Skip to main content
AI Jobs·Report

ANALYSISReported by AI Jobs Report

Labs urged to slow, but hiring favors build: 1,760 eng vs 46 safety

Amid calls to “I think the answer is we have to shut the labs down” and a push for safety, our tracker shows employers adding engineering and data roles far faster than evaluation and safety.

Read the original at AI Jobs Report
9 min read28 viewsBy AI Jobs Report

Amid calls to “I think the answer is we have to shut the labs down” and a push for safety, our tracker shows employers adding engineering and data roles far faster than evaluation and safety.

The split-screen week: shut it down vs ship it

Gary Marcus led with a stark line he attributes to Jensen Huang: "I think the answer is we have to shut the labs down". He sharpened that into a principle of his own: "If a company can’t control its software, we should shut it down." In another post, Marcus said global leaders at the UN Security Council converged on five safety planks: "The need for preflight safety testing." "The need for transparency auditing." "The need for liability." "The need for global dialogue and international cooperation." "The need for immediate action."

At the same time, Zvi Mowshowitz focused on a rapid drumbeat of capability upgrades. On Anthropic’s latest, he wrote: "Introducing the world’s most powerful model, at least by some measures like Artificial Analysis or any standard benchmark list, which is now Claude Opus 5.5. Anthropic is claiming Opus 5.5 is outright as good or better than Fable 5.1, while being actively cheaper than Opus 5. That means it’s time for a good old system card reading." He added a note of caution on early takes: "The quick feedback from the internet is that Opus 5.5 is very good. I need more time before I am willing to offer comment."

Simon Willison highlighted a concrete drop in cost and friction for synthetic media: "Google released two new Gemini text-to-speech models today - gemini-3.8-flash-tts and gemini-3.8-flash-lite-tts." He reported hands-on performance and price: "It took ~20 seconds to generate 1m 18s of audio using Gemini 3.8 Flash TTS (not the cheaper Flash-Lite), at a cost of 2.74 cents."

Ethan Mollick captured the pace: "Stuff is happening quite fast." He urged attention beyond benchmark proofs, writing: "All of this effort from the AI labs pouring into proofs, but there are so many other interesting problems in other fields" and showed hands-on creation in public, from a zine experiment ("Not bad. I appreciate it mocking my prompt & itself.") to game prototyping ("There is an abandonware game I loved as a kid called Rescue Raiders. I asked Claude to create a modern & updated version of the game with new graphics, goals, tech trees, etc. It iterated back-and-forth with critic & art agents until I got this. Its fun!").

Alex Hanna pulled in the labor lens: "“AI” IS impacting labor, but not in the ways you think. In some cases, people generating “workslop” are making their colleagues’ jobs more difficult."

So the week’s views line up as pressure to slow and test, a flood of new capability, creator-grade tools becoming cheaper, and mixed accounts of workplace impact. Here is what the jobs signal says.

What our tracker shows now

Our tracker counts 4,409 open AI roles across 178 employers, taken from employers’ own job feeds.

By role family:

  • Modelling and engineering: 1,760 open, 1,155 opened and 698 closed in 30 days, across 143 employers
  • Data: 760 open, 540 opened and 349 closed in 30 days, across 125 employers
  • Infrastructure: 476 open, 238 opened and 81 closed in 30 days, across 97 employers
  • Research: 236 open, 69 opened and 23 closed in 30 days, across 50 employers
  • Product and design: 103 open, 54 opened and 21 closed in 30 days, across 49 employers
  • Evaluation and safety: 46 open, 16 opened and 3 closed in 30 days, across 20 employers

Top employers:

  • Accenture: 743 open, 848 opened in 30 days
  • Capital One: 205 open, 245 opened in 30 days
  • OpenAI: 176 open, 51 opened in 30 days
  • Amgen: 156 open, 46 opened in 30 days
  • Anthropic: 132 open, 36 opened in 30 days
  • PwC: 103 open, 72 opened in 30 days
  • Waymo: 102 open, 31 opened in 30 days
  • Databricks: 79 open, 28 opened in 30 days

Safety rhetoric vs safety hiring

Marcus amplified a consensus list that begins with "The need for preflight safety testing" and ends on "The need for immediate action." Our hiring data shows movement on evaluation and safety, but on a very different scale to build-out work. There are 46 open roles in evaluation and safety, against 1,760 in modelling and engineering, and 760 in data. Only 16 evaluation and safety roles were opened in the last 30 days, across 20 employers, compared to 1,155 modelling and engineering openings across 143 employers.

That mismatch is the clearest divergence between commentary and our tracker this week. While more leaders are talking about audits, liability, and testing, employers are still prioritizing core model engineering, data pipelines, and production infrastructure by orders of magnitude. Even among the frontier labs named in the posts, the openings skew toward growth rather than pause: OpenAI lists 176 open roles and Anthropic 132, with 51 and 36 opened in 30 days respectively.

Cheaper tools, broader build-out

Willison’s concrete price/time datapoint on TTS ("It took ~20 seconds... at a cost of 2.74 cents.") captures why enterprise demand keeps tilting toward integration. When a minute of voiced audio costs pennies and a few seconds to produce, teams will need engineers to wire it into products, data professionals to govern inputs and outputs, and infrastructure hires to scale it.

Our role-family mix aligns with that pattern. Infrastructure roles stand at 476 open, with 238 opened in 30 days. Data roles are at 760 open, with 540 opened in 30 days. Product and design, at 103 open roles, is a smaller slice but still active.

Who is hiring most also speaks to diffusion. Accenture’s 743 open roles (with 848 opened in 30 days) and PwC’s 103 open (72 opened in 30 days) point to client implementation work driving volumes. Capital One’s 205 open (245 opened in 30 days) signal continued adoption in financial services. This is consistent with Mollick’s point that the field’s impact now reaches beyond lab benchmarks into varied domains, even if he framed that as a missed opportunity by labs: "All of this effort from the AI labs pouring into proofs, but there are so many other interesting problems in other fields".

The “workslop” complaint and what hiring can and cannot say

Hanna argues that "people generating “workslop” are making their colleagues’ jobs more difficult." Our listings cannot tell us whether a given workflow creates rework on the ground. They do, however, show that companies are investing heavily in the job families tasked with making these systems usable and reliable. The 760 data roles and 476 infrastructure roles, plus 103 product and design roles, are the very categories organizations expand when they are trying to reduce churn from low-quality outputs and harden pipelines.

We do not see evidence in the postings of a broad retrenchment prompted by quality concerns. If anything, the opened-in-30-days figures show an acceleration in the last month toward engineering and data hires.

Speed, pressure, and the absence of a pause

Mollick’s "Stuff is happening quite fast" is reflected in the open-and-opened counts. Modelling and engineering saw 1,155 roles opened in the last 30 days. Data saw 540. Infrastructure saw 238. Accenture alone opened 848 roles in that period. Even the labs at the center of debates about shutdowns or moratoriums are hiring, not pausing: OpenAI opened 51 roles in 30 days, Anthropic 36.

Zvi Mowshowitz’s framing of Opus 5.5 as cheaper and stronger, and Willison’s demonstration of near-instant, cents-per-minute TTS generation, also speak to why hiring is not slowing. Lower unit costs push adoption; adoption demands build-out.

Our headline indicator and how to read this moment

We track jobs because they are a leading indicator of where organizations are actually placing bets. Our headline jobs-created indicator is counted from job listings and calibrated to the World Economic Forum’s Future of Jobs Report 2025 (11M AI roles created, 9M displaced, by 2030). This week’s mix of postings says employers are still predominantly investing in building and deploying systems, with some but comparatively limited investment in evaluation and safety.

Marcus’s checklist from the UN briefing is unlikely to materialize at scale until we see the evaluation and safety category grow in line with the engineering and data hiring we are counting now. Until then, the market signal is clear: the center of gravity is on shipping and integrating, not pausing.

Bottom line

  • Calls to halt or slow development are getting louder. Marcus quotes a bright line: "If a company can’t control its software, we should shut it down." He also reports broad agreement on "The need for preflight safety testing" and "The need for immediate action."
  • Capability and cost curves continue to improve. Zvi highlights a cheaper, stronger frontier model. Willison shows voice synthesis in seconds for pennies.
  • Employers are voting with requisitions. 1,760 modelling and engineering roles are open versus 46 in evaluation and safety. 1,155 modelling and engineering roles were opened in the last 30 days.
  • Diffusion is led by integrators and regulated industries. Accenture and Capital One top new openings.

If safety hiring and governance roles rise to meet the rhetoric, we will see it first in our evaluation and safety counts. For now, the hiring reality supports Mollick’s summary: "Stuff is happening quite fast."

What we read

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

](https://bsky.app/profile/emollick.bsky.social/post/3mw7g7v3uuc2z), I propose this as the official replacement for the famous (and now sat, I just got the first copies of my new book, Co-Existence (out October , [Stuff is happening quite fast.

When asked in September 2025 (yes, 202](https://bsky.app/profile/emollick.bsky.social/post/3mwa2mjely225)

More on this

Today's briefing