Skip to main content
AI Jobs·Report

ANALYSISReported by AI Jobs Report

Safety talk surges; just 48 safety roles in our tracker

Posts this week pressed for AI safety and celebrated cheaper models; our hiring tracker shows most demand is still for engineers and services, with only 48 safety roles open across 178 employers.

Read the original at AI Jobs Report
9 min read22 viewsBy AI Jobs Report

Posts this week pressed for AI safety and celebrated cheaper models; our hiring tracker shows most demand is still for engineers and services, with only 48 safety roles open across 178 employers.

The week’s split-screen: rising risk rhetoric, faster and cheaper models

Zvi Mowshowitz declares, "This was the month the world took notice that AI might kill everyone." Gary Marcus, speaking at the UN, rejects extinction talk but warns of near-term harms: "What we are facing near-term, anyway, is not extinction. And it’s not superintelligence… It’s wholesale deepfaked disinformation, and unreliable but persistent AI systems stealing credentials and launching cyberattacks, at scale." Nathan Lambert, after a long debate with JS Denain of Epoch AI, offers a humbler stance: "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."

At the same time, Simon Willison tracks a fresh price war and new model shapes. On Jev, a "System One" or decision model, he quotes TypeSafe’s positioning: "Think of Jev as a frontier-intelligence function call: unstructured state in, typed probabilistic decisions out." He also notes industry-wide cuts: "It's going to take a while to get a good read on all of these new models, but here are my impressions so far." And about the latest prices, "It's hard to overstate how competitive this pricing is."

Our tracker lets us test which narratives are showing up in hiring. Right now, we count 4,433 open AI roles across 178 employers, taken from their own job feeds. Most demand is concentrated in builders. Evaluation and safety roles are thin.

What the hiring shows now

By role family in our data:

  • Modelling and engineering: 1,767 open, 1,081 opened and 645 closed in 30 days, across 142 employers
  • Data: 786 open, 541 opened and 310 closed in 30 days, across 128 employers
  • Infrastructure: 459 open, 231 opened and 77 closed in 30 days, across 95 employers
  • Research: 235 open, 69 opened and 21 closed in 30 days, across 51 employers
  • Product and design: 105 open, 57 opened and 17 closed in 30 days, across 49 employers
  • Evaluation and safety: 48 open, 18 opened and 3 closed in 30 days, across 21 employers

The top employers by open roles skew toward integrators and regulated industries, not just frontier labs:

  • Accenture: 735 open, 777 opened in 30 days
  • Capital One: 206 open, 217 opened in 30 days
  • OpenAI: 170 open, 49 opened in 30 days
  • Amgen: 163 open, 45 opened in 30 days
  • Anthropic: 131 open, 34 opened in 30 days
  • PwC: 103 open, 91 opened in 30 days
  • Waymo: 102 open, 29 opened in 30 days
  • Databricks: 79 open, 29 opened in 30 days

Safety fears vs safety jobs

Mowshowitz sees the policy spotlight intensifying: "You may not be interested in politics. But when you want to save the world, or change it, and you start to get traction, politics is going to get interested in you." Marcus argues "Hope is not a strategy," and calls out practical threats that should motivate action now.

Our hiring numbers do not reflect a surge in safety capacity yet. Evaluation and safety has 48 open roles across 21 employers. That is a small share of the 4,433 openings we track and much smaller than modelling and engineering, which has 1,767 open roles. Even among labs in our top employers, OpenAI (170 open) and Anthropic (131 open) are not driving enough safety postings to shift the overall mix.

The commentary says safety should be front and center; our data says safety is a niche hiring category today. On this question, the rhetoric is ahead of the requisitions.

Open weights, new form factors, and where the jobs actually are

Lambert briefed Congress on open weights, writing: "Open language models are AI models where their weights are publicly available for inspection or downstream use." He points to popular open-weight families and the licensing regimes that govern them: "These models are governed by licenses, governing documents dictating what is allowed with downstream use, and are often accompanied by inference code in libraries such as Transformers, VLLM, SGLANG, etc."

Willison’s coverage highlights how the economics are shifting in ways that should enable more downstream deployments. Jev is "also very fast, and really cheap," with "the input price of their first model is $0.042 per million tokens - cheaper even than OpenAI's GPT-5 Nano ($0.05/million)." He also flags dramatic price cuts from the latest general-purpose models and concludes, "It's hard to overstate how competitive this pricing is."

If cheaper models and open weights were already flipping the jobs mix toward application roles, we would expect a surge in product and design postings. We do not see that yet: just 105 product and design roles are open, against 1,767 in modelling and engineering and 786 in data. The hiring signal says organizations are still building foundational plumbing, integrations, and model-facing engineering talent more than they are scaling product teams around new LLM capabilities.

That lines up with the employer list. Accenture with 735 open roles and PwC with 103 suggest companies are leaning on services partners to thread together the fast-moving model landscape into production stacks. The hiring is following the integration work.

Infrastructure keeps climbing

Willison also flags platform shifts that could expand who can deploy AI at the edge. On Cloudflare, he quotes the release claim that "Python is now a first-class, fully supported language on the Cloudflare Developer Platform" and describes how it runs via WebAssembly and Pyodide.

Our infrastructure category sits at 459 open roles, with 231 opened in the last 30 days across 95 employers. That is a notable middle tier between modelling/engineering and data. Combined with 79 open roles at Databricks and 102 at Waymo, we read this as continued investment in runtime, data pipelines, and on-device or near-user inference. The platform story shows up in the jobs.

Geopolitics and uncertainty: is hiring preserving options?

Jack Clark summarizes RAND’s advice for an uncertain path: "It’s not clear exactly what is optimal, so the US should spend money now to preserve options" and to "maintain a “Freedom of Action” strategy." Lambert’s framing echoes the uncertainty: "so much uncertainty with exactly where we are heading."

Our data is consistent with an options-preserving posture. Employers are hiring across six role families and 178 organizations. That is a distributed bet, not a concentrated push into any one narrative like agents or safety. We also note that Lambert and Denain planned to examine "What Chinese job postings reveal about their labs" as part of their discussion. Our tracker does not include a breakdown of Chinese lab postings, so we cannot test claims about a U.S.–China hiring gap. On geopolitics, our data is compatible with the keep-options-open thesis but cannot resolve cross-border comparisons.

Agents, realism, and the pace of productization

Marcus revisits a prior agent push and warns against premature narratives. On Facebook’s Project M, he writes: "It was all more or less a ruse." The reveal, "No more than 30% of requests served by AI!" is a reminder that agent claims can overrun operational reality.

Today’s model cost curves and tools are far better. Willison even shipped a plugin to tap Jev’s typed decisions: "I built this new plugin for LLM to add support for TypeSafe AI's new Jev model." But our hiring signal suggests employers remain cautious on productization. With only 105 product and design roles open, firms are not staffing massive new agent product lines yet. They are prioritizing engineering, data readiness, and infrastructure that could support those products later.

Corporate claims vs real headcount

Azeem Azhar observes a change in tone on earnings calls: "Companies are backing their AI claims with more numbers." His extract aims to surface "What corporate AI claims reveal – and leave unanswered" by "analyzing earnings calls from S&P 500 companies and the claims that they make."

Our tracker picks up where calls leave off. The biggest hiring footprints are in services and regulated verticals. Accenture’s 735 open roles and 777 opened in 30 days, PwC’s 103 and 91, Capital One’s 206 and 217, and Amgen’s 163 and 45 show demand is not confined to AI labs. It is enterprises and their integrators converting strategy talk into headcount for engineering, data, and infrastructure. Frontier labs are still hiring, but they are not dominating the job market: OpenAI has 170 open roles and 49 opened in 30 days; Anthropic has 131 and 34.

Where the posts meet the postings

  • Alignment: Clark’s and Lambert’s emphasis on uncertainty matches a diversified hiring portfolio across role families and employers.
  • Divergence: Mowshowitz’s and Marcus’s calls for safety and security do not yet show up as broad safety hiring. Only 48 evaluation and safety roles are open across 21 employers.
  • Emerging but not yet scaled: Willison’s pricing and model-shape coverage points to lower costs and new decision-oriented APIs. Our data shows builders being hired now, while product and design hiring remains modest. Integrators are leading the way.

The commentary this week is animated by risk and rapid change. The jobs are following the buildout. If and when the risk conversation turns into budget, we will expect to see it where we do not today: in a larger evaluation and safety category, and a bigger shift toward product headcount. Until then, the numbers say the center of gravity is still engineering, data, and infrastructure, not safety or splashy agent teams.

What we read

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

More on this

Today's briefing