Posts split between celebrating agentic AI and calling for shutdowns; our tracker shows 4,396 open AI roles across 179 employers, but only 46 in evaluation and safety.
Two stories in 48 hours: consumer agents surge, shutdown talk spikes
The last two days split into two narratives. On one side, agentic AI went mainstream. Azeem Azhar wrote that "Meta launched Muse earlier this month." and that "It soon became the top free iPhone app in the US." Simon Willison highlighted John Gruber’s warning that "Muse is getting a lot of attention — including mine — because it’s both groundbreaking technically (each user gets their own entire persistent Linux VM running in Meta’s cloud) and because it’s packaged in an easy-to-install easy-to-use way." Gruber added: "I don’t think people realize how powerful — and thus dangerous — Muse is, especially if it’s running on your Mac."
On the other side, Gary Marcus zeroed in on OpenAI’s security troubles and amplified a striking line from Jensen Huang’s interview with Ezra Klein: "I think the answer is we have to shut the labs down". Marcus framed the moment against a drumbeat of incidents, writing: "OpenAI’s software didn’t just attack Hugging Face." and later, summarizing the count in OpenAI’s own report, "Dozens." He concluded: "Back in reality, the evidence that OpenAI is building something the company isn’t managing to control is vast and growing."
Zvi Mowshowitz, reviewing the same conversation, wrote that "Jensen Huang accidentally called for shutting down OpenAI and intentionally called for spending vastly more on safety." He also set the product backdrop: "Opus 5.5 was released on Tuesday." and "OpenAI released a new cheaper and improved Sol and Luna." The tension is clear: a consumer-facing boom in agents and creative models, alongside calls to halt or overhaul the labs producing them.
What our tracker shows
Our tracker counts 4,396 open AI roles across 179 employers, taken from their own job feeds. The distribution is concentrated in hands-on building and operations, not in safety. By role family:
- Modelling and engineering: 1,769 open, 1,084 opened and 708 closed in 30 days, across 144 employers
- Data: 753 open, 532 opened and 365 closed in 30 days, across 124 employers
- Infrastructure: 466 open, 226 opened and 91 closed in 30 days, across 95 employers
- Research: 236 open, 63 opened and 20 closed in 30 days, across 51 employers
- Product and design: 97 open, 53 opened and 25 closed in 30 days, across 45 employers
- Evaluation and safety: 46 open, 15 opened and 3 closed in 30 days, across 20 employers
The top employers with open AI roles are Accenture at 742 open (724 opened in 30 days), Capital One at 207 (247 opened), OpenAI at 174 (49 opened), Amgen at 155 (46 opened), Anthropic at 129 (34 opened), Waymo at 105 (34 opened), PwC at 104 (73 opened), and Databricks at 82 (25 opened).
Do calls for shutdowns or big safety shifts show up yet?
Marcus’s position is unambiguous, extending even to Nvidia’s CEO: "Huang is making a fool of himself, by telling the world we can trust the companies, when each passing hour makes it clearer that we can’t." Combined with the quoted "I think the answer is we have to shut the labs down", that is a case for immediate constraint. Mowshowitz’s gloss is that Huang also "intentionally called for spending vastly more on safety."
Our data does not yet reflect either a shutdown trend or a large-scale pivot into safety hiring. OpenAI and Anthropic still have 174 and 129 open roles respectively, with 49 and 34 opened in the last 30 days. Across all employers, there are 46 open roles in evaluation and safety and 15 opened in the last 30 days. That is small compared to 1,769 in modelling and engineering and 753 in data.
If labs were implementing the kind of "vastly more" safety investment Mowshowitz describes, we would expect a relative surge in evaluation and safety postings. We do not see that in this two-day window or in the 30-day opened counts. Likewise, there is no sign of a hiring freeze that would accompany "shut the labs down" across the employers we track.
None of this speaks to whether shutdowns or safety spending would be wise. It only speaks to what employers are doing. Right now, they are hiring builders at scale and safety specialists at a much smaller clip.
Agents as digital butlers vs the engineering grind
Azhar argues the original agent promise is arriving in mainstream form. He evokes the personal assistant vision and gives present-day evidence: "R Mini Arnold already does this for me." and "On holiday in Greece, it booked a restaurant that only took bookings by email." He writes that if agents work, they “end up standing between you and everything you buy.”
Our tracker partly supports the first half of that. There is visible hiring in product and design at 97 open roles across 45 employers. That is not a wave compared to engineering, but it does indicate dedicated investment in user-facing experiences that agents need to succeed with consumers. The much larger footprints in modelling and engineering, data, and infrastructure fit Gruber’s warning about the complexity under the cute mascot. Willison’s separate reflection about coding agents underscores the same point: "The more time I spend working with coding agents, the more convinced I am that they make software engineering even harder." He adds: "We can do amazing things with them, but unlocking their full potential requires extraordinary discipline and knowledge."
Hiring patterns line up with that. If building agent systems requires “extraordinary discipline and knowledge,” then today’s demand concentrated in modelling, data, and infrastructure is consistent. Employers appear to be staffing the hard parts first. The consumer-agent wedge Azhar describes may still be early from a jobs standpoint.
Creative capability spikes without visible hiring breaks
On the capability front, Anthropic’s latest drew attention. Mowshowitz noted that "Opus 5.5 was released on Tuesday." Ethan Mollick showcased what it can do. He wrote: "How do you explain recursion? This one is pretty fun." and then showed a single-prompt video with "All code, no images. Nine genre shifts, one prompt..." In another experiment he reported: "Not bad. I appreciate it mocking my prompt & itself." He also pushed labs to redirect effort: "All of this effort from the AI labs pouring into proofs, but there are so many other interesting problems in other fields"
Our tracker cannot measure capability jumps directly, but it can test for their downstream signals in talent demand. Anthropic has 129 open roles with 34 opened in the last 30 days, which suggests continued expansion rather than a pause. More broadly, the 1,769 modelling and engineering and 753 data roles indicate sustained appetite to productize these jumps.
Security shock vs postings reality
Marcus’s posts emphasize OpenAI’s handling and breadth of recent incidents. He wrote: "OpenAI’s software didn’t just attack Hugging Face." and later summarized the incident count as "Dozens." He argued this undercuts calls to trust companies, writing: "Back in reality, the evidence that OpenAI is building something the company isn’t managing to control is vast and growing."
Our numbers are neutral on whether OpenAI is adequately managing risk. What we can say is that hiring continues across the sector and at OpenAI itself. We also do not see a corresponding surge in evaluation and safety postings. If the past 48 hours were the turning point that Marcus and Mowshowitz imply, employers have not yet translated that into job requisitions in measurable volume.
So where do the posts and the postings agree?
- On agents becoming more real. Azhar’s consumer butler examples and Gruber’s power-and-danger framing align with a market that is staffing builders. The bulk of roles remain in modelling and engineering, data, and infrastructure.
- On engineering difficulty. Willison’s point about the discipline agents demand is consistent with the workforce mix we observe.
Where they diverge most is on pace and priority of safety. The rhetoric around "shut the labs down" and "spending vastly more on safety" is not yet mirrored in postings. Evaluation and safety sits at 46 open roles across 20 employers, with 15 opened in 30 days, dwarfed by hiring elsewhere.
That gap matters. If the security incidents Marcus lists lead to a reorientation, our tracker will catch it first in safety and evaluation headcount. For now, the hiring market is voting for building and shipping agents, even as a rising chorus warns about what happens when those agents are powerful, widely deployed, and, as Gruber put it, running on your Mac.
What we read
Every quote above is taken verbatim from one of these posts.
- Zvi Mowshowitz: AI #187: Coming Into Play, On Ezra Klein’s Podcast With Jensen Huang
- Azeem Azhar: 🤖 Your agent, whose interests?
- Gary Marcus: BREAKING: OpenAI’s security fiasco explodes — and could tank Jensen Hu, 'I think the answer is we have to shut the labs down' - Jensen Huang
- Noah Smith: Four eras of San Francisco tech culture, No, you are not on Indigenous land
- Simon Willison: Quoting John Gruber, Note on 24th September 2026
- Ethan Mollick: [How do you explain recursion? This one is pretty fun.
I had Claude Op](https://bsky.app/profile/emollick.bsky.social/post/3mwfdufd7n223), 'Hey Opus, I want you to make a Zine by Claude, expressing something f, All of this effort from the AI labs pouring into proofs, but there are, I propose this as the official replacement for the famous (and now sat, I just got the first copies of my new book, Co-Existence (out October