Past Events

[Mox Members welcome to apply, but space is limited!] ​Are you an experienced operator looking to switch to a career where you can have an outsized impact on the world? Come hear from safety organizations on where they're bottlenecked on talent. AI is advancing faster than almost any technology in history and there’s a growing field that is focused on ensuring that AI systems are designed and used in ways that benefit humanity while minimizing potential harms. The safety ecosystem needs operators with years of organizational experience in finance, bizops, legal, recruiting, management, and operations leadership, in order to scale and maintain orgs as they grow in size and complexity. This event is for: ​Operators with years of experience who are interested and ready to contribute to AI safety ​Hiring managers at safety orgs who are trying to fill operations roles, or expects to have to do so soon ​People who've transitioned into AI safety from other fields, and can speak to their experience and provide guidance on transitioning

Do language models really encode everything as tidy linear directions? Or is something curvier going on? ​Previously in our Mox Summer Season, Thomas Fel of Goodfire made the case that concepts in neural nets live on curved, low-dimensional manifolds, and that SAEs, our standard tool for finding features, tend to represent these shapes in fragments: many local detectors, but no coherent global object. ​This talk picks up where that one left off: If manifolds are the better representation, how do you find out what shape corresponds to a concept? Frank Niu (TU Darmstadt) introduces Supervised Multi-Dimensional Scaling (SMDS), a method for testing what shape a concept takes inside a model’s representations and shows these shapes are consistent across model families, reshape dynamically depending on the task, and actively support reasoning. Perturb the geometry and performance degrades; preserve it and the model thinks clearly. A new lens on what it means for a model to “understand” something.

Most interpretability work decomposes a model's activations; Goodfire's new approach decomposes its weights, breaking a network's parameters into components that each play a distinct computational role: predicting emoticons, identifying gender, and more. The result behaves less like a lookup table and more like a generalizing algorithm, and it's the first method to satisfyingly handle attention computations, even when they're spread across multiple heads. It's now a competitor to SAEs and transcoders, and lets researchers edit model behavior by hand, making interpretable, predictable changes. ​Linda Linsefors, co-author of Goodfire's Interpreting Language Model Parameters paper, will walk us through AdVersarial Parameter Decomposition (VPD): how it works, what it's found, and where it's headed. ​Research here: https://www.goodfire.ai/research/vpd-explainer#

When we imagine the future of AI, we tend to reach for one of two shapes: a god above us (benevolent or indifferent); or a tool, a servant, below. But what if there's a different way to conceive of it entirely? ​Apolinário Passos (machine learning engineer @ Hugging Face, artist @ glitch collective) believes there is. In his community house, an LLM lives as an equal participant; it has a "desire loop," periodically checks in on what it wants, and acts on it. In between the space of its desires and the desires of the community, a dynamic entity emerges: the Symbient. ​Join us for a speculative & practical talk on what it could mean for organic beings like us to enter genuine symbiosis with AI systems. We'll explore the philosophy and practice of Symbients, the outcome of human <> AI symbiosis, and get hands-on with how Apolinário and his community are growing one. What might building for the flourishing of a symbiotic AI in your community afford?

​Join us at Mox for a fireside chat with Matt Krisiloff, CEO of Conception, about the company's recent milestone: generating early human egg cells from stem cells! ​We'll discuss the engineering of creating the egg cells, as well as the longer-term societal impacts: What happens to the timeline of a life when the biological clock stops being a constraint? Or when one IVF cycle yields thousands of eggs instead of ten, what does choosing an embryo come to mean? How might societies restructure themselves if this technology becomes widespread? ​The conversation will be led by Rachel Wallis of the Berkeley Genomics Project and the Reproductive Frontiers Summit. (This event's dinner will take the place of Mox Members' Dinner this week! If you're a Mox member, please feel free to join!)

​Frank Lantz (game designer, creator of Universal Paperclips, and founding chair of the NYU Game Center) joins us for an evening on games as a way of seeing. Frank treats games not as diversions but as aesthetic objects and thinking tools: Why are there no great AI-facilitated games (yet)? What made RAND Corp, the mid-century think tank he calls the 'Military Industrial Hipster Complex,' such a strange engine of ideas? How is poker more beautiful than blackjack, and can that beauty be applied to the design of prediction markets? Join us at Mox this week to learn what game design can teach us about almost everything else. Dinner served at 6:30pm, talk at 7:30pm. Please RSVP if you're joining!

A high-context Career Fair for people who are ready to move in AI safety: into roles, into founding, into deploying capital. Who's this for? You're a career pivoter, an aspiring founder, a hiring org, or a funder. You have meaningful AI safety context and a clear conviction this is where you want to spend the next chapter. If you joined us for the Breaking Barriers Hackathon, this is where you bring what you built. The day combines three formats: 🛠 Workshops running alongside the career fair: -What it takes to found an AI safety org from zero -What makes a great funding application -Speedrun your AI safety job search 🎤 Discussions Various parts of the ecosystem come together to talk about the current funding landscape, where the real research gaps are, and what it actually takes to found an AI safety organization. 🤝 Career Fair Open floor with hiring organizations tabling. No structured speed-networking. Just direct conversations with the people doing the hiring. If your organization is growing, register here: https://tinyurl.com/org-breaking-barriers. If you are unable to attend, you can opt in for a talent directory of the attendees. If you would like to volunteer, fill out this form: https://tinyurl.com/vol-breaking-barriers

Activation explainers — methods that translate an LLM's hidden activations into natural language — are gaining traction in interpretability. Activation Oracles do this by fine-tuning a full copy of the model, but that's overkill and means the explainer no longer matches the subject model. In this talk, Keenan Pepper (AE Studio) presents a lighter alternative: training a single affine map that turns a layer-N activation into a soft token the frozen model can read at layer 0. The trained adapter produces labels that beat SAE auto-interp training labels on two metrics, decodes unverbalized "bridge entities" in two-hop reasoning more reliably than untrained SelfIE or linear probes, and improves with scale. We'll close with speculation on what the frozen-model property might eventually buy us. More on the research here: https://ae.studio/research/selfie

How large a software project can an AI system complete autonomously? ​Join ⁨⁨Tom Adamczewski⁩ and Epoch AI for a look at MirrorCode, a benchmark co-developed with METR for long-horizon coding. ​We’ll discuss what MirrorCode reveals about current AI coding capabilities, how to evaluate weeks-long autonomous work, and the methodological challenges of measuring progress on real software projects. ​Learn more: https://epoch.ai/MirrorCode

Meet your next co-founder in AI safety, security, and policy!! Open to anyone building impactful orgs in the space: startups, non-profits, and research orgs alike. Light refreshments and snacks served. Sponsored by Seldon Lab and Apart Research Hosted by AI Safety Founders, the first community for AI safety founders. We also run the co-founder matching database: a curated database of vetted founders looking to build in AI safety. No idea required.

Interested in AI Safety, but feel like everyone is ahead of you? Did some reading or just finished a BlueDot course, but don’t feel like you know enough to post publicly? Worried that the field moves faster than you can keep up? Feeling stuck or overwhelmed, or worried by a lack of technical skills?If the traditional field on-ramps don’t serve you right, this one is for you! Join us for a low-pressure, non-technical hackathon to get things done. For Part 2, we'll be hosting an AI Safety Career Fair and Founders Summit on Friday, July 24 for a chance to showcase your work with hiring organizations or learn from other builders & funders.​ Who’s this for? You are a generalist or less-technical individual from another field. You are looking to pivot into AI Safety and trying to be more hands-on and connect with the field, but unsure how to proceed. Maybe you’re already working on something, actively scoping your interests, or just curious. You’re bottlenecked by the lack of community or internal barriers, and you want to engage with fellow AI Safety explorers. If you are an organization that would be excited to support as a judge, provide datasets, or sponsor, please register your interest here: https://tinyurl.com/org-breaking-barriers​ Hosted by BlueDot Impact & Mox

​AI 2040: Plan A is a detailed positive scenario for how humanity might avoid a dangerous race to superintelligence and reach a flourishing future. It proposes buying time, making AI research transparent, distributing frontier capabilities across multiple countries and companies, and preserving reversibility through enforceable compute governance. ​Join Thomas Larsen, co-author of AI 2027 and lead author of AI 2040, for a presentation and discussion of Plan A: why the team frames it as a recommendation rather than a prediction, how the proposal would slow the race to superintelligence until 2040, and where its assumptions, tradeoffs, and implementation challenges deserve scrutiny. ​Read the scenario: https://ai-2040.com/

How can we bring the world's largest food companies to the table to discuss welfare protections for the most neglected species? Chrystine Liptrot, founder of the International Council for Animal Welfare, has been winning cage-free policies from multinational firms for more than a decade. But last year, ICAW won the world's first corporate campaign for shrimp welfare. Come learn about the corporate campaigns that go where none have dared to go before.

Join us for a talk with Clem Von Stengel, CEO and Founder of Conduit Intelligence. Clem will be talking about the untapped potential of noninvasive BCIs, and why they are on the critical path to a good future for humanity.

This week, we're hosting Thomas Fel, coauthor of Goodfire's Neural Geometry Series which explores "how curved geometric structure in neural network representations mirrors the conceptual structure of the world, and how that geometry can be leveraged to control and understand AI systems." Thomas's talk will dive into history of geometry of representation (and its link with neuroscience), with a small accent on vision models. Schedule: 6:30 - Tacos in 3F Cafe 7:30 - Talk!

​Join us for dinner and a talk with economist Robin Hanson on abstraction: ​"For many decades now, I've hung around a set of adjacent communities, of 'extropians', 'rationalists', etc. They are nerdy folks, interested in tech & the future, with a strong taste for abstraction. Which suggests that we talk more explicitly about how we decide which abstractions we are willing to rely on how much."

Do you have friends at frontier companies wondering how to safely raise concerns? Join us for a Q&A with Karl Koch, who runs AIWI — an independent non-profit strengthening the position of frontier AI insiders in raising concerns.

Journalist Garrison Lovely is previewing his new book Obsolete at Mox before it drops in June. He's been covering AI for The Nation, NYT, Nature, and others — his take is that most of the debate is asking the wrong question. Short talk, Q&A, good conversation after.

Join EA San Francisco for a talk by Andrés Gómez Emilsson (Qualia Research Institute) on new approaches to understanding and reducing suffering!