’s going to pop out on the other side. Someday, it’s going to be Russian Roulette, and if you ever find yourself playing Russian Roulette, I have a recommendation: Stop.” He outlined three recommendations to the House of Lords: (1) strict liability for damage caused by AI systems that apply not just to the user, but also to the developer; (2) a cap on computing power set at 10^24 FLOP, or roughly the size of ChatGPT; and (3) a global ‘kill switch’ whereby governments build the infrastructure to be able to shut down powerful AI systems deployments and training runs if needed.
What’s interesting? Moves to apply political pressure are increasing in tandem with efforts to build public support for major regulatory interventions. Vox reported on the results of another poll from the AI Policy Institute by concluding that Americans simply “don’t want” AGI, while Axios introduced a “scorecard” from the same group arguing that all currently proposed legislation is insufficient as judged by public opinion. And a post from Holly Elmore on the Effective Altruism forum made the case for AI safety advocacy to the public. It’s been a big week for AI public engagement.
What else? There’s a lot to say here, starting with Conjecture’s comments to the House of Lords. While I’m not going to get into how we ought to apportion liability across the value chain or the appropriate ceiling for a hypothetical compute cap, I will say that I was a bit confused by the proposal for a ‘global kill switch’. In practical terms, this would require (1) either unilateral action on behalf of one state to prevent another from executing a large training run; or (2) global coordination to find, assess, and secure multilateral agreement in order to prevent a training run. I’ll reserve full judgement until I see a fleshed out plan (which I’ve heard is on the way) but I think this one has some challenges.
Best of the rest
Friday 22 September
AI developing too fast for regulators to keep up, says Oliver Dowden (The Guardian)
The EU AI Act Newsletter #36 (Substack)
Top Fed official sees "cautious optimism" in economic impact of AI (Axios)
Governments race to regulate AI tools (Reuters)
Opinion: The Copyright Office is making a mistake on AI-generated art (Ars Technica)
Thursday 21 September
OpenAI teases DALLE-3 (OpenAI)
How the U.N. Plans to Shape the Future of AI (TIME)
Microsoft hopes people won’t become ‘over-reliant’ on its AI assistant (The Verge)
Anthropic’s Dario Amodei on AI’s limits: ‘I’m not sure there are any’ (TechCrunch)
Top GOP senator teams up with key Dem on ‘light-touch’ AI bill (Politico)
Wednesday 20 September
AI’s $200B Question (Sequoia)
Midjourney vs. Ideogram, ML product companies, preventing AI winter, DALL·E 3 tease (Interconnects, Substack)
Will Hurd Releases A.I. Plan, a First in the Republican Presidential Field (The New York Times)
Exclusive: First AI plan in Republican primary (Axios)
How Google taught AI to doubt itself (The Verge)
Tuesday 19 September
OpenAI Red Teaming Network (OpenAI)
Opinion: America’s framework to lead on AI (The Hill)
Britain invites China to its global AI summit (Reuters)
Biden plans to work with world leaders to ensure AI’s use as a tool of ‘opportunity’ (CNBC)
DeepMind’s New AI Can Predict Genetic Diseases (Wired)
Monday 18 September
Notes on Existential Risk from Artificial Superintelligence (Michael Nielsen)
Proposed principles to guide competitive AI markets and protect consumers (UK Government)
Course: Understand How AI Impacts You and Your Government (aPolitical)
OpenAI Hustles to Beat Google to Launch ‘Multimodal’ LLM (The Information)
AI Startup Writer Raises $100 Million to Pen Corporate Content (Bloomberg)
One for the road
