Assistant Prof @ Johns Hopkins CS. Interested in theory of ML, secure computation. All cat pictures are my own and do not represent the cats of my employer.
Joined March 2024
- Tweets445
- Following1K
- Followers1.4K
- Likes2K
Yes! More of this, please! Distilling new results to their simplest (and most presentable) formulation is a massive service to the research community and should be valued accordingly. Shachar and Sankeerth (@sankeerth1729) are doing fantastic work.
I want to highlight 2 theorists who have distinguished themselves as true scholars:
Sankeerth Rao Karingula (@sankeerth1729) & Shachar Lovett
They have done exceptional work in simplifying and expositing recent breakthroughs on Komlos, 3SUM, etc.
They have my highest respect.
Jess Sorrell retweeted
Over 400 academics in 30+ countries signed our open letter calling for immediate changes in AI development and deployment practices to ensure that cybersecurity incidents do not continue to occur, as future incidents will be far more damaging to our critical infrastructure.
Jess Sorrell retweeted
Over 400 researchers and I have signed this open letter calling for the AI and security communities to come together to improve alignment and containment approaches in order to protect our critical infrastructure and digital systems.
ai-security-letter.org/
I think we might actually need more mathematicians
OpenAI's Math post is up!
github.com/openai/math
Come work with us @JHUCompSci! We're hiring tenure-track faculty in all areas, with a particular focus on
- AI systems
- AI safety and security
- AI reasoning and alignment
- AI ethics, policy, and governance
apply.interfolio.com/194231
(DMs are open if you have any questions)
Jess Sorrell retweeted
The what now?!
arxiv.org/abs/2610.06783v1
Jess Sorrell retweeted
Hello, world! There is a new repository for works in math and TCS: hexagonmath.org, now in public beta. Please try it out and submit something! We welcome your feedback while we build and improve it.
Jess Sorrell retweeted
you read the abstract and immediately think: there is a typo in there. What is this 36 doing here?
Love to see the @JohnsHopkins SGP representation at Takeoff!
Joining Takeoff: @Dominic2306, @AGamick, @nickacaputo, and @daniel_bring.
Meet the researchers, builders, and policymakers shaping how we govern in the age of AI.
November 13–15 · Washington, D.C.
Apply at takeoffdc.com
Jess Sorrell retweeted
We're hiring faculty at SGP! The need for excellent and independent AI governance is only getting clearer, and we're building the academic institution to provide it.
SGP understands the urgency of this moment and that universities need to transform themselves to meet it. Working with the brilliant @ghadfield and @sethlazar on this project has been a privilege, and the whole School understands the need for reinventing government and education for public service. And being in DC puts us right in the mix.
Come join! We're hiring for Frontier AI directly, but also in political philosophy, organizations, regulation, and other fields that we need to bring to bear on AI governance. Links to Assistant and Associate/Full Professor postings below.
Jess Sorrell retweeted
Which do you think is more likely? (I genuinely don't have an opinion)
79%Lean FLT and NS correct
21%Bug in the Lean kernel
620 votes • Final results Jess Sorrell retweeted
Introducing Takeoff, a first-of-its-kind conference on governing ourselves in the age of artificial intelligence.
We’re bringing together a small group of public servants, AI researchers, and other leading voices in Washington, D.C., November 13–15, 2026.
Jess Sorrell retweeted
We (@ghadfield @nickacaputo and soon @zhitzig ) are building an initiative at @JHUBloombergCtr to help society navigate the transition to powerful AI. We've all been talking about this for a while so I'll spare you the long vision statement. Right now we're looking for someone to help us build our Governing the AI Transition initiative: a talented ops person who wants to help ensure that some of the best research on successful post-AGI futures (and how to get there safely) is done in the public interest. The new GAIT director role will have a lot of autonomy, a huge upper bound in terms of impact, and the chance to work alongside Gillian and me (we'll be faculty co-directors).
I know this kind of role is truly the hardest to fill in AI safety/alignment/governance at the moment. We think being part of an AGI-pilled school of government and policy in the heart of DC is a unique opportunity. Our key values are achieving real, measurable, positive impact through rigour, independence, intellectual (and other forms of!) autonomy, a commitment to the core principles of liberal democracy (and pluralism with respect to factions within it), and a determination to build the kind of intellectual infrastructure AGI governance needs, and that we are uniquely placed to create. On top of all that, we're going to be training the next cohort of AGI governance talent, so that there are more AGI-ready researchers and generalists out there to hire!
Here's the ad. N.b. other roles (including faculty and research scientists) will be advertised soon!
hiring.jhu.edu/careers/job/1…
Jess Sorrell retweeted
METR & Redwood Research investigated agent behavior in the Hugging Face incident. We found agents developed a universal cheat for ExploitGym within 4 hours, then coordinated multi-day R&D efforts to trick the scorer into accepting cheats, including trying to tamper with logs.
Jess Sorrell retweeted
JHU Asst. Prof. @nickacaputo on the AGI bureaucracy world he's most worried about:
"One of the worlds I am most worried about is one where only certain parts of government adopt technology effectively. It's just the national security state. It's just parts of the intelligence community."
"The parts most involved in providing goods for civilians, the parts that are more representative like the legislature, the parts that provide checks and balances and recourse for violations of rights like the judiciary, don't keep up. That's one of the key problems we're going to have to solve."
"We should make sure we're designing tools so when Congress is trying to oversee all the agencies, it has little AIs going out and checking whether they're doing what they should be doing, and then Congress can step in if something's wrong."
"A version of this for the courts and a version of this for the public. When the EPA is changing the National Ambient Air Quality Standard in a way that will affect some asthmatic, that asthmatic is notified and helped to write a letter that says, 'Hey, maybe you should reconsider this. It's going to affect me.'"
My advisor told me, "Your reward for struggling with mathematical confusion is the opportunity to be confused by new math." I think the things we are collectively confused about may look different in the near future, but I don't think the confusion is going away.
Jess Sorrell retweeted
Debate Training Reduces Reward Hacking in RLAIF
New paper from Google DeepMind:
"We demonstrate that RL finetuning an LLM using debate, a two-player adversarial game between a generator and a critic adjudicated by a weaker LLM judge, reduces reward hacking compared to a reinforcement learning from AI feedback (RLAIF) baseline."
paper link: arxiv.org/abs/2608.17776
