Skip to main content
AI Jobs·Report

ANALYSISReported by AI Jobs Report

Pacing talk, fast hiring: safety roles still a sliver

Commentators called for pacing, warned of fear, and flagged regulation. Our tracker shows capability hiring surging, safety roles small, and no sign of an AI-led hiring pause.

Read the original at AI Jobs Report
9 min read20 viewsBy AI Jobs Report

Commentators called for pacing, warned of fear, and flagged regulation. Our tracker shows capability hiring surging, safety roles small, and no sign of an AI-led hiring pause.

Speed versus pacing: what was said

Zvi Mowshowitz summarizes the case for slowing down frontier progress in “We Must Pace The Frontier,” writing: “As in, we need to slow the rate at which AIs increase their capabilities, to allow for the necessary alignment and safety work.” He is not optimistic about the road ahead, adding: “The odds are still against us. The situation remains grim.” Yet he still frames it as a turn: “But this is Actual Progress. The work can begin.”

Gary Marcus argues that even with political noise, a US framework for screening models is emerging, if opaque. He writes, “Remember, even as Trump shouts from the rooftops of Truth Social that he is violently opposed to regulating AI … … he is in fact regulating AI, with a secret AI security framework for evaluating which “frontier” models can be released.” After reviewing heavily redacted documents, he concludes, “Either way, U.S. policy on screening AI remains opaque for another day.” In a separate post about the much-discussed “slowdown,” he offers a blunt read of the corporate stance: “Translation of the highlighted bit? We will go as fast as we can without going to jail or getting sued out of existence, but for optics we will it call it “pacing”.” He also cites Elon Musk’s line: “It’s certainly some crazy 4d chess to say there’s whatever a 10% chance of annihilating humanity. But by the way, um, how much allocation would you like in our IPO?”

The safety debate overlaps with rising public anxiety. Simon Willison reflects on how doomsday narratives can spread, quoting a claim that many Anthropic researchers believe AI “could kill us all by the end of the decade”. He relays Bryan Cantrill’s warning against vague alarms that outrun expertise, citing examples like “hacking critical infrastructure” and “extinction-level bioweapons.” The admonition is clear: “The lesson that I learned (shamefully) decades ago is that domain experts, by way of their expertise, implicitly hold the public’s trust — and we must not abuse it.” Margaret Mitchell points readers to a cautionary explainer, posting: “If you read just one thing about the realities of modern AI and agents going "rogue", read this: aiguide.substack.com/p/misleading... Brilliant piece from @melaniemitchell.bsky.social”

Product releases and compute ambition point to acceleration

Even as pacing enters the vocabulary, product cycles are not pausing. Simon Willison notes, “Google released Gemini 3.8 Live and 3.8 Live Extended Thinking today - two new speech-to-speech models that are a similar shape to OpenAI's GPT-Live family.” He built a web UI to test them and highlights the maturity of the developer surface: “You can select a model and voice preset, enter an optional system prompt and then start a voice conversation through your browser, including the ability to interrupt the model while it is talking.” The engineering friction is dropping too: “The implementation uses no libraries.”

On the capital side, Azeem Azhar flags an enormous scale-up in compute commitments: “Anthropic has signed compute agreements worth up to $517 billion in the last 11 months,” and observes broad corporate claims of impact: “Every week, we share the latest updates on the State of the AI Economy based on our proprietary research.” He adds, “Around a third are directly attributing financial impact to their implementation of AI.” In a separate reflection on Adam Smith, he offers a framing likely to resonate with enterprise buyers and workers alike: “I reckoned he’d love the technology as a productivity tool; he’d be skeptical about companies working together on any issue; and he’d be concerned about giving moral agency to machines.”

What our tracker shows on hiring now

Our tracker counts 4,318 open AI roles across 178 employers, taken from their own job feeds.

By role family:

  • Modelling and engineering: 1,674 open, 1,042 opened and 554 closed in 30 days, across 141 employers
  • Data: 841 open, 545 opened and 226 closed in 30 days, across 125 employers
  • Infrastructure: 418 open, 184 opened and 66 closed in 30 days, across 95 employers
  • Research: 230 open, 66 opened and 20 closed in 30 days, across 50 employers
  • Product and design: 97 open, 47 opened and 12 closed in 30 days, across 49 employers
  • Evaluation and safety: 44 open, 15 opened and 5 closed in 30 days, across 20 employers

Top employers:

  • Accenture: 631 open, 813 opened in 30 days
  • Capital One: 192 open, 144 opened in 30 days
  • OpenAI: 174 open, 51 opened in 30 days
  • Amgen: 164 open, 56 opened in 30 days
  • Anthropic: 126 open, 47 opened in 30 days
  • PwC: 114 open, 98 opened in 30 days
  • Waymo: 90 open, 15 opened in 30 days
  • Databricks: 88 open, 26 opened in 30 days

Across every role family, more jobs were opened than closed in the last 30 days. The largest gap is in modelling and engineering, where 1,042 roles opened and 554 closed. Data roles show a similar pattern, with 545 opened and 226 closed. Even evaluation and safety is net positive with 15 opened and 5 closed, but the absolute count is small at 44 open roles across 20 employers.

Does the safety turn show up in jobs yet?

Not much. The rhetoric of pacing and safety does not correspond to a hiring pivot. Evaluation and safety roles are only 44 open out of 4,318, while modelling and engineering accounts for 1,674. Our data therefore does not support the idea that companies are materially shifting headcount toward safety, even as leaders discuss pacing and governments draft screening policies.

This gap aligns more with Marcus’s skepticism about corporate incentives than with a substantive rebalancing toward safety. His read that companies will push speed up to legal boundaries — “we will go as fast as we can without going to jail or getting sued out of existence, but for optics we will it call it “pacing”.” — is consistent with the continued surge in capability hiring and product releases. Zvi Mowshowitz’s claim that the safety camp has achieved “Actual Progress” may be true in discourse and policy attention, but our listings show only a narrow channel for dedicated safety hiring so far.

Regulatory opacity, as Marcus notes with “U.S. policy on screening AI remains opaque for another day,” is not yet visible as a drag on recruitment at leading labs. OpenAI lists 174 open roles with 51 opened in the last 30 days, and Anthropic lists 126 open with 47 opened. That suggests active expansion rather than a pause.

Misuse, threat models, and where headcount goes

Zvi’s read of Anthropic’s misuse report is cautious but upbeat: “A lot of bad guys try to use Claude to do bad things. Mostly they fail. We think.” He flags a different long-run concern: “Indeed, distillation is by far the most important threat in this report, and the part of the report that will have the most impact.” If companies internalize that diagnosis, we would expect more hiring in red-teaming, evals, and model security. Our tracker does not yet show that shift at scale. Evaluation and safety remains a tiny slice compared to data and engineering. If threat hardening becomes a top priority in practice, we should see more openings in that category across far more than 20 employers. We do not see it yet.

Layoffs, productivity, and the shape of work

Alex Hanna challenges a popular narrative about job cuts: “The second myth is that "AI" is the reason for recent layoffs, but a deeper look tells a different story.” Our tracker cannot speak to layoffs directly, but the live demand picture does not show a hiring freeze in AI roles. Accenture alone lists 631 open, with 813 opened in the last 30 days, and major adopters in finance, consulting, and biotech continue to expand.

Ethan Mollick points to real productivity shifts: “Study from Google about AI use in science, complex impacts: acceleration (7 hours saved per week) along with shifts in the kind of work (more verification) and what research gets done (possibly safer topics).” He adds, “It is very clear that This Time is Different in terms of AI compared to previous innovations.” That resonates with where companies are adding roles. Product and design has 97 open, and infrastructure has 418 open with 184 opened in 30 days, reflecting investment in getting systems into production and keeping them reliable. Simon Willison’s observation that the new speech models are accessible to developers — “The implementation uses no libraries.” — underscores why infra and product roles matter.

Laurie Voss, quoted by Willison, sketches 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. What's left of making software is finding out what people actually want, defining it precisely, and making it pleasant to use. That cost is per piece of software and doesn't transfer, so as the amount of software goes to infinity, which it will because there's no ceiling on demand, that cost becomes the whole job.” Our listings show steady demand for data, product, and infra talent that fits this picture.

Bottom line

Commentators argued for pacing, warned of fear contagion, and pointed to secretive policy moves. Our data shows companies still hiring aggressively for capability work, with safety hiring positive but small in absolute terms. Until we see evaluation and safety roles rise much closer to data and engineering, the labor market read is speed, not slowdown.

What we read

Every quote above is taken verbatim from one of these posts.

More on this

Today's briefing