Three AI Labs, 708 Open Roles, And a Glimpse of the Future
Robots, interpretability, and FDEs.
I had some down time over the holiday and so decided to do what any data nerd does— look at some data. I wanted to get a sense of hiring at the major labs: Anthropic, DeepMind and OpenAI, the three labs most directly competing on frontier models. The thesis here is simple: jobs postings are statements about priorities and analyzing them can reveal the bets labs are making… or not making. And while I agree that most armchair theories about strategy are wrong1, it is my job to make bets about the future.
I pointed Claude Code to the career pages of the top AI labs and had it do what it does best: analyze text. I interrogated most findings, pushed for clarity on others, straw-manned and steel-manned for a bit, and here we are, 708 roles later. (N.B. It is truly a wonderful time to be alive if you’re a person who enjoys the journey of knowledge- and perspective-building.)
Let’s start with a caveat: job postings are an imperfect signal. They don’t capture internal mobility, acqui-hires, or roles filled through back-channels. But public postings remain one of the few consistent, comparable data sources across labs.
High level trends
We start by categorizing each open role into a major function across companies using a radar plot, which allows us to easily see how companies compare and contrast.
Starting with DeepMind (blue): More than a third of open roles are in research, at 36%. Only 6% of open roles are in GTM. Most notably, DeepMind is the only lab with open roles in Hardware & Silicon. They are building capabilities and letting Google handle distribution through Gemini integration across Search, Android, Cloud, and Workspace. Interestingly, there are 20 Product roles open, all Gemini App-focused:
Product Managers for: Personalization, Notifications, Discovery, Monetization, Activation, Growth
Content designers for model UX
Mobile engineers
This is fascinating and quite the combo.
Anthropic and OpenAI are now neck-and-neck on GTM. Both labs are at ~40% GTM/Sales, but the composition differs dramatically:
Anthropic is building a traditional enterprise sales motion (AEs close deals). OpenAI is deploying engineers directly to customers (FDEs implement solutions). OpenAI has 12.5x more open FDE roles open than Anthropic.
Taking a look at safety, Anthropic has the most open roles: dedicated ML engineers are actively building safety into their products via the “Safeguards” team. OpenAI’s Alignment team exists but isn’t publicly hiring. Their safety hiring is traditional security, not AI safety. This hiring data corroborates public signals: OpenAI disbanded its Superalignment team in May 2024, and its AGI Readiness team in October.2 DeepMind is making some investments here, but it’s nothing noteworthy afaict.
Research Agendas
Job reqs reveal fundamentally different priorities across labs.
At DeepMind, I see a lab betting on embodiment and longer-term (5-10 years) applications. They’re the only lab with Hardware & Silicon roles (3%). These folks are likely working on TPU-related projects. In 2025, DeepMind launched Gemini Robotics—vision-language-action models that translate instructions into motor commands—and released the ASIMOV dataset for robotic safety.3 Robots, alongside emotional-social AI, and even weather suggest a deeper investment in a physically-oriented approach. The Multi-modal focus may reinforce (and strengthen) this.
At Anthropic, I see research concentrated on understanding what’s happening inside models. There are 3 roles dedicated to interpretability, and honesty is a research area (!) Their ‘Scaling Monosemanticity’ paper (2024) and circuit tracing work (2025) represent the field’s most ambitious attempts to understand what happens inside large models. The Safeguards team (ML engineers building safety into the product) represents a distinctive approach, with safety as product feature. Alignment is obviously an overriding focus. There’s heavy RLHF investment, as well as pre-training. It’s AGI or bust, so Anthropic will continue to move research forward. Not surprisingly, there’s v little investment in multimodal compared to other labs. Maybe interpretability is seen as the necessary precursor to multimodal?
At OpenAI I see a corporation hiring for few research roles and instead rapidly productizing existing research. Maybe fundamental research is done, or done enough? While they have 10 named research teams, they have only 7 research job postings across those teams. Why so few research postings? I see three possibilities: (1) research teams are fully staffed, (2) they’re hiring through non-public channels, or (3) the research phase is over and it’s time to ship. 31 open Product Engineering roles suggest the latter.
What to Expect in Coming Months
I’ve always held that the gold standard of data work is prediction, so here’s what I expect at each lab in the next 6-12 months.
At DeepMind, open hardware roles suggest investment in physical robots, but with longer timeframe to impact. We may see improved emotional/personality capabilities in Gemini. Getting these right could make any robots more… likable? The 20 open Product roles (all Gemini App-focused) on growth and monetization suggest they do not intend to be a bit player in the consumer/chat/video wars. Again, Google’s advantages in distribution (4B users!), funding, compute (TPUs), data (huge training data advantage, esp wrt to computer use) and perhaps most importantly, patience (they’re a public co that prints money, after all), will continue to compound. Continued investment in research will lead to major research breakthroughs.
Revenue will continue to surge at Anthropic, driven by major enterprise deals, especially in regulated industries. Interpretability supports this well: When a bank or hospital deploys AI, regulators ask ‘how does it work?’ Anthropic is the only lab that appears interested in that question. Claude DevEx is also strong, and the MCP ecosystem will continue to expand. All of these are reasons to think Anthropic is positioned to win enterprise AI. The Safeguards team hiring suggests “safety as product feature” will become more visible—expect honesty features, uncertainty detection, and calibration in Claude. I expect to see major research publications on interpretability, and honesty features in Claude may become defining. I
Finally at OpenAI, I expect a consumer hardware reveal given camera/connectivity engineer hiring (and recent “leaks”). The Forward Deployed Engineer model (25 roles—12.5x more than Anthropic) suggests OpenAI is betting on implementation depth over traditional sales. Like Anthropic they will go after regulated industries. Expect more “we’ll build it for you” enterprise motions. We will see various product launches and integrations/partnerships to try to move the needle on un-monetized users. On this dimension though (Consumer AI), I expect Google to win given structural advantages.
Up Next
I’ll share some thoughts what these findings mean for startups in a follow-up piece.
Appendix
Data Sources
Anthropic: https://www.anthropic.com/careers/jobs
DeepMind: https://job-boards.greenhouse.io/deepmind
OpenAI: https://jobs.ashbyhq.com/openai
How we mapped departments
GTM/Sales: Sales, Marketing, Forward Deployed Engineers, Solutions Architects, Partnerships
Research: Fundamental AI Research, Frontier AI, Interpretability, Alignment
Product: Product Management, Product Design, Developer Relations
Product Engineering: Software Engineers building user-facing products
Infrastructure: Compute, Scaling, Platform Engineering
Hardware & Silicon: Chip design, Electrical/Mechanical Engineering
Safety/Security: AI Safety, Trust & Safety, Security Engineering
G&A: HR, Finance, Legal, IT, Operations
https://www.cnbc.com/2024/05/17/openai-superalignment-sutskever-leike.html, https://www.axios.com/2024/05/17/openai-superalignment-risk-ilya-sutskever
https://deepmind.google/blog/gemini-robotics-brings-ai-into-the-physical-world/, https://github.com/google-deepmind/open_x_embodiment
What strikes me is how *few* jobs these are—we must be looking at 50x talent in these roles, not just 10x