Andrew J Lohn

Hey all, big-compute and big-AI get a lot of airtime but it's also worth spending a minute to think about when AI is forced to be small and what that implies for reliability and so forth. @KyauMill21 @CSETGeorgetown cset.georgetown.edu/publication/on…

New AI & international security paper just dropped. 🔥 Let's take a look at how confidence-building measures can mitigate some of the international security risks associated with foundation models. arxiv.org/abs/2308.00862

New paper just out that I got to watch take shape. One encouraging point that stuck out to me was that there is still room to improve safety across many dimensions; looks like we don't usually have to trade off one type of safe for another yet. arxiv.org/pdf/2307.10586…

Today, the White House announced voluntary commitments about safety and security from seven major AI companies. John Bansemer and I recently wrote a short blog post about what that might look like. Might be worth a gander. cset.georgetown.edu/article/securi…

✨New Policy Brief✨ CSET's @AndrewJLohn & Krystal Jackson, in partnership with the @turinginst, investigate the future of cybersecurity: Autonomous Cyber Defense Agents. Learn more in their new report ⬇️ cset.georgetown.edu/publication/au…

If you're in the 'Compute for AI' world this is worth a quick look. I, for one, was surprised by some of the main results.

✨New Report✨ AI systems are being rapidly deployed, but many have vulnerabilities that are hard to study + mitigate. In our report produced w/ @StanfordCyber, 17 coauthors from policy research, law, industry + government propose steps to reduce risks: cset.georgetown.edu/publication/ad…

Who likes tons of data compiled into an easy-to-use and free interface? Want to do bibliometrics or understand the semiconductor supply chain or know which country is doing what in AI? Visit eto.tech

In a world where some hackers are making millions of dollars at a time, there’s also the question of who would turn them in at all. That said, it’s not unheard of for hackers to turn on each other, CSET's @AndrewJLohn told @washingtonpost's @timstarks. washingtonpost.com/politics/2022/…

✨New issue brief✨ As the United States faces a growing number of threats in the cyber domain, high school cybersecurity competitions can help to train the next generation of domestic cyber talent, write Kayla Goode, @ali_craw, and Christopher Back ⬇️: cset.georgetown.edu/publication/u-…

Wow! Way to go RAND. I can't think of a better person for this role than Jason (and I had been trying).

Some best practices to consider for deploying large language models from @OpenAI, @CohereAI, and @AI21Labs. Worth a look if you're sending them out into the world. txt.cohere.ai/best-practices…

CSET's @AndrewJLohn is testifying LIVE before the @SASCDems and @SASCGOP Cybersecurity Subcommittee where he will outline #AI’s capabilities and vulnerabilities within cyberspace defenses and offenses.

🎉New Policy Brief🎉 Managing vulnerabilities in machine learning is going to require a few changes. A new brief by @andrewjlohn and Wyatt Hoffman discusses a few of those along with what to expect for the future of cybersecurity. cset.georgetown.edu/publication/se…

I think our new one is a fun and short read. On its face it's obvious that such a steep exponential can't go forever but it's not so obvious when it might stop or why.

@Jason_Healey Point them to us at CSET too, especially the CyberAI Project. We'd love to have more practical skills/experience in our ranks. cset.georgetown.edu/job/research-f…

Any think tanks or academic cyber programs have spots for fellows, especially from the practitioner community? Our provost has been clamping down on these appointments and there's some amazing people looking for opportunities...

Big news for CSET and for the broader AI policy world!

@Miles_Brundage Not saying it's a common case, just saying it's convenient to reason about.

@Miles_Brundage Picture you have years of data on some daily/weekly/whatever thing that happens but then it gets wiped or you find out it's poisoned. Now you need X data points to train your model and after 24 hours or 1 week or whatever you have 1 data point. Tomorrow 2...

We usually envision AI conflict, crisis, and escalation as being fast but what counts as fast? Apples and Oranges but: State Dept figures Iran nuclear breakout time is 3 months. Wall clock to train a 1T-parameter language model is also 3 months.

@Miles_Brundage I think computing is a bad measure because it's not usually the limiting factor. But it's more predictable than most so it's somewhat better for setting bounds. My favorite is when data is necessarily generated slowly so that it is both the limiting factor and predictable.

@AndrewJLohn (and of course compute purchasing/manufacturing timelines! Lots of ways of slicing the issue, though I think compute is the common currency of many, or at least any that aim at high precision)