Posts by Gary Marcus, Azeem Azhar, Simon Willison, Ethan Mollick and David Ha stress risk, meaning and cost, but our tracker shows 4,226 open AI roles led by engineering and just 47 in evaluation and safety.
The weekend’s split-screen: risk, meaning and costs vs product momentum
A cluster of posts in the past two days pulled in opposite directions. Gary Marcus pressed for tougher governance and flagged a controversial attitude to harms. Azeem Azhar said "AI is really forcing some big questions into the open." Simon Willison argued for strict cost controls as agents scale. In the same window, Ethan Mollick showcased a working game prototype and announced a book aimed in part at AI readers, and David Ha promoted a research symposium.
Our job tracker cuts through the noise. We count 4,226 open AI roles across 180 employers, taken from their own job feeds. The mix is clear: modelling and engineering lead by a wide margin, infrastructure remains strong, research is steady, and evaluation and safety are a small slice. The pattern in hiring does not match the scale of the safety and meaning debates these authors are surfacing.
What they said
Gary Marcus previewed testimony for a New York City Council hearing and underscored a values clash he wants the city to confront. He highlighted Politico’s report that Sam Altman said, “we believe that the world should accept some bad things happening for the benefits of this technology and people having the agency”. Marcus’s framing is unequivocal: "We stand at a pivotal moment in history, and New York City has a chance to improve our chances, both for itself, and for the world." His metaphor is blunt: "Contemporary AI is like a bull in a china shop, powerful and impressive, but difficult to control."
Azeem Azhar pulled on two threads. First, in a weekend curation he wrote, "AI is really forcing some big questions into the open." After summarising debates about effective altruism and even theology around machine consciousness, he closed that item with "These are huge questions." In a separate note on energy transition, Azhar argued that progress can be easy to miss if you look at the wrong measure: "The electrification of the world economy is further ahead than most people think." And: "Which bar you watch determines how far along you think we are."
Simon Willison focused on something concrete: runaway bills from autonomous agents. "Here's a product feature which the world is going to need a whole lot more of over the coming months and years: default hard budget caps." He warned that "These need to be hard limits." and that soft warnings are insufficient: "Soft caps, "after $X/month, send me a warning email", will not cut it." The driver, in his view, is straightforward: "Coding agents, and personal agents (coding agents wrapped in a less threatening UI), greatly reduce the friction of spinning up code that can do useful things." His prescription is simple: "I think hard budget caps need to be the default." with a UI affordance to opt out: "Have a nice, clear checkbox somewhere prominent: Remove the budget cap."
On the product and culture side, Ethan Mollick blended media and artifacts. He wrote: "I think my upcoming book, Co-Existence, might be the first to include a blurb written specifically for AI readers, in this case from Tyler Cowen (whom I thought AIs would respect) The book website (with elaborate pre-order bonus) also has a page for AIs. Its out October 20: co-existence.ai" He also showcased a working demo: "Here is Fable 5.1, a little over a year since GPT-5 below & a single shot (compared with many rounds with GPT-5.6) Quite good: brut-city.netlify.app"
David Ha’s note points to the research community’s drumbeat: "Registration now open: From World Models to Real-World AI Luma: luma.com/kjf0z516 Join us on October 26 in Tokyo for a symposium with Jürgen Schmidhuber, who recently joined Sakana AI as Chief Scientific Advisor."
What the hiring tape says
Against this backdrop, employer hiring looks practical and execution-heavy. Our tracker counts 4,226 open AI roles across 180 employers. By role family:
- Modelling and engineering: 1,683 open, 889 opened and 778 closed in 30 days, across 137 employers
- Data: 733 open, 478 opened and 405 closed in 30 days, across 120 employers
- Infrastructure: 467 open, 230 opened and 101 closed in 30 days, across 97 employers
- Research: 234 open, 49 opened and 22 closed in 30 days, across 53 employers
- Product and design: 101 open, 41 opened and 22 closed in 30 days, across 48 employers
- Evaluation and safety: 47 open, 14 opened and 3 closed in 30 days, across 20 employers
Top employers are skewed to scaled enterprise deployment. Accenture has 658 open, 522 opened in 30 days. Capital One has 214 open, 255 opened in 30 days. OpenAI has 179 open, 47 opened in 30 days. Amgen has 158 open, 44 opened in 30 days. Anthropic has 137 open, 37 opened in 30 days. Waymo has 112 open, 35 opened in 30 days. PwC has 100 open, 58 opened in 30 days. Databricks has 86 open, 14 opened in 30 days.
Where posts and numbers align
Willison’s warning about agents quietly spinning up costly activity shows up in the demand for infrastructure roles. With 467 infrastructure openings, and 230 opened in the past 30 days across 97 employers, companies are staffing the plumbing that would both enable and constrain usage. That is consistent with his call for budget guardrails to be a default product capability, not an afterthought.
Mollick’s artifacts point to a maturing product surface. But our counts show only 101 open product and design roles, across 48 employers. That is a relatively small pool compared to modelling and engineering’s 1,683 openings. The implication is that most hiring remains under the hood, building and scaling systems that prototypes can sit on top of, rather than flooding into new AI-first product management or UX hires.
Ha’s symposium announcement rhymes with our research openings. There are 234 research roles open, with 49 opened and 22 closed in 30 days, across 53 employers. That is not a hiring frenzy, but it is persistent. It suggests a steady investment in fundamental and applied research, consistent with continued community events.
Azhar’s broader point about measurement frames is a useful caution for reading job tapes too. "Which bar you watch determines how far along you think we are." If you focus on safety discourse, the market might look frozen. Focus on engineering requisitions and it looks busy. Our data shows activity concentrated in implementation roles, which fits an adoption phase rather than an abstract ethics phase.
Where commentary and our data diverge
The most striking mismatch is between the prominence of risk and safety in the conversation and the small number of safety jobs. Marcus’s preview underscores that the stakes are high and that a major city is moving on policy. Yet evaluation and safety account for only 47 open roles, with 14 opened and 3 closed in 30 days, across 20 employers. For all the urgency in "Contemporary AI is like a bull in a china shop, powerful and impressive, but difficult to control.", employers are not yet posting at scale for people to build those controls.
Azhar’s second post goes even deeper into meaning. "These are huge questions." Our tracker cannot adjudicate consciousness debates, but it can show whether companies are buying capacity to test and evaluate system behavior. On that, the needle is thin. The heavier lift is still in modelling and engineering. If safety hiring were to surge, we would expect to see a measurable uptick from 47 open roles and broader distribution beyond 20 employers. We do not see that yet.
Macro context: how we calibrate the signal
Our headline jobs-created indicator is counted from employer listings and calibrated to the World Economic Forum’s Future of Jobs Report 2025, which projects 11M AI roles created and 9M displaced by 2030. That calibration anchors the long-term arc. The short-term picture from this week’s tape is simpler: companies are hiring people to build, deploy and maintain systems, not to staff safety teams at scale.
What to watch next
- If Marcus’s policy push gains traction, the share of evaluation and safety roles should rise above the current 47 open and 14 opened in 30 days.
- If Willison’s budget-cap argument takes hold, infrastructure hiring could remain elevated or accelerate from 467 open and 230 opened in 30 days, as teams invest in usage governance and controls.
- Enterprise adoption remains the big engine. Accenture’s 658 open roles and Capital One’s 214 open hint that services and financial firms are the chief buyers of AI talent right now, more than research labs.
The weekend’s posts make the philosophical and governance stakes vivid. Our numbers make the operational reality plain. For now, risk talk is loud, but engineering jobs are louder.
What we read
Every quote above is taken verbatim from one of these posts.
- Gary Marcus: Coming soon: New York City’s hearing on AI risks
- Azeem Azhar: 🔮 The transition is hiding in plain sight #604, 🔮 Quick weekend reads: the big AI questions
- Noah Smith: The second Trump presidency is a world-historic failure
- Simon Willison: We're going to need default hard budget caps on pretty much everything, September sponsors-only newsletter
- Ethan Mollick: I think my upcoming book, Co-Existence, might be the first to include , 'create the ultimate brutalist city builder. you should have lots of c
- David Ha: [Registration now open: From World Models to Real-World AI
Luma: luma.](https://bsky.app/profile/hardmaru.bsky.social/post/3mx3kq6snoc2v)