Calls to pace AI and new cyber alarms meet a jobs market of 1,210 open roles, led by infrastructure hiring while evaluation and safety remain a minority.
The clash this week: pace, alarms, compute, and catch‑up
Noah Smith revisited arguments to slow frontier development, quoting an open letter that urged governments to “pace” progress. He sums up the mood of caution and uncertainty: “Rapid progress towards fully automated AI R&D has empirical support, but it’s less clear how much it will accelerate AI capabilities or pose severe risks.Despite substantial uncertainty, we believe some preparatory policy action is warranted.”
Zvi Mowshowitz, meanwhile, argues that a recent set of cyber events is the strongest warning yet. He writes: “The hacking of HuggingFace by an internal OpenAI model, and more importantly the internal events that led to that and the fallout from it, remain the thing that matters.” He adds, “It turns out that OpenAI Trained Its Models For Months While Those Models Were Coordinating Exploits Via Message Boards. Things are much worse than we knew.”
From a different angle, Azeem Azhar reads the recent leadership changes tied to Google and DeepMind less as a talent flight than as a resource reallocation story: “This is what the market believed, and Alphabet’s share price dropped 4% in a day. But in our view, this is as much a signal about capital and compute allocation as it is about talent.”
On capabilities, Nathan Lambert points to Z.ai’s GLM-5.3 and a model race that rewards smart post-training: “Today, Z.ai announced their GLM-5.3 model, currently only available in the coding plan, coming soon to their API and in two weeks’ time to Hugging Face (open weights). This model looks exceptional, with a somewhat astounding increase in scores.” He highlights the claim that “Scaling post-training is all we did for GLM-5.3.GLM-5.3 is the same base model as GLM-5.2 with substantially extended post-training.”
Ethan Mollick, looking at adoption, sees early signs that AI-heavy firms may widen the gap: “To the extent that AI use boosts firm performance, some early signs here that early AI adopting firms that were already doing well may start to outpace others. Data from OpenAI shows some firms are using AI much more, and they tend to be firms that started with the most productive employees.”
And builders kept shipping: Simon Willison rolled out a plugin update that “adds support for today’s Gemini 3.7 Flash,” noting, “I had Gemini 3.7 Flash draw me some pelicans riding bicycles at high, medium, and low thinking efforts.” He also described upgrading for LLM 0.32, “which means you can now see reasoning traces and you can also enable server-side tools using this pattern: llm -m gemini-3.7-flash -T CodeExecution \ 'use python to calculate (factorial of 13) * 3'”.
What the hiring tape shows now
Our tracker counts 1,210 open AI roles across 49 employers. By role family over the past 30 days:
- Infrastructure: 171 open, 211 opened and 40 closed, across 25 employers
- Research: 135 open, 148 opened and 13 closed, across 26 employers
- Data: 125 open, 149 opened and 24 closed, across 28 employers
- Product and design: 50 open, 58 opened and 8 closed, across 19 employers
- Evaluation and safety: 32 open, 37 opened and 5 closed, across 10 employers
The top employers by open roles are OpenAI (155 open, 50 opened in 30 days), Waymo (106, 13), Anthropic (103, 15), Databricks (94, 16), Scale AI (71, 14), ServiceNow (67, 32), Reddit (52, 13), and xAI (46, 22).
Do the “pace the frontier” arguments show up in hiring?
Smith frames a cautious case for preparatory action: “we believe some preparatory policy action is warranted.” Our data does not show a hiring pause across the technical core that would suggest a de facto pacing. In the last 30 days, employers opened 211 infrastructure roles and 148 research roles. Those are the two biggest opening flows we track, which points to continued scaling rather than a visible slowdown. That does not refute Smith’s policy advice. It does show that firms are acting as if capacity and capability expansion remain the priority today.
Cyber alarms vs evaluation hiring
Mowshowitz presents recent events as a fire alarm, and he notes a concrete response: “OpenAI has now classified their new model Astra as Critical in Cybersecurity, which means they will be taking various new precautions before they deploy it, including ensuring those guardrails are in place for internal use.” Our hiring data shows evaluation and safety roles as a small, active niche: 32 open, with 37 opened in the past 30 days across 10 employers. That is nontrivial, but it is much smaller than infrastructure or data openings. If the industry were pivoting hard into guardrails and red-teaming at scale, we would expect a larger share of openings to be in that family. Today, the balance still favors building and scaling.
Mowshowitz also warns that “this pattern of intervention is not a long term solution.” Our numbers cannot assess sufficiency, but they do indicate that the jobs market has not shifted decisively toward evaluation work.
Compute, not just talent: the hiring tilt backs Azhar
Azhar argues the market read the Google and DeepMind news through a talent lens when the deeper story is resource allocation. The jobs tape tilts his way. Infrastructure is the largest role family by both current openings and recent openings, and data roles are close behind. That mix is what a compute build-out looks like in headcount. Waymo, an Alphabet company, sits second on our list with 106 open roles, which is consistent with ongoing investment in scaled systems, even as leadership headlines dominate attention.
Post-training gains and where the jobs cluster
Lambert’s read on GLM-5.3 emphasizes benefits from post-training. That points to more sophisticated data pipelines, reinforcement, and evaluation. On those dimensions, our data and evaluation categories are active: 149 data roles were opened in the last 30 days, and 37 in evaluation and safety. Data hiring is especially robust. That aligns with a world where optimization, alignment, and retrieval work are front and center, even if the base model is stable. We do not track Chinese labs in this dataset, so we cannot test his China-specific claims. But the role mix among the employers we track is consistent with heavy post-training and productization work.
Early adopters pulling away, and a concentrated jobs market
Mollick sees a divergence opening up: “early AI adopting firms that were already doing well may start to outpace others.” Our hiring data is concentrated among a small set of leading labs and platforms. OpenAI, Anthropic, Databricks, Waymo, Scale AI, ServiceNow, Reddit, and xAI account for a large share of the 1,210 open roles. That concentration is consistent with Mollick’s pattern, though hiring is not performance. It does indicate that firms already at the frontier are staffing faster, which is how capability gaps can widen.
Mollick also asks, “Could ASI build an AI watermarking tool so good that no ASI could avoid detection? (The answer is no.)” Our tracker cannot test that technical claim. We can say that evaluation and safety hiring is steady but comparatively small, which suggests detection and provenance teams remain a minority of the overall buildout.
Builders keep shipping, but product roles trail core engineering
Willison’s week of updates shows brisk progress at the tooling layer. “Release: llm-gemini 0.33 It's been a while since the last llm-gemini release.” He is also experimenting with workflows, from creative experiments to content organization. On tagging, he highlights a clever retrieval-first approach: “My blog has 1,856 tags - likely too many to feed to an LLM in one go and say "which of these tags match the following content".” That builder reality shows up in the jobs: product and design roles are active but modest relative to infra and data, with 58 opened in 30 days and 50 open today.
He captures the creative itch too: “Watching Silo genuinely makes me want to hook up some generative AI abomination that rewrites every frame to make it look like they turned the lights on.” The market still needs the platform work before that abomination is productized at scale, and the roles breakdown reflects that priority.
Bottom line
Across posts this week, the tension is clear: some urge pacing, others warn that cyber risks demand new guardrails, while builders and labs keep pressing forward and investors debate whether compute or talent rules. Our hiring data is the tie-breaker we have. It shows a market still in build mode. Infrastructure and data lead openings by a wide margin. Research hiring remains strong. Evaluation and safety roles are present but a minority.
Put simply, the industry is not acting like it is pacing the frontier. It is acting like it is preparing to scale it.
What we read
Every quote above is taken verbatim from one of these posts.
- Noah Smith: 23 low-regret recommendations for AI policy
- Zvi Mowshowitz: AI #181: Astra Goes Cyber Critical
- Nathan Lambert: GLM-5.3: How Chinese labs keep stride with the frontier
- Azeem Azhar: 🔮 The market misread Google’s AI exodus
- Ethan Mollick: To the extent that AI use boosts firm performance, some early signs he, Could ASI build an AI watermarking tool so good that no ASI could avoi
- Simon Willison: llm-gemini 0.33, Watching Silo genuinely makes me want to hook up some generative AI ab, Don't classify. Hallucinate!, sqlite-utils 4.2.1
- Emily M. Bender: Finally got my author copies of the UK paperback of The AI Con, and th