Commentators saw safety consensus and a price war; our tracker shows engineering and data roles dominate while evaluation and safety hiring remains scarce.
Safety consensus in the headlines, but not yet in headcount
Zvi Mowshowitz opened the week with a stark line: “This was the month the world took notice that AI might kill everyone.” In another post, Gary Marcus celebrated a rare alignment of voices at the UN Security Council, writing that speakers converged on “The need for preflight safety testing.” Along with that, he highlighted “The need for transparency auditing.” “The need for liability.” “The need for global dialogue and international cooperation.” and “The need for immediate action.”
Zvi added that “Politicians on both sides took notice and were saying sensible things.”
Our hiring tracker does not yet reflect a surge in the specific kind of safety and evaluation work those posts call for. Out of 4,371 open AI roles across 178 employers we track, only 48 are in evaluation and safety, with 18 opened and 3 closed in the last 30 days across 21 employers. By contrast, the role families tied to building and deploying systems are far larger: modelling and engineering has 1,735 open roles (1,104 opened and 686 closed in 30 days) across 143 employers; data has 783 open (537 opened and 325 closed) across 126; infrastructure has 467 open (239 opened and 82 closed) across 97.
The commentary says safety is finally center stage. Our numbers say hiring is still mostly about shipping, not evaluating. That is not a judgment on whether safety should scale faster, only a record that it has not yet done so in staffing.
Cheaper, stronger models and the rush to build
If safety talk is up, so is the capabilities race. Zvi’s system-card read starts: “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.” Simon Willison put the economic shift plainly: “GPT-6 Sol and Luna are half the price of their GPT-5.6 equivalents” and later, “It’s hard to overstate how competitive this pricing is.” He also showed what lower latency and cost mean in practice on the application side: “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.”
This chorus of stronger-for-less aligns closely with what employers are hiring for. The heaviest demand is for the teams that can take advantage of cheaper inference and faster iteration. In the last 30 days alone, modelling and engineering roles saw 1,104 openings and data roles saw 537. Infrastructure roles added 239 openings. Product and design remains a smaller slice at 104 open roles (54 opened and 20 closed) across 49 employers, suggesting that many organizations are still laying the technical foundations rather than scaling polished user experiences.
The top employers in our tracker also point to a deployment wave more than a pause for reflection. Accenture has 678 open AI roles and opened 785 in the last 30 days. Capital One has 208 open and 235 opened in 30 days. OpenAI has 172 open roles and 48 opened; Anthropic has 134 open and 38 opened. PwC has 103 open and 72 opened. Waymo sits at 103 open and 31 opened. Databricks has 79 open and 29 opened. Those mixes, especially the outsized appetite at consultancies and a major bank, match a world where model prices fall and organizations race to wire them into operations.
On this point, the commentary and our data agree. The pricing pressure Simon describes and the performance drumbeat Zvi notes are being met by hiring for the people who implement, integrate, and run these systems at scale, not a pivot to pause.
Agents, hidden humans, and what hiring does not show (yet)
Gary Marcus used Meta’s agent news to caution against hype, starting with: “For the last few days, all the buzz has been about Meta’s new agent, Muse:” and ending his look back at Facebook M with “It was all more or less a ruse.” His point is that agent demos have a history of concealing human labor and brittle scaffolding. That is a hiring story too, but not one our categories capture directly. We do not break out human-in-the-loop operations as a role family, so we cannot confirm from the tracker whether the new agent wave is staffing up that layer.
Simon Willison is watching the agent space from the builder’s side. He promoted an upcoming “Birds of a Feather Session on Agentic Engineering,” inviting people “building weird and interesting things with and on top of coding agents.” That grassroots energy is consistent with our finding that infrastructure and engineering roles dominate. If agents are to become robust products, the current hiring emphasis on infrastructure and core engineering is a necessary precondition. Whether it also leads to a back-office of human handlers, as Marcus worries, will show up later in customer support, operations, or trust-and-safety postings. Those are not visible at scale in our AI-specific counts today.
China, uncertainty, and the limits of our data
Nathan Lambert framed his conversation with JS Denain by saying, “A big takeaway of this podcast is how JS and I both have so much uncertainty with exactly where we are heading, and this was our best effort at stating our observations today.” Among their topics was “What Chinese job postings reveal about their labs.” Our tracker does not include a breakdown that would let us test that specific point. We track 178 employers and list top-line global firms by name as they appear in their own job feeds, but we do not slice by geography or publish a dedicated view on Chinese labs in this dataset. On the US side, the demand signal is clear: the largest share of openings sits with engineering, data, and infrastructure teams, and the employers leading in volume are global consultancies, a major US bank, and US-based AI labs and platforms.
Work quality, arts, and whether adoption slows
Alex Hanna posted that ““AI” IS impacting labor, but not in the ways you think. In some cases, people generating “workslop” are making their colleagues’ jobs more difficult.” That is a claim about quality and productivity effects on teams. Our hiring counts cannot validate or falsify the effect on coworkers. What they do show is that employers have not pulled back. In the last 30 days, we see net additions across the core build tracks: modelling and engineering opened 1,104 roles vs 686 closed, data opened 537 vs 325 closed, infrastructure opened 239 vs 82 closed. If poor-quality outputs were creating immediate drag large enough to slow adoption, we would expect to see a sharper contraction or pause in those categories. We do not see that in this snapshot.
Ethan Mollick, meanwhile, emphasized creative and educational uses. “There is an abandonware game I loved as a kid called Rescue Raiders.” He described having Claude iterate with agents to produce a modern version and reported that “Its fun!” In a second post he added, “There are opportunities to use AI as a bridge to arts appreciation.” That squares with the small but real product and design footprint in our data, as organizations explore experiences beyond system plumbing. Still, with only 104 open roles in product and design across 49 employers, the hiring tilt remains technical rather than curatorial or community-building.
Noah Smith connected the current public debate to a longer Bay Area current: “Much to my amusement, the world seems to have discovered Effective Altruism.” He noted that many people who now work at AI companies, “especially Anthropic,” were shaped by those conversations. Our tracker shows Anthropic with 134 open roles and 38 opened in 30 days. That demand is consistent with an expanding lab, but it does not, on its own, assign motive. Whether the EA conversation translates into more safety headcount will be visible if evaluation and safety roles grow from today’s 48 open across 21 employers.
Where commentary and hiring diverge, and what to watch next
Across these posts, two narratives run in parallel. One calls for a step change in safety practice and governance. The other celebrates a step change in capability and cost. Our hiring data says the latter is resourced far more heavily than the former right now. Evaluation and safety roles are a thin slice of the market. Engineering, data, and infrastructure dominate openings and new postings. The biggest employers by volume are implementers and deployers.
Where claims align with the numbers, we should expect more of the same: cheaper, stronger models and builder energy are showing up in hiring. Where they do not, the gap is instructive: public safety consensus has not yet become widespread safety staffing. If that changes, we will see it in a sustained rise from today’s 48 open evaluation and safety roles and a broader spread of employers adding those positions alongside the build teams.
Until then, the market is voting for builders.
What we read
Every quote above is taken verbatim from one of these posts.
- Nathan Lambert: Debating RSI, the US-China Gap, and Jaggedness with JS Denain of Epoch
- Zvi Mowshowitz: Politics Gets Interested In Those Trying Not To Die, Claude Opus 5.5: The System Card
- Simon Willison: Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war, llm-anthropic 0.29, llm-typesafe 0.1a0, Gemini 3.8 TTS Playground, SF October 14th: A Birds of a Feather Session on Agentic Engineering
- Noah Smith: The problem(s) with utilitarianism
- Gary Marcus: The secret behind Meta’s Muse, Historic UN Security Council Briefing on AI
- Alex Hanna: 'AI' IS impacting labor, but not in the ways you think. In some cases,
- Ethan Mollick: [There is an abandonware game I loved as a kid called Rescue Raiders.
](https://bsky.app/profile/emollick.bsky.social/post/3mw7g7v3uuc2z), For all the tension between AI and the arts, I have received enthusias