Posts from Zvi Mowshowitz, Gary Marcus, Ethan Mollick and others warn of safety chill, valuation wobble and on-device parity; our tracker shows steady demand led by modelling and engineering at major employers.
The week’s split-screen: crisis talk versus practical momentum
Several prominent voices painted a picture of AI in flux. Gary Marcus argued that liability and legal headwinds are deflating frontier-lab ambitions. Zvi Mowshowitz highlighted a safety blow-up at OpenAI. Ethan Mollick said on-device systems have eclipsed last year’s benchmark. Jack Clark spotlighted persistent reasoning gaps. Simon Willison shipped tools that make everyday AI workflows faster. And Noah Smith noted how AI can become a prop in political theater, writing: “These include demanding that artificial intelligence be renamed “artificial superintelligence”, bellowing that he should have won the Nobel Peace Prize, and conducting an execution on live TV.”
Against that background noise, our tracker counts 4,192 open AI roles across 181 employers taken from their own job feeds. The hiring signal is clearest in modelling and engineering (1,592 open). Evaluation and safety roles are smaller in absolute terms (51 open) but measurable and growing across 22 employers. Below, we test the week’s claims against those figures.
Valuation worries and liability vs actual hiring at frontier labs
Gary Marcus framed the funding mood bluntly: “LLM’s might well someday be renamed Large Liability Modules.” He added, “Against all of this, it’s harder and harder to fathom how the companies could really merit the enormous valuations they are seeking.”
If liability and valuation doubts were already choking off frontier hiring, we would expect to see a clear slowdown at the biggest labs. Our data does not show that. OpenAI has 174 open roles, with 40 opened in the last 30 days. Anthropic has 138 open roles, with 39 opened in the last 30 days. Those are among the highest recent-opening counts in our tracker. The broader market also shows sustained activity in modelling and engineering (1,592 open, 792 opened and 700 closed in 30 days across 135 employers).
Our read: Marcus is describing valuation sentiment and legal risk, not recruiting pipelines. On the hiring side, our numbers show frontier labs and their ecosystem are still adding roles at pace.
Safety chill: is it visible in the jobs data?
Zvi Mowshowitz reported: “Three OpenAI safety researchers, Tomek Korbak, Mikita Balesni and Jasmine Wang, were fired, ostensibly for ‘mishandling sensitive company information.’” He concluded, “OpenAI has three less good safety researchers. Trust will be lower. Speech will be chilled.”
We cannot adjudicate the case, but we can check whether evaluation and safety roles are shrinking or expanding in the aggregate. Our tracker shows 51 evaluation and safety roles open today, with 18 opened and 3 closed in the last 30 days, across 22 employers. That is a small but active category. The low closures suggest these roles are not being filled quickly, consistent with a niche talent pool or more cautious processes. Meanwhile, OpenAI’s overall hiring remains elevated (174 open; 40 opened in 30 days). On balance, our data does not show a broad pullback in safety and evaluation jobs. If anything, we see continued, measured demand.
On-device models “better than GPT-4” vs the composition of open roles
Ethan Mollick wrote: “GPT-4 was the undisputed best model in the world for a full year and we all debated whether anyone could catch up or if OpenAI had some secret formula that nobody else knew about. Now there are are 19 AI models that can run on a modern smartphone that are better than GPT-4!”
If on-device parity were collapsing employer demand for core model work in favor of pure application layers, we would expect product and design to dominate new postings. Our tracker shows the opposite. Modelling and engineering has 1,592 open roles (792 opened, 700 closed in 30 days), while product and design sits at 96 open (44 opened and 28 closed in 30 days). Infrastructure roles are also strong at 465 open (225 opened and 120 closed in 30 days across 96 employers). The center of gravity in hiring still leans toward core technical roles, not primarily end-user features and packaging.
Mollick’s claim points to capability democratization. Our data says that, if this is happening, employers are still investing heavily in engineering and infra to harness it at scale.
Visual reasoning gaps and the demand for research and eval
Jack Clark highlighted a fresh benchmark stressor: “New benchmark shows how AI models lack intuitive visual reasoning:” and “Humanity’s Sixth Sense highlights some shortcomings of current AI systems” with “AI startups Elorian and Scale AI have built and released Humanity’s Sixth Sense, a benchmark that “measures how a model interprets and reasons through information a visual scene implies but doesn’t explicitly show.” HSS contains 522 questions that cover 288 images and 234 short video clips (totalling 17.6 hours of video).”
Those gaps line up with hiring for research and evaluation. Research roles sit at 234 open, with 46 opened and 23 closed in the last 30 days across 55 employers. Evaluation and safety, while smaller, shows active openings (51 open; 18 opened in 30 days). Clark’s observation about persistent weaknesses is consistent with our data: employers continue to fund research and eval to close those capability gaps.
Developer ergonomics and the quiet enablers of scale
Simon Willison’s week was about shipping. “Python 3.15 is here!” he wrote, adding “My LLM CLI tool implements plugins.” He also released a small web utility: “I had Codex and GPT-6 Astra Extra High build this Python web app which provides a web UI for viewing and editing files in a folder.” His excitement about lazy imports was practical: “Lazy imports directly address this startup cost.”
This focus on performance tuning and operational ergonomics shows up in our job families. Infrastructure has 465 open roles, with 225 opened in 30 days. Data roles are at 756 open (460 opened and 387 closed in 30 days across 119 employers). These are the functions that make model deployments reliable, fast and compliant in real settings. The continued hiring here aligns with Willison’s developer-centric lens: optimization and integration work is not optional, it is being staffed at scale.
Health tech eclipse narratives vs cross-sector AI hiring
Azeem Azhar argues the spotlight may have shifted: “In August 2024, we said GLP-1s “will be the most impactful technology in the US over the next two years, with more meaningful and measurable impacts than either artificial intelligence or renewable technologies.” He also wrote, “By the end of 2025, GLP-1 sales exceeded generative AI revenue, reaching $76 billion, versus our $63 billion for generative AI.”
We do not track GLP-1 jobs and cannot test revenue comparisons. What we can say is that AI hiring remains broad-based. Accenture leads our employer table with 608 open roles and 390 opened in the last 30 days. Capital One has 201 open and 256 opened in the last 30 days. OpenAI sits at 174 open; Anthropic at 138; Waymo at 115; PwC at 103; Databricks at 86. Notably, life sciences is represented too, with Amgen at 147 open and 38 opened in 30 days. Even if GLP-1 revenue currently outpaces generative AI revenue, employers across services, finance, pharma and core AI are adding AI positions.
Self-improvement and the supervision bottleneck
David Ha highlighted a research direction: “New Sakana AI paper: Recursive Self-Improvement through Multi-Agent Self-Supervision (MASS).” He asked, “How do we keep improving AI when tasks become too complex for humans to reliably evaluate? The model eventually hits a supervision bottleneck.”
Our numbers suggest employers are resourcing around that bottleneck. Research openings total 234, and evaluation and safety 51, with recent openings in both. That hiring pattern is consistent with the need to design scalable eval methods and guardrails as models take on tasks that are harder for humans to score directly.
Politics and perception vs labor market reality
Noah Smith’s political framing includes AI as a stage prop: “These include demanding that artificial intelligence be renamed “artificial superintelligence”, bellowing that he should have won the Nobel Peace Prize, and conducting an execution on live TV.” Whatever the impact of such rhetoric on public sentiment, we do not see it reflected in employer demand. Across role families, we count:
- Modelling and engineering: 1,592 open, 792 opened and 700 closed in 30 days, across 135 employers
- Data: 756 open, 460 opened and 387 closed in 30 days, across 119 employers
- Infrastructure: 465 open, 225 opened and 120 closed in 30 days, across 96 employers
- Research: 234 open, 46 opened and 23 closed in 30 days, across 55 employers
- Product and design: 96 open, 44 opened and 28 closed in 30 days, across 44 employers
- Evaluation and safety: 51 open, 18 opened and 3 closed in 30 days, across 22 employers
This is a hiring market with breadth and churn. The additions and closures indicate active pipelines rather than stasis.
How our indicator is calibrated
Our headline jobs-created indicator is counted from employer listings and calibrated to the World Economic Forum’s Future of Jobs Report 2025 (11M AI roles created, 9M displaced, by 2030). That calibration anchors our interpretation of the flow of new roles without presenting it as a raw count divorced from macro context.
Bottom line: where commentary meets the numbers
- Marcus’s liability and valuation skepticism does not yet show up as a hiring freeze. OpenAI and Anthropic continue to open many roles.
- Mowshowitz’s account of a safety chill at one lab contrasts with ongoing openings in evaluation and safety across 22 employers, and strong overall hiring at the same lab.
- Mollick’s on-device leap claim has not shifted hiring away from core engineering; modelling and infra remain the largest pools.
- Clark’s benchmark points to real capability gaps; employers are staffing research and eval to close them.
- Willison’s developer improvements mirror robust infrastructure and data hiring.
- Azhar’s GLP-1 eclipse narrative may hold for revenue, but employer behavior in AI remains expansionary across sectors.
The conversation this week is about turbulence. The labor market signal we see is continued investment in building, evaluating and operating AI systems at scale.
What we read
Every quote above is taken verbatim from one of these posts.
- Noah Smith: Why Donald Trump won in 2024, A Nobel for New Keynesian macro, at last!
- Zvi Mowshowitz: OpenAI Fires Three Safety Researchers
- Jack Clark: Import AI 476: Creativity and RSI; worse-than-China open weight models
- Azeem Azhar: 📈 What is bigger than AI?
- Gary Marcus: Could this be the Frontier Labs’ WeWork moment?
- Simon Willison: Hello Python 3.15, goodbye 3.10, and now I can use 3.11, web-edit 0.1a0
- Ethan Mollick: GPT-4 was the undisputed best model in the world for a full year and w
- David Ha: New Sakana AI paper: Recursive Self-Improvement through Multi-Agent Se