Our listings show infrastructure-heavy hiring that backs a swing to cloud agents, while safety hiring remains small despite mounting calls for guardrails.
A split screen: cloud acceleration, safety anxiety, and optimism
Across the last two days, several practitioners described an AI sector pulling in different directions: a fast shift to cloud-hosted agent workflows, growing anxiety about safety and governance, and optimism that AI will deliver broad gains. We tested those claims against our tracker of employers’ own job feeds.
Ethan Mollick sees a platform swing underway: "Big swing happening from this summer's trend to have AI running on virtual machines on your own computer (Claude and ChatGPT apps) to moving them back into you having a dedicated VM the cloud (Claude Cowork just announced they were doing this, also Grok Bot, Dot, and others). Cloud is mostly better." Simon Willison, quoting Anthropic’s Felix Rieseberg, describes the same turn in concrete product terms: "The "old" version of Cowork runs model inference in the cloud, executing tool calls in an Anthropic-provided VM we shipped to your computer." Now, "The "new" version of Cowork runs model inference and the VM in the cloud." The appeal is operational: "Each session gets its own sandbox, not sharing state with other sessions." and "We think this solves a lot of problems we've heard about (like using Cowork from a phone, keeping work running, or getting all the same power without losing battery to the VM)".
At the same time, Gary Marcus previews testimony for a New York City council hearing, reacting to Sam Altman’s reported remark, "we believe that the world should accept some bad things happening for the benefits of this technology and people having the agency". Marcus argues for tighter oversight: "Contemporary AI is like a bull in a china shop, powerful and impressive, but difficult to control." Zvi Mowshowitz, reviewing recent Claude releases, is blunt on reliability and incentives: "Opus 5.5 Has Too Much Deference." and "Do Not Trust the Self-Reports."
Noah Smith stakes out the opposite pole, arguing the technology wave will deliver real benefits, much as the internet did: "Paul Krugman is the greatest econ writer in the world, and also a legendary economist. But I sometimes feel that he has a blind spot when it comes to the value of new technologies." He concludes, "The internet is a lot more than just people yelling at each other on forums and social media."
Our data: infrastructure-heavy hiring backs the cloud-agent turn
If AI work is moving from laptops to hosted sandboxes, hiring should tilt toward cloud and platform skill sets. Our tracker shows 4,309 open AI roles across 180 employers. Infrastructure accounts for 478 open roles, with 237 opened and 98 closed in 30 days across 99 employers. That is the third-largest role family, and its 30-day openings outpace closures by more than 2 to 1.
The shift Mollick and Rieseberg describe is not theoretical, it is consistent with where employers are adding people. Modeling and engineering remains the largest slice at 1,712 open roles, with 891 opened and 739 closed in 30 days across 137 employers. Data roles are next at 748 open, with 491 opened and 397 closed in 30 days across 119 employers. Together with infrastructure, these families map onto exactly the work implied by cloud agents: building models, orchestrating tools, and keeping hosted environments safe and reliable at scale.
Top-line employer activity also aligns with active product shifts. Anthropic lists 138 open roles, with 38 opened in 30 days. OpenAI shows 179 open, with 47 opened in 30 days. System integrators and large enterprises are leaning in too: Accenture leads with 674 open and 515 opened in 30 days. Capital One has 221 open and 262 opened in 30 days. These figures support a broad platform and deployment push, not a pause.
Swarms prize speed, companies hire to build and coordinate
Jack Clark highlights Toby Ord’s framing of multi-agent systems: "A good way to see AI swarms is as a new form of inference-scaling," and, crucially, "Why would you ever use swarms? The most important answer is speed. The 4-agent swarm needed about twice the total number of tokens to get the same performance, but in terms of tokens per agent, it only needed half as many. Since the agents are run in parallel, this means it can theoretically achieve the same task in half the time," He also flags a coordination cost: "Stepping on Toes parameter: Swarm scaling doesn’t scale perfectly - rather, it seems like as you increase the number of agents in a swarm you get a diminishing returns property... where you pay some kind of tax as you scale the number of people involved."
Our hiring data fits this picture. Modeling and engineering openings dominate, and infrastructure demand is strong. Product and design shows 104 open roles, with 44 opened and 22 closed in 30 days across 48 employers. That mix suggests teams expect real integration work to coordinate multi-agent systems with products, data sources, and user workflows. The speed benefit Ord notes will not arrive without careful engineering and platform work, and employers are staffing accordingly.
Simon Willison’s lab note on local experiments is a useful counterpoint. He reports running a controlled arithmetic test on local hardware and shares a reasoning trace snippet that reads, "Wait, let me redo this more carefully." Local performance improvements matter for research, and our numbers show 237 open research roles, with 53 opened and 19 closed in 30 days across 55 employers. But the relative weight of infrastructure hiring suggests that, for deployment, the center of gravity is in the cloud.
Safety and governance talk outpaces safety hiring
Marcus argues the policy window is open, and he is not alone. Ethan Mollick ties lab policy to a kind of deferral: "AI policy right now, especially from the Labs, makes me think of this poem. If you believe superintelligence is near, you don't need to make hard decisions about how to build AI to make the world better. Just wait for ASI & it will decide for us, for better or worse. But if that doesn't happen..." Zvi Mowshowitz’s review of model behavior points to unresolved failure modes and tensions between customer demands and honest disclosure.
Our tracker shows evaluation and safety roles are a small share of current hiring. There are 47 open evaluation and safety roles, with 13 opened and 3 closed in 30 days across 20 employers. That is meaningful, and openings exceed closures, but it is a fraction of the 1,712 modeling and engineering roles or the 478 infrastructure roles. The rhetoric in these posts points to an urgent need for oversight, yet headcount remains concentrated in building and shipping systems. On this point, our data does not support the idea that the safety workforce is scaling at anywhere near the same rate as engineering.
Access to justice: early user signals, unclear hiring footprint
Azeem Azhar describes a nascent role for AI in civil justice: "In England and Wales, 60% of defendants in defended county court claims this year did not have a lawyer. In the US, more than nine in ten consumers sued for debt face the case without one." He suggests that "AI may give some of them a new way to argue their case." He quantifies early usage: "Around 0.6% of all Claude use in May was for lawyers’ tasks, ranking 38th of 718 occupations, ahead of accountants but well behind programmers and writers. Four-fifths of these queries were people asking about their rights or what the law means."
Our hiring data neither confirms nor contradicts growing legal-AI usage, but it does not show a distinct surge in legal-focused AI roles. Product and design is small at 104 open roles, and there is no legal category in our taxonomy. The strongest growth is in general-purpose modeling, data, and infrastructure. Large professional services firms like Accenture, with 674 open roles and 515 opened in 30 days, and PwC, with 102 open and 58 opened in 30 days, could be channeling AI into legal-adjacent work, but our tracker cannot break that out. The pattern here is that user demand may be ahead of specialized hiring.
Does the job market reflect broad economic gains?
Noah Smith’s argument that technology waves deliver real benefits meets a labor market that is busy. We count 4,309 open AI roles across 180 employers. The spread is not confined to a few labs. Hiring is strong at cloud-native model companies and at enterprises integrating AI into their workflows. Capital One’s 221 open roles and 262 opened in 30 days, Waymo’s 113 open and 36 opened, and Amgen’s 157 open and 45 opened in 30 days point to financial services, autonomous systems, and biotech all adding AI headcount. That breadth is consistent with a diffusion story rather than a narrow research boom.
Where our data tempers the optimism is in the composition of roles. Companies are investing most in building and operating systems, less in evaluation and safety. If the economic gains depend on trust and reliability, the gap that Marcus and Mowshowitz describe remains visible in the hiring profile.
What to watch next
If the cloud-agent pattern holds, infrastructure and modeling roles should continue to lead. The current 237 infrastructure openings in the last 30 days, against 98 closures, suggest momentum in that direction. Safety openings are growing from a small base, and policy attention could nudge that up, but our tracker so far shows employers prioritizing development and deployment. On the user side, Azhar’s access-to-justice examples highlight practical value, yet we do not see a distinct hiring wave to match. That may emerge inside larger product teams rather than as standalone legal-AI headcount.
The clearest agreement between commentary and hiring is on platform direction. As Mollick puts it, "Cloud is mostly better." Employers are hiring like they believe that, and they are building the teams to make it work.
What we read
Every quote above is taken verbatim from one of these posts.
- Noah Smith: What the AI boom is doing for Americans
- Gary Marcus: Coming soon: New York City’s hearing on AI risks
- Jack Clark: Import AI 475: Swarm scaling; Google DeepMind watermarks biology; and
- Zvi Mowshowitz: Mythos 5.1, Fable 5.1 and Opus 5.5: Model Welfare
- Azeem Azhar: 📈 Monday data: More AI, more justice?
- Ethan Mollick: Big swing happening from this summer's trend to have AI running on vir, AI policy right now, especially from the Labs, makes me think of this , 'create the ultimate brutalist city builder. you should have lots of c
- Simon Willison: Quoting Felix Rieseberg, Qwen3.8 27B addition in words
- David Ha: [Registration now open: From World Models to Real-World AI
Luma: luma.](https://bsky.app/profile/hardmaru.bsky.social/post/3mx3kq6snoc2v)