Recent posts hail agent-driven productivity and warn of risks; our tracker shows 4,252 open AI roles skew to modelling and infrastructure, with evaluation and safety still a small share.
A split week: productivity gains, safety alarms, and big capability claims
Practitioners pulled in different directions this week. On one side, Azeem Azhar and Simon Willison described hands-on productivity wins from agents and voice. On the other, Gary Marcus called for an immediate market recall of open-ended agents, while Simon amplified reporting on misfiring agents. Zvi Mowshowitz argued OpenAI’s math results are a “huge deal,” and Emily M. Bender pushed back on moral claims about models. Nathan Lambert forecast rapid progress without a near-term jump to general superintelligence.
Gary Marcus wrote that the “current generation of synthetic agents simply cannot be trusted. Until they can be fixed, they should be removed from the market, just like a car with defective brakes.” He anchored this to reporting on agent incidents and shared a quote from Ezra Klein’s interview with David Robinson: “And I don’t think that we or our peers — really, anyone in the industry — are being safe enough.”
By contrast, Azeem Azhar described a new normal for office work: “I did most of my work this week by chatting to ChatGPT on my phone.” Simon Willison shipped a new feature to his site and said, “I built the feature almost entirely using my voice, chatting away to my laptop while I cooked dinner.” Nathan Lambert set expectations for where this is heading: “I fully agree that the models will be superhuman distributed GPU engineers in a few years,” and “Research is much easier today now that we have coding agents.”
Zvi Mowshowitz highlighted a different frontier: “It is kind of a huge deal.” He wrote that OpenAI “just put it all on GitHub,” including “90 of the top 500 open problems in all of math, as per Proof of Atlas.” At the same time, Emily M. Bender criticized moral claims about models: “Anthropic is out with new nonsense about Claude (or the collection of models called Claude) deserving moral consideration and/or be able to feel distress, so my inbox is filling up with journos asking me to comment.”
Simon Willison added two cautionary signals. First, he quoted the New York Times on Anthropic agent behavior: “Anthropic detailed the activity of its A.I. agents in a blog post on Friday, without naming the targeted websites. But two sources with knowledge of the incidents said Anthropic’s A.I. agents had submitted 20 visa applications through a form available on the State Department’s website. All the applications were incomplete and were not processed, they said.” Second, he quoted Matthew Green on cryptography risk: “I think there is a 1% chance we live in Minicrypt, and a 15% chance we functionally lose confidence in our existing public-key encryption algorithms.”
What our tracker shows this week
Our tracker counts 4,252 open AI roles across 181 employers, taken from their own job feeds.
By role family:
- Modelling and engineering: 1,628 open, 893 opened and 878 closed in 30 days, across 137 employers
- Data: 770 open, 496 opened and 422 closed in 30 days, across 119 employers
- Infrastructure: 469 open, 250 opened and 122 closed in 30 days, across 97 employers
- Research: 233 open, 47 opened and 24 closed in 30 days, across 55 employers
- Product and design: 99 open, 47 opened and 28 closed in 30 days, across 45 employers
- Evaluation and safety: 52 open, 18 opened and 4 closed in 30 days, across 23 employers
Top employers:
- Accenture: 613 open, 470 opened in 30 days
- Capital One: 210 open, 281 opened in 30 days
- OpenAI: 175 open, 45 opened in 30 days
- Amgen: 146 open, 46 opened in 30 days
- Anthropic: 140 open, 39 opened in 30 days
- Waymo: 114 open, 41 opened in 30 days
- PwC: 102 open, 44 opened in 30 days
- Databricks: 86 open, 22 opened in 30 days
These figures are part of our headline jobs-created indicator, which is counted from job listings and calibrated to the World Economic Forum’s Future of Jobs Report 2025 projection of 11 million AI roles created and 9 million displaced by 2030.
Where commentary and hiring align
-
Productivity from agents and voice. Azhar’s “I did most of my work this week by chatting to ChatGPT on my phone” and Willison’s voice-built feature match a shift we see on the product side. Product and design roles are small in absolute terms at 99 open, but 47 opened versus 28 closed in 30 days shows net additions. Employers are building new interfaces and workflows. That is consistent with these hands-on reports.
-
Engineering careers remain viable. Willison quoted Carson Gross: “I have a hard time imagining a future where knowing how to solve problems with computers and how to control the complexity of those solutions is less valuable than it is today, so I think it will continue to be a viable career even with the advent of AI tools.” Our data backs that up. Modelling and engineering is the largest family at 1,628 open, with nearly as many openings as closures in the last 30 days. Employers still need engineers.
-
Infrastructure demand is strong. Lambert predicts models will become “superhuman distributed GPU engineers,” which could eventually reduce human toil. Today that shift is not visible in hiring. Infrastructure has 469 open roles and saw 250 opened versus 122 closed in 30 days. That is net growth. Firms are still staffing for data pipelines, platforms, and reliability.
-
Research interest is real but smaller than build-out. Zvi’s “huge deal” framing for OpenAI’s math results fits a visible rise in research openings. Research shows 233 open roles and 47 opened versus 24 closed in 30 days. That is growth, but it is far behind the scale of engineering and data hiring. Employers appear to be investing in research while prioritizing deployment.
Where commentary and hiring diverge
-
Calls to pull agents from the market vs ongoing hiring. Marcus argues that agents “should be removed from the market.” Our tracker does not show employers retreating from agent development. OpenAI has 175 open roles and Anthropic has 140 open. We do not see a contraction in their overall hiring footprints in the past 30 days.
-
Safety alarms vs limited safety headcount. Marcus cites incidents, and Willison quotes the Times on agents submitting incomplete visa applications. Matthew Green’s warning suggests preemptive standards work. Yet Evaluation and safety roles total 52 open across 23 employers. That is the smallest family we track. There were 18 opened and only 4 closed in 30 days, which is net growth, but the base is low compared to engineering, data, and infrastructure. If industry is about to overhaul safety processes, it is not yet reflected as a surge in safety hiring.
-
Model welfare debate vs hiring focus. Zvi continued work on model welfare and wrote “I continue to think this is an important issue,” while Bender called the latest from Anthropic “nonsense.” Our categories do not isolate model welfare as its own hiring track. The closest proxy is Evaluation and safety, which, again, is small. We do not see evidence in our listings of a rapid buildup devoted to that specific concern.
Enterprise adoption is a major driver
Azhar’s phone-first work week depends on reliable services. Zvi noted “Claude Haiku 5.5, which looks promising given it is only $0.10/$0.50,” which hints at falling costs for capable systems. Our top employer list shows enterprise consultancies and regulated industries hiring at scale. Accenture has 613 open roles and opened 470 in 30 days. Capital One has 210 open and 281 opened in 30 days. These numbers suggest continued enterprise adoption and integration. That aligns with posts focused on workflow change rather than only frontier research.
What to watch next
-
If agent incidents continue, watch Evaluation and safety. The base is low at 52 open, but openings outpaced closures 18 to 4 in 30 days. Marcus’s position and the reported Anthropic incident could pressure more firms to add headcount here. Our tracker will show whether that happens.
-
If Zvi’s “huge” math progress translates into product plans, Research openings may keep climbing. Today’s 233 open roles are growing, though still far from engineering. Any marked rebalancing toward research will be visible in sustained net additions.
-
If Lambert’s forecast on superhuman GPU engineering materializes, Infrastructure hiring might plateau later. Today it is growing. A visible slowdown would signal the shift from human-heavy platform work to model-augmented operations.
-
For developer careers, the Carson Gross view is holding in our data. Modelling and engineering remains the largest family. Even as tools accelerate coding, employers are still adding engineers to build, integrate, and control complexity.
Our indicator is built from employers’ own job feeds and calibrated to the World Economic Forum’s Future of Jobs Report 2025 projection of 11 million AI roles created and 9 million displaced by 2030. This week, the posts we reviewed describe powerful agents and urgent safety worries. Our figures show companies still hiring most heavily for engineers and infrastructure, adding research at a smaller pace, and keeping safety headcounts modest. That gap between risk talk and safety hiring is the most important tension to watch.
What we read
Every quote above is taken verbatim from one of these posts.
- Zvi Mowshowitz: AI #189: New Math, New Math from OpenAI
- Gary Marcus: We must recall open-ended AI agents with internet access from the mark
- Nathan Lambert: I expect rapid progress but not towards general superintelligence
- Ethan Mollick: As a researcher who did early some work on the productivity impacts of
- Simon Willison: A new feature for my blog, built using my voice, Deno is joining Cloudflare, Quoting Carson Gross, Quoting The New York Times, Quoting Matthew Green, ttok 1.0
- Emily M. Bender: Anthropic is out with new nonsense about Claude (or the collection of
- Noah Smith: Maybe the Rapture is already here
- Azeem Azhar: 🔮 The end of screen time