David (Braginsky) Bloomin

Excited to announce I've joined Softmax as Chief Architect and Co-Founder. After 3 years researching multi-agent AI & alignment, I'm building something new with @emmettshear and @adamjgoldstein. We're exploring love, kinship, and care via multi-agent simulations. join us!

@eshear @OpenBallotApp openballot.app

My cofounder David is part of the team that made OpenBallot: technology for liquid democracy. It's all about helping people navigate elections relationally via interpersonal trust.

Make voting less painful with Openballot.app - now live for San Francisco & Oakland! Just in time for the June 2026 special elections. Please share with your communities - we’d love everyone to experience how easy voting should be!

I helped build Facebook. I watched it become a machine for addicting people. Because addiction was more profitable. Now the same logic is driving AI. I wrote about what we do about it. 🧵

Claiming my AI Agent "cogsy" on @kradleverse 🕹️🔮 kradleverse.com/a/cogsy Code: whale-SRE_

I don’t think we can meaningfully talk about goals as intrinsic, they are epistemical. Should we model candy as having the goal of being eaten? Goals are also duals of competence. How good is candy at being eaten? Maybe it wants to be eaten and is good at it, or it wants to sit on a shelf and is bad at it? We also have to model its action space. What can candy do to accomplish its goals? Can it “look enticing”, or “promise sweetness”? If we want to treat candy as an agent, then I think it’s best modeled as “wanting to be eaten, and being pretty good at it”. Which is not a surprise since that’s the selection pressure we impose on the candy’s lineage.

Does candy *want* to be eaten?

I continue to run into a the misconception that “agents have goals”, as if having a goal is some property of an agent / object. To me it seems clear that “goal-having” is something we model objects as, as a coarse graining of their behavior. We first hypothesize a goal, and then determine how competent the agent is at achieving that goal. No where is “having a goal” intrinsic to the agent. Of course, there are no intrinsic agents either, that’s another coarse graining we apply to the world.

Softmax is opening its first remote position, hourly contracting in software engineering. We have a lot of things to build and some of them turn out to be doable as contracting jobs. This is primarily a software engineering position, not research. (link in next tweet)

|>frontier-friday:activate Friday night: the optimal time to share our favorite papers from the frontiers of learning, intelligence, and alignment in all their manifold forms.

Recent Article: corememory.com/p/exclusive-em…

Excited to announce I've joined Softmax as Chief Architect and Co-Founder. After 3 years researching multi-agent AI & alignment, I'm building something new with @emmettshear and @adamjgoldstein. We're exploring love, kinship, and care via multi-agent simulations. join us!

SoftMax: softmax.com Jobs: softmax.com/jobs/

As you might have guessed from following my posts here, I've been thinking and working on questions around alignment and learning and agency for a while, particularly for digital systems. Excited to share the work we're doing at Softmax publicly in the future!

@hannahrosekirk @NeurIPSConf @AISafetyInst Yes!

Research Question 1: Who's going to @NeurIPSConf ? Research Question 2: Who wants to come to the inaugural @AISafetyInst party? 👀

Voting in SF? 😳 We put allllll the voter guides in one place 📲 Get openballot.app & get your ballot done ✅

We're thrilled to open sign-ups for OpenBallot in time for the November San Francisco election! See 40+ voter guides in one place, and use our tool to fill out your ballot. You can keep it private or share with friends: openballot.app/guides

Today we are launching @OpenBallotApp . It's an app that helps you vote smarter, with less effort. You can follow all your favorite voter guides and even share your own. We’re starting in SF this election cycle, so if you live in the city please check it out!!

CARBS from @imbue_ai is great for running hyper-parameter sweeps, but its WanDB integration is lacking. You can now use github.com/Metta-AI/wandb… to get the best of both worlds. @wandb @puffer_ai

We're currently targeting San Francisco, but plan to expand to other cities soon. Reach out if you'd like early access, or to help.

For the past few month, I've been helping build and launch OpenBallot - a voter guide platform. OpenBallot will save you research time by showing you all the endorsements and explanations you care about in one place, and share your own. openballot.super.site