Welcome! I’m Jason Benn, and I like to build big harmonious systems with code and people. This page describes my interests in systems engineering, AI-assisted reasoning, cooperative AI, community building, and a smattering of miscellaneous non-professional interests.

🐦 Twitter | 🐙 Github | 🔗 LinkedIn | 📫 Email

ML systems engineering

ML model training pipelines. I was the sole Founding MLE at Sourceress (YC S17), an AI for recruiting startup, which peaked at $2M ARR. I architected an ML experimentation, training, and deployment platform that enabled the comparison of various model architectures across 400+ proprietary datasets — basically, an internal Kaggle.

Distributed systems. I spent 2 years on distributed A/V systems at Minerva. Imagine building Zoom in 2015 for classrooms distributed across the world in places with poor Internet, like Bangalore and Buenos Aires, with tons of synced state (like whiteboards). These mixed availability and consistency requirements also had to tolerate dozens of network partitions per day. Some of the solutions we implemented are STUN/TURN servers, bundling A/V streams, optimistic updates,

AI evals. I’m currently collaborating with Prof. Daniel Kang on a test of RL generalization. I also helped his team implement CVE-Bench, the only cybersecurity benchmark that actually reimplements exploits from the real world. I also spent a year at Generally Intelligent, the successor to Sourceress and predecessor of Imbue (now valued at $1B), helping an AI step through infant cognitive development milestones.

MLE Club. I organize a club for senior ML engineers to explore and hack. We're kind of like Recurse Center for machine learning, or South Park Commons for engineers instead of founders.

image.png

AI-assisted reasoning

How can AI help me do my best thinking? What would it look like to have my aspirational self (or a groupchat of positive influences) as a thinking partner? How can AI help me complete my hierarchical reviews and goal-setting? Can AIs help us forecast the future more accurately and make better policymaking decisions with futarchy? What are practical “thinking algorithms”? How can I filter, integrate, and remember the content I consume into published artifacts and durable behavior change? I would love to collaborate on projects along these lines!

billion-dollar-pile.jpg

<aside> ➡️ List of projects related to AI-assisted reasoning

</aside>

AI coordination

Can AIs help us coordinate with each other (and with AIs) more effectively? Even if we succeed in aligning AI, it won’t matter unless we can solve long-standing global coordination problems.

Meal matchmaking. By adapting the Gale-Shapley algorithm to a variant of the Stable Roommates problem, and generating a pairwise preferences matrix with GPT-4, I was able to introduce conference attendees to the 18 people they most wanted to meet. At the end of an event, one attendee told me that it “would be hard to improve” with tears in his eyes…!

image.png

<aside>

➡️ Meal matchmaking algo & results

</aside>

Experts collectively snapshotting the frontier of a field. As an experiment in collective intelligence, I organized a conference and asked everyone to ELO scoring important, neglected, and tractable ideas, and starting opt-in email threads of folks that like each idea.

image.png

<aside>

➡️ Video overview of snapshotting tool

</aside>

A Marauder’s Map for housemates. I took inspiration from scripts released at the DEFCON security conference to triangulate the positions of housemates’ devices by pinging Bluetooth MAC addresses from an IoT network. In Emily’s words: "The best Big Brother app ever! I check it every day"

image.png

<aside>

➡️ Flock: evolution and takeaways

</aside>

AI CRM. Can we collectively source events? Yes: by chains of warm introductions to people that “you want to be more like in some way”. I built a pipeline to automate the research and outreach process.

image.png

<aside>

➡️ Video overview of research pipeline

</aside>

Explorer’s Clubs. A Google Sheets script and event format in which everyone writes down an inarticulate hunch, rates other hunches, and is then matched with each other in 3 iterated rounds to flesh out said hunches.

image.png

<aside>

➡️ Explorer’s Club

</aside>

Community building

The Neighborhood. Is it possible to reproduce the vibes of a university campus, except multigenerational, by kickstarting a bunch of community spaces in a single square mile? I founded a 501c3, was funded for 3 years by Schmidt Futures, ran 4 conferences and started 6 community spaces, and convinced 6 other organizations and ~200 people to move into a one-square mile area in San Francisco. The idea was to reproduce the vibes of a university campus, but multigenerational.

<aside> ➡️ My quarterly updates through the entire Neighborhood journey

</aside>

<aside>

**➡️ My working notes**, summarizing much of what I learned about unconferences, community, real estate, and housing

</aside>

Real estate. It’s key to my cohousing dream so I got a real estate license (but if you want to buy you should talk to actual professionals like John and Lisa) and I spent the summer of 2024 developing a novel real estate investing strategy for cohousing. Real estate will still be scarce post-singularity, after all.

<aside>

➡️ Looking to buy a house? I recommend John & Lisa

</aside>

The Archive. I co-organized the Archive coliving house. Living around people that inspire me destroyed my limiting beliefs even as we all grew more connected and inspired the Neighborhood. Several of us even went on live national TV to get interviewed by Megyn Kelly — 3 weeks before she got canceled. I like to think that it’s because she asked us “so coliving means orgies, right?” No Megyn, that is not right.