×
Softmax
AI alignment research lab pursuing "organic alignment", how intelligent agents learn to cooperate without being controlled
Last verified August 17, 2026 · Updated daily
What Softmax is building
Softmax is a research lab and eventual product company built around a single thesis: that AI alignment doesn't have to be hierarchical and control-based. Instead, the founders believe intelligent systems can learn to align themselves, the way cells form organisms, ants form colonies, humans form teams. They call this "organic alignment." In practice, that means running reinforcement learning experiments with small-scale agents in virtual environments to study how cooperative behaviours, role differentiation, and shared goals emerge naturally from multi-agent interaction. The models are small (millions of parameters, not billions) and the work is firmly in the research phase, but it's generating outputs: published writing, multi-agent frameworks, and an experimental codebase in active development.
Why this matters
The dominant approach to AI alignment, RLHF, constitutional AI, hard-coded constraints, works by imposing values from the outside. Softmax's argument is that this produces brittle alignment that breaks as systems get smarter. Their alternative is grounded in biology and collective intelligence research: if you build the right conditions for cooperation, alignment emerges as a property of the system rather than a constraint on it. This matters because the field is bifurcating: one branch is scaling models faster than safety understanding can keep up, another is trying to solve safety before scaling further. Softmax sits in the second camp with a genuinely novel theoretical framework, not a tweak to existing approaches, but a different starting point. The a16z backing signals that at least one major institution believes this framework can eventually commercialise. Plural, a European VC focused on social and governance technology, adds further signal that this isn't purely academic, there's a product arc being sketched.
Open roles at Softmax
3 positions we're tracking. Roles are re-checked daily and removed when filled.
Know when Softmax is hiring before anyone else
A role stays uncontested for about four days. Here's the window — and where we put you in it.
From $9/month, cancel any time.
Watching Softmax
0 applicantsRole spotted & verified
1You get the alert
1You've applied
~8Hits the job boards
250+Hiring outlook
Research lab with a live experiment pipeline, 10 people, 2 open roles publicly posted, unusually few for a team at this stage with this investor pedigree. The next hires are coming, they just haven't listed them yet.
Working at Softmax
Since , Softmax has built AI alignment research lab pursuing "organic alignment", how intelligent agents learn to cooperate without being controlled. The team is now <50 people. For a AI company this size, the reality is broad remit, direct access to founders, and equity that still means something if the company works out.
Most openings are based out of San Francisco.
How to actually get hired at Softmax
Why applying the normal way doesn't work
At <50 people, Softmax has no recruiting team. Your application lands with a founder who is also running sales, product and payroll. The obstacle isn't a queue or an ATS, it's being seen at all. Cold outreach outperforms the form here, consistently.
Who to contact at Softmax
Founding team
Emmett is one of the few people in tech who can say he built a company from zero to acquisition, ran a major AI lab through a crisis, and then pivoted into pure research, and makes each of those moves feel like part of a coherent intellectual project rather than a résumé. He co-founded Justin.tv in 2007, spun out Twitch in 2011, and ran it as CEO for 16 years, through the Amazon acquisition in 2014 and well into its maturity as the dominant live-streaming platform. What's less often noted is that Emmett is deeply ideological: he has been writing and thinking publicly about AI consciousness, AI rights, and the philosophical foundations of alignment for years before Softmax existed. When he was named interim CEO of OpenAI during the November 2023 board crisis, he described it afterward as "a lot of thrash without much real change", a characteristically blunt assessment that tells you he's not in this for the prestige. He describes current AI systems as analogous to bacteria or fish on the scale of general intelligence, "less than" human, but approaching the line, and argues that the time to build alignment infrastructure is now, before the line is crossed. His published writing on softmax.com and LessWrong is serious philosophy, not founder-brand content: he writes about the nature of consciousness, the frame-dependence of reality, and why he thinks the "enslaving the wish-granting God machine" approach to AI is going to fail badly. He's based in San Francisco, in-person with the Softmax team, and is known to engage with people who come to him with substantive intellectual challenges rather than credentials.
David is the builder on a team of thinkers, and that role is more important than it sounds. He has two decades of large-scale infrastructure experience at Google, Facebook, and Asana, the kind of engineering background that knows what it takes to go from experiment to production system. He co-founded the Plurality Institute, a research initiative exploring novel governance models and multi-agent AI frameworks, which means his intellectual interests were already firmly in the territory Softmax is exploring before the company existed. He leads the MettaGrid project at Softmax, the simulation environment where the organic alignment experiments actually run. If you're a research engineer or systems person looking to engage with Softmax, David is likely your first call: he understands what "technically viable" means in a way that complements Emmett's philosophical framing and Adam's biological research. His public profile is deliberately low-key; he doesn't do a lot of public writing under his own name, but his GitHub activity and the Plurality Institute work give you a clear picture of where his head is.
What to show them
This is the core of what Softmax is doing. The lab runs RL experiments with agents in virtual environments. The existing team (David Bloomin leads engineering; Goldstein and Shear do research and strategy) needs people who can build the experiment infrastructure, run the simulations, and iterate on the agent architectures. The MettaGrid project is the current technical centre of gravity. Core skills: Python, JAX or PyTorch, multi-agent RL (MARL), simulation environments (PettingZoo, MeltingPot, OpenSpiel), emergent behaviour analysis, distributed training, research engineering (not just research). For proof of work, fork the MettaGrid repo on GitHub (github.com/Metta-AI/mettagrid if public, or equivalent Softmax-adjacent work), run the existing experiments, and write a 2-page technical memo documenting: what the agents currently learn, where cooperative behaviour breaks down in your runs, and one architectural change you'd make to test whether a different reward structure produces more stable role differentiation. Send it with your email.
A cold email that works at Softmax
What Softmax screens for
The two listed roles are research-oriented, which tracks. But a lab that describes itself as "part research lab, part aspiring money maker" will need more than researchers. The product commercial layer hasn't been built yet. Shear has said explicitly he doesn't yet know what form the commercial product will take, which is the honest founder answer, and also the tell that product and commercialisation infrastructure doesn't exist yet. a16z's investment pattern with research-leaning AI companies (Mistral, various safety labs) pushes these teams toward three things: publishing research publicly, building developer communities, and finding the path from research to deployable product. All three of those require hires that aren't researchers.
Customize your CV for the Softmax role. Matching the job description language helps clear ATS filters.
Don't make these mistakes
The biggest mistake with Softmax is coming in as a true believer in the AI safety movement without engaging with what makes Softmax's approach different from other alignment efforts. These founders are not just building another safety lab. They have a specific, technically grounded, philosophically distinct framework, and they will immediately identify whether you've engaged with it or not. Don't reference RLHF or constitutional AI as the frame for what Softmax is doing, that misses the entire point. Don't position yourself as someone who cares about AI safety; position yourself as someone who has a specific thought about organic alignment, multi-agent cooperation, or the biological models they're drawing from. Emmett responds to intellectual challenge and direct disagreement; he explicitly said in his first Softmax interview that he wants to be pushed on the ideas. Treat that literally.
Mistakes that kill Softmax applications
A recycled CV gets rejected fast at Softmax (<50 people). They notice.
Skip 'I'm looking for...' — start with The dominant approach to AI alignment, RLHF, constitutional AI, hard-coded constraints, works by imposing values from the outside and your specific angle on solving it.
Applying and waiting = silence. Follow up at day five — it roughly doubles your odds of a reply.
Applying to Softmax? Get the contact, not the form.
The Softmax interview process
3 stages · 7 days typical · take-home: yes · modelled from similar companies
We don't yet have verified candidate reports for Softmax. What follows is the typical process for a <50-person AI company — treat it as a model, not confirmed detail.
Interview stages
Intro Call
Video call · 30 min
Culture fit and role expectations
Founder or hiring manager
Technical Deep Dive
Video call or in-person · 60 min
Past projects and problem-solving approach
Technical founder or lead
Final Round
In-person or video · 45 min
Team fit and offer discussion
Founding team
Softmax take-home assignment
Softmax includes a take-home exercise in their interview process. For AI roles, this typically involves a practical problem that takes 2-4 hours. Focus on clean, working code over premature optimization. They're evaluating how you think and communicate, not just the solution.
Softmax interview timeline
Timeline: ~ days. That's faster than the AI median (10 days at <50 people).
Interviewed at Softmax?
Tell us how it went — stages, questions, timeline. Takes 90 seconds and it's how this page stays accurate for the next person.
Submit your Softmax interview experience →Softmax jobs, frequently asked questions
How many jobs does Softmax have open?
3 open roles at Softmax, last checked March 2025.
Does Softmax hire remotely?
Softmax doesn't have remote openings at the moment. All roles are in San Francisco.
What roles is Softmax hiring for?
Softmax is hiring across Engineering, Other, Data. The most recent opening is Research Engineer (Multi-Agent RL).
How do I apply for a job at Softmax?
Use the apply links above, or check our guide to getting hired at Softmax.
Does Softmax respond to cold emails?
Not enough data yet on Softmax's cold email response rates.
Who is the hiring manager at Softmax?
At this size, hiring is usually run by a founder or department head.
How competitive is it to get hired at Softmax?
Typical applicant count for AI roles (<50 people): 50-100 in two weeks. Apply fast.
How many rounds is the Softmax interview?
3 stages: Intro Call, Technical Deep Dive, Final Round.
Is the Softmax interview hard?
Focus on technical depth and system design, not puzzles. Technical Interview is reportedly the most challenging round.
Does Softmax give a take-home task?
Yes, Softmax includes a take-home assignment.
How long does Softmax take to get back to you?
Around 7 days across the full process.
What should I prepare for the Softmax interview?
technical depth and system design is the priority. Show you can work autonomously — that matters more than algorithms at <50 people.
Where is Softmax based?
Softmax is headquartered in San Francisco.
Get Softmax roles before they're posted
A role stays uncontested for about four days. Here's the window — and where we put you in it.
From $9/month, cancel any time.
Watching Softmax
0 applicantsRole spotted & verified
1You get the alert
1You've applied
~8Hits the job boards
250+