Bill Anderson-Samways

Research Analyst at the Institute for AI Policy and Strategy (IAPS)

[Institute for AI Policy and Strategy (IAPS) @iapsAI]

🚀 The Trump Administration unveiled its comprehensive AI Action Plan on Wednesday. Experts at the Institute for AI Policy and Strategy reviewed the plan with an eye toward its national security implications.

Read our analysis:
https://t.co/EZfEPjvaVR

[Institute for AI Policy and Strategy (IAPS) @iapsAI]

VIRTUAL EVENT | July 24 | 2:30pm ET

Join IAPS for a virtual panel on the Trump Administration's AI Action Plan featuring @MarkBeall of the AI Policy Network, Tanya Das of @BPC_Bipartisan, @jgeltzer of @WilmerHale, & IAPS's Jenny Marron.

🔗 Register: https://t.co/j4hWgsYfkT

[Oscar Delaney @oscar__delaney]

NEW @iapsAI REPORT: Will the US government lead a project to build the first advanced AI system? @BillSamways and I gathered forecasters & experts to address this critical question for AI governance. Bottom line: participants predicted 34%, with high uncertainty. 🧵/1

[Bill Anderson-Samways @BillSamways]

@Jess_Riedel @robertwiblin (For context, I'm one of the report's authors :)) That would be my guess based on participants' qualitative comments, though we didn't ask participants to make quantitative estimates conditioning on longer development timelines.

[Rob Wiblin @robertwiblin]

The US Government’s Role in Advanced AI Development: Predictions and Scenarios: https://t.co/BC3WSQmxUD

[Rob Wiblin @robertwiblin]

Will the US government develop AGI itself?

Six forecasters and five domain experts on average gave it a ~1 in 3 chance.

(If AGI comes soon and occurs in the USA.) 1/

[Centre for British Progress @BritishProgress]

The UK has a growing AI testing + assurance industry. With some changes to regulation and a small amount of public investment, this industry could be worth £18bn to the UK economy.

New paper with @SMFthinktank @iapsAI from @BillSamways @jkraprayoon

https://t.co/A80DL28jD5

[David Lawrence @dc_lawrence]

The @UKDayOne proposals are coming in thick and fast! Our latest sets out the case for AI assurance market shaping. This could be a great opportunity for @peterkyle to shoot for growth + safety at the same time.

Thanks to @SMFthinktank @iapsAI @BillSamways @JKraprayoon

[Jam Kraprayoon @JKraprayoon]

Out now: how the new UK govt can kickstart a world-leading AI assurance industry, supercharging AI adoption and growth.

This memo from @BillSamways and I outlines a market-shaping programme around AI assurance tech for the UK govt [1/6]

https://t.co/xRGiLsqBKa

[Bill Anderson-Samways @BillSamways]

Finally, a disclaimer: our piece intends to suggest case-studies that researchers or policymakers could examine further. Our findings should not be interpreted as endorsements of specific regulatory approaches, nor as definitive rankings of the case-studies identified. [4/4]

[Bill Anderson-Samways @BillSamways]

Which case-studies can inform the regulation of advanced AI?

New paper from myself, @oscar__delaney, Ashwin Acharya and @zoehtwilliams undertakes a first-of-its-kind systematic search for relevant regulatory precedents.

https://t.co/wS1IfEc2Xs

Summary in the thread👇[1/4]

[Bill Anderson-Samways @BillSamways]

We identified regulatory case-studies using quantitative measures of five variables that seem relevant to AI regulation: (1) intensiveness; (2) expertise; (3) enforcement against powerful companies; (4) use of risk-assessments; and (5) focus on uncertain phenomena. [3/4]

[Bill Anderson-Samways @BillSamways]

Finally, Anthropic’s RSP is a commendable first step towards managing AI risks. It performs well on many UK govt recommendations, such as planning to pause model development if adequate risk-mitigation measures are not in place. Other companies should adopt such measures. [6/6]

[Bill Anderson-Samways @BillSamways]

Our third rec is that Anthropic and other AI companies should detail thresholds for alerting government authorities of identified risks - again, ideally based on the SR tolerances previously mentioned. [4/6]

[Bill Anderson-Samways @BillSamways]

Fourthly, Anthropic and other companies should commit to external scrutiny of both their evaluations results and their evaluations methods at a predefined risk-threshold, equivalent to Anthropic’s ASL-3 threshold. [5/6]

[Bill Anderson-Samways @BillSamways]

Secondly, Anthropic and other AI companies should specify thresholds in their RSPs for a granular subset of risks – for example, not just “misuse” but “biological misuse” as opposed to “cyber misuse”. These thresholds should be pegged to SR tolerances, as suggested above. [3/6]

[Bill Anderson-Samways @BillSamways]

How does Anthropic’s Responsible Scaling Policy compare to UK govt guidance?

New brief from me and @iapsAI colleagues, @shaunkeee, @__J0E___, @mariedbuhl & @zoehtwilliams.

https://t.co/xlTLFAWPpG

Recs for AI companies & UK/US govts in thread - esp. on risk thresholds 👇 [1/6]

[Bill Anderson-Samways @BillSamways]

Our top rec: Anthropic, other companies, and governments should set societal risk (SR) tolerances for RSPs, covering multiple-fatality events. SR tolerances for catastrophic events (≥1,000 fatalities) in other domains range from 1 E-04 to 1 E-10 such events per year. [2/6]

[Bill Anderson-Samways @BillSamways]

FSAP shows that “checklist”-style approaches are good at addressing known threats with clear control measures. Effective AI regulations could therefore combine checklist-style approaches for known threats with risk-based measures addressing less certain threats. [5/5]

[Bill Anderson-Samways @BillSamways]

What lessons does the core US biosecurity policy hold for AI regulation?

Among other things, myself and Ashwin Acharya find that the Federal Select Agent Program (FSAP) offers a precedent for an R&D-phase AI licensing regime.

https://t.co/C0WsV1LTqX

Summary in thread. [1/5]

[Bill Anderson-Samways @BillSamways]

FSAP covers a list of known dangerous pathogens. That misses less certain risks from enhanced potential pandemic pathogens. A more "risk-based” approach would be better here. AI regulations should also have a risk-based component, given the uncertainty around many AI risks. [4/5]