Commentators called for slowing AI and warned of x-risk, but our tracker shows hiring still centered on core engineering and enterprise rollout, with safety and evaluation a small share.
The week’s split-screen: slow the frontier vs ship the future
Zvi Mowshowitz opened with a jolt: “The first Millennium Prize, Navier-Stokes, has fallen to AI.” He framed it as a sign of rapid capability gains and urged attention to frontier pace, writing, “As in, we need to slow the rate at which AIs increase their capabilities, to allow for the necessary alignment and safety work.” Gary Marcus cheered that turn: “We should all salute their tentative agreement to slow down the acceleration of this technology, given that not one of the AI companies seems to have good control over it.”
At the same time, others focused on what AI is already doing in the economy. Azeem Azhar wrote, “There are decades when nothing happens. This week, I am allowing myself that cliché.” He described a dramatic shift in enterprise adoption: “This year, nearly every single hand went up; I estimate some 95%.” Ethan Mollick emphasized broader impacts beyond existential risk: “Existential AI risk is obviously critical, but it is not the only AI thing that requires policy.” He also argued that frontier AI can be profitable in serving models: “There seems to be a persistent belief that frontier AI companies are unprofitable serving models but it appears that Anthropic has 80%+ gross margins on inference.”
Noah Smith described the public mood shift: “This was the week that AI safety hit the big time.” Simon Willison warned against fueling panic: “These ghoulish claims strike brazenly at the hearth, and given the obvious importance of AI, it is unsurprising that they have leapt into the mainstream, with people asking the natural question: how would that happen?” Margaret Mitchell pointed readers to a reality check on “agents going "rogue"”. Marcus added a note of skepticism about both doom and hype: “AI is not really going to take down the internet, and (as discussed in another earlier essay) it’s not going to cure cancer either, at least not anytime soon or on its own.”
These are big claims, often pointed in opposite directions. Here is what employer hiring tells us right now.
What employers are actually hiring for
Our tracker counts 4,261 open AI roles across 178 employers, taken from their own job feeds.
The mix is overwhelmingly toward building and deploying systems, not guarding them. By role family:
- Modelling and engineering: 1,646 open, 1,022 opened and 565 closed in 30 days, across 140 employers
- Data: 816 open, 511 opened and 219 closed in 30 days, across 123 employers
- Infrastructure: 416 open, 180 opened and 62 closed in 30 days, across 93 employers
- Research: 229 open, 64 opened and 19 closed in 30 days, across 50 employers
- Product and design: 95 open, 45 opened and 12 closed in 30 days, across 48 employers
- Evaluation and safety: 41 open, 12 opened and 5 closed in 30 days, across 19 employers
That is the core finding this week. For every evaluation and safety role in employers’ own feeds, there are roughly forty modelling and engineering roles. If “Pacing the Frontier” becomes the organizing principle for the ecosystem, it has not yet shown up in the hiring mix. The last 30 days saw 1,022 new modelling and engineering openings vs 12 in evaluation and safety.
Enterprise demand matches the customer story
Azhar’s claim that enterprise adoption turned a corner finds support in who is hiring. Accenture has 600 open AI roles and 796 opened in 30 days. PwC has 122 open and 97 opened in 30 days. Capital One shows 197 open and 142 opened in 30 days. These are the kinds of employers that reflect broad deployment rather than research milestones. Azhar wrote, “I’m not the only one to see an avalanche of customers.” Our data is consistent with that: the consultancies and large adopters are where the openings are thickest.
The role mix also lines up with implementation at scale. Data (816 open) and infrastructure (416 open) roles are second only to engineering. Those are the seats you fill when you are putting systems into production and integrating them with legacy software, not just when you are prototyping. That supports Mollick’s point that “We don’t need better models for AI to have wide impacts on jobs & society and we need to be preparing to encourage good outcomes & mitigate bad.” Even without new frontier breakthroughs, employers are hiring to build and operate AI in core workflows.
Frontier labs are hiring, but not signaling a pause
The frontier labs in this week’s posts are adding headcount, not freezing it. OpenAI has 170 open roles, with 49 opened in 30 days. Anthropic has 125 open, with 45 opened in 30 days. That does not answer Mollick’s profitability claim, which our data cannot test, nor does it verify Azhar’s statement that “Anthropic has signed compute agreements worth up to $517 billion in the last 11 months, compared with $180 billion in server rental through 2029 promised last December.” But it does show ongoing investment in talent at the core builders.
If the “pacing” commitments described by Zvi Mowshowitz and praised by Marcus were biting already, we would expect early signals in the hiring mix: more evaluation and safety roles relative to engineering, or at least a slowdown in new openings for core modelling. We do not see that yet. In the last 30 days, employers opened 1,022 modelling and engineering roles vs 64 in research and 12 in evaluation and safety.
Safety talk vs safety staffing
Noah Smith’s read that “AI safety hit the big time” and Willison’s warning about how we communicate risk reflect the public conversation. Mitchell’s post pointing to work on “agents going "rogue"” adds to that focus. But employers are not staffing safety at scale. Evaluation and safety roles are 41 open across 19 employers. That is small next to 816 open data roles across 123 employers and 416 open infrastructure roles across 93 employers.
This is not an argument about what should happen. It is a report of what is happening in hiring. Right now, the market is building. As Marcus puts it on the tempered-expectations side, “AI is not really going to take down the internet,” but companies are certainly putting it into the network and into processes. The demand signal is for people who can ship, not for people whose primary job is to slow things down or vet them. If “the work can begin,” as Zvi wrote, that work in practice is still dominated by engineering.
Skills on the ground: productization is not yet the whole job
Simon Willison surfaced Laurie Voss’s observation about the shift in software work: “The cost of writing code collapsed, and the cost of reviewing, fixing and operating it is following, and I'm assuming it gets there.” Voss argues that what remains is product definition and UX. Our data suggests product hiring is growing but still small relative to engineering. Product and design has 95 open roles across 48 employers, compared to 1,646 in modelling and engineering. If Voss is right about where value pools will land, hiring has not fully pivoted there yet. It looks more like a buildout phase in which organizations are still creating the technical capacity to deliver AI features, with productization demand likely to rise later.
Azeem’s roundup also flagged skill retention with AI assistants among lawyers. We cannot test that with job postings, but we can see employers looking for people who can operationalize these tools: the mix of data, infra and engineering roles is what you would expect if AI is being threaded through existing teams’ work, not cordoned off to skunkworks.
What we can and cannot infer from hiring
- Capability milestones. Mowshowitz’s claim that a “Brand New AI Solves a Millennium Prize” is beyond the scope of our tracker. Hiring does not confirm or deny it.
- Profitability and capex. Mollick’s claim about “80%+ gross margins on inference” and Azhar’s compute figures are not visible in job feeds. The closest proxy we have is that frontier labs are still adding roles.
- Pace the frontier. Calls to slow capability growth are not yet reflected in a shift toward evaluation and safety hiring. Employers opened 12 such roles in the last 30 days, compared with 1,022 in modelling and engineering.
Bottom line
AI’s center of gravity this week looked split between calls to slow down and evidence that adoption is accelerating. On the hiring side, there is no split. Employers are hiring to build and deploy. Consultancies and large adopters are opening the most roles. Frontier labs are still hiring. Safety and evaluation remain a small fraction of the market.
Azhar wrote, “This year, nearly every single hand went up; I estimate some 95%.” Mollick added, “It is very clear that This Time is Different in terms of AI compared to previous innovations.” Our numbers support the direction of travel they describe, and they do not yet support the idea that “Pacing the Frontier” has shifted how employers staff. If that changes, it will show up first in the roles. For now, the market is voting for builders.
What we read
Every quote above is taken verbatim from one of these posts.
- Zvi Mowshowitz: Brand New AI Solves a Millennium Prize, We Must Pace The Frontier
- Noah Smith: Two missing pieces in the AI safety discussion
- Azeem Azhar: 🔮 Look up, the curve turned #601, 📈 Anthropic’s $517 billion shopping list
- Simon Willison: The contagion of fear, Quoting Laurie Voss, shot-scraper 1.12
- Gary Marcus: Two cheers (out of three) for Dario Amodei, President Trump’s Date With Destiny?
- Ethan Mollick: Existential AI risk is obviously critical, but it is not the only AI t, There seems to be a persistent belief that frontier AI companies are u, It is very clear that This Time is Different in terms of AI compared t
- Margaret Mitchell: If you read just one thing about the realities of modern AI and agents