New posts tout agent swarms and warn of AI risks; our tracker shows surging hiring in engineering and infra, but only 49 open evaluation and safety roles across 21 employers.
Agents get easier, models get cheaper, and AI beats junior accountants
Across the last two days’ posts, there is a shared story of accelerating capability and growing unease about its implications for work.
Ethan Mollick says the hard problems of coordinating agents may already be melting away. After anticipating managers would need to carefully orchestrate agent teams, he now writes simply: "Nope." He adds: "I fell prey to The Bitter Lesson, the hard truth, learned over and over again, that things that we thought required elaborate human rules and thinking can be solved with the brute force of better machine learning systems and more AI." That squares with the move toward ever more autonomous agents—and the arrival of products like "Dots" referenced elsewhere this week—but it also raises the stakes for what kinds of human jobs get augmented next.
On that front, Mollick shares a stark data point from new research on finance workflows: “We find that on medium-length, well-defined accounting tasks, frontier AI models are now faster and more accurate than junior accountants, even the best one in our study.”
Zvi Mowshowitz, meanwhile, frames the model competition as red hot and price-driven. "Is Google back?They claim that they are back." He notes "Gemini 4 Argon is rolling out, with competitive frontier-level benchmarks, at $2/$10" and that OpenAI is countering with new options at similar prices. On quality, he is emphatic: "The new hotness remains Claude Opus 5.5. This model rocks. It has made me considerably more productive and made my day more pleasant. It should raise your ambitions."
At the same time, Mowshowitz’s second post argues that public and insider concern over AI’s trajectory is spreading: "Those who know keep warning us about AI risks. People are listening." He cites a fresh resignation: "Robert O’Callahan: I resigned from Google today." O’Callahan’s explanation, quoted by Mowshowitz, grounds that concern directly in compute progress: "I enjoyed my work and loved the people, but my GDM team was working on a new generation of chips to make AI much faster and cheaper, and I think AI is already progressing too fast, so I had to quit." Emily M. Bender and Timnit Gebru push for guardrails from the other direction of the pipeline: "It is high time for us to regulate usage of systems sold as “AI”, not because they are magical machines, but because they are error-prone products that shouldn’t be used in high stakes scenarios."
Alex Hanna and Bender also spotlight who gets left out when data is scraped indiscriminately. Hanna writes: "Indigenous languages are often hidden to LLMs, for good reason." Bender amplifies a related conversation: "What a thrill to get to interview Keoni Mahelona of Te Hiku Media for Mystery AI Hype Theater 3000!" The theme is that stewardship and sovereignty over data are labor issues, not just technical ones.
And in a historical provocation, Dwarkesh Patel draws a line from conquest to disruption: "We often forget how wild human history is. People talk about AI takeover as if it would be some crazy, unprecedented sci-fi outcome."
What the hiring pipeline shows right now
Our tracker counts 4,325 open AI roles across 180 employers, taken from their own job feeds. The center of gravity is not moving away from builders:
- Modelling and engineering: 1,735 open, 971 opened and 795 closed in 30 days, across 137 employers
- Data: 754 open, 529 opened and 415 closed in 30 days, across 122 employers
- Infrastructure: 467 open, 240 opened and 116 closed in 30 days, across 96 employers
- Research: 235 open, 55 opened and 23 closed in 30 days, across 53 employers
- Product and design: 103 open, 47 opened and 24 closed in 30 days, across 49 employers
- Evaluation and safety: 49 open, 15 opened and 3 closed in 30 days, across 21 employers
On the face of it, Mollick’s point about agents becoming easier to coordinate is not yet showing up as a shift away from core engineering hiring. Modelling and engineering remains the largest family by a wide margin. If anything, the steady demand in Infrastructure—467 open, with 240 opened in the last 30 days—tracks with O’Callahan’s description of the industry push to make systems "much faster and cheaper," even as he personally opted out.
At the employer level, the firms that would productize automation for clients are prominent. Accenture has 692 open AI roles and 555 opened in 30 days. PwC has 100 open and 61 opened in 30 days. Capital One shows 216 open and 263 opened in 30 days. This is consistent with Mollick’s accounting evidence. If frontier models are already above junior-accountant performance on "well-defined accounting tasks," then consultancies and financial services players need people to integrate those capabilities into real workflows. Our data supports that integration wave.
On the labs side that Mowshowitz tracks closely, OpenAI has 180 open AI roles, with 48 opened in 30 days. Anthropic lists 137 open, with 40 opened. Those are sizable pipelines that underline Mowshowitz’s claim that competition at the frontier is intense and ongoing. They do not suggest a hiring pause in response to the alignment worries he mentions. That is a contrast worth noting: "OpenAI was forced to pull what would have been GPT-6.1 Astra due to alignment failures," Mowshowitz writes, yet the hiring tap at the labs he names remains open in our counts.
Safety talk is loud; safety headcount is still small
Two related claims are that the policy mood is shifting and that systems should face tighter controls. Mowshowitz describes a "‘morally binding’ White House accord on AI Safety" and Bender and Gebru argue for regulation because systems "are error-prone products that shouldn’t be used in high stakes scenarios." Our tracker does show activity in Evaluation and safety, but it is modest: 49 open roles across 21 employers, with 15 opened and 3 closed in 30 days. Against 4,325 open AI roles overall, that is a small fraction. The policy conversation may be accelerating, as Mowshowitz says, but the hiring response on safety and eval is lagging the engineering buildout.
Research hiring, at 235 open roles and 55 opened in 30 days, also remains secondary to application and platform work. If a new White House mood music were already translating into headcount for embedded evaluators, we would expect to see a sharper rise in those eval and safety openings. We do not see that yet.
Agents without orchestration? Hiring does not show a pivot
Mollick writes that "A huge amount of effort went into building elaborate computer systems to feed AIs the right information at the right time, but AI systems have learned to seek out information themselves." He continues: "The same thing happened to prompting. People built elaborate templates and chains of prompts that walked the AI through a task one step at a time." If agent swarms now demand much less scaffolding, you might expect a pullback in roles focused on prompt-chains and orchestration, and a relative rise in product roles to package agentic behavior.
We cannot see that pattern clearly in our families. Product and design roles are 103 open, with 47 opened and 24 closed in the last 30 days across 49 employers. That is healthy, but still small beside modelling and engineering. The data neither confirms nor contradicts Mollick’s claim about how teams will work with agents. It does, however, show that the hiring market continues to prize builders of core capability and data pipelines most.
Data sovereignty and language stewardship: hiring signal is diffuse
Hanna urges caution on corporate narratives about preservation: "Indigenous languages are often hidden to LLMs, for good reason." Bender flags a conversation about "technosolutionism and colonialism in language technology." Our tracker’s Data family has 754 open roles across 122 employers, with 529 opened in 30 days. That scale certainly includes work on dataset curation and governance. But our taxonomy does not let us isolate roles centered on Indigenous data stewardship or language sovereignty. We cannot claim a specific hiring uptick matching Hanna and Bender’s focus from these figures alone.
Method note
Our counts are from employers’ own job feeds. Our headline jobs-created indicator is counted from job listings and calibrated to the World Economic Forum’s Future of Jobs Report 2025 (11M AI roles created, 9M displaced, by 2030).
What to watch next
- Do evaluation and safety openings rise from today’s 49 as policymakers and companies move from statements to staffing? Mowshowitz’s report of a "‘morally binding’" accord and Bender and Gebru’s call for regulation have not yet been matched by headcount in our data.
- Does consulting and finance hiring keep pace with Mollick’s accounting result? Accenture, PwC, and Capital One are already among the most active AI employers in our tracker.
- Does infrastructure hiring stay elevated in the face of compute concerns? O’Callahan’s resignation highlights a moral line some engineers draw. Our counts show many employers still recruiting to make systems "much faster and cheaper."
The through line in this week’s commentary is acceleration, often without as much orchestration as expected. Our hiring data matches that in where firms are putting people: core engineering, data, and infrastructure. The cautionary voices are getting louder. For now, the jobs market is still building.
What we read
Every quote above is taken verbatim from one of these posts.
- Zvi Mowshowitz: AI #188: Gemini Dot Argon, The AI Preference Cascade Reaches Farther
- Ethan Mollick: The Dot and the Swarm, 'We find that on medium-length, well-defined accounting tasks, frontie
- Dwarkesh Patel: Si Sheppard – How did a few hundred Spanish soldiers topple two empire
- Emily M. Bender: On the next Mystery AI Hype Theater 3000 livestream, we’re so excited , What a thrill to get to interview Keoni Mahelona of Te Hiku Media for , [New from @timnitgebru.blacksky.app and me:
'It is high time for us to](https://bsky.app/profile/emilymbender.bsky.social/post/3mwsxbeju2kxj)