Risk talk rises; hiring skews to builders, not safety

Commentators sound alarms on risk and secrecy, but our tracker shows 1,681 modeling roles vs 44 in evaluation and safety, plus strong enterprise demand.

The split-screen week: existential alarms, secret rules, and shipping products

"Everyone in a position to know is freaking out about AI potentially killing everyone this decade and wants to pace the frontier," writes Zvi Mowshowitz, adding that "people are finally listening." Gary Marcus came out of a BBC hit unimpressed by soothing reassurances, saying "Went on BBC last night and nearly vomited in the electronic green room," and reminding readers of Sam Altman’s prior admission on trust: "Bloomberg Reporter: Even you would acknowledge, you have an incredible amount of power. Why should we trust you? Sam: You shouldn’t." Marcus also says the U.S. government is already regulating AI via a "secret AI security framework," but "U.S. policy on screening AI remains opaque for another day."

At the same time, product builders are, well, building. Simon Willison notes that "Claude Cowork and chat are now one Claude" and observes, "I guess this means Claude is becoming a general agent in its own right." He also tried Google’s new speech-to-speech models: "Google released Gemini 3.8 Live and 3.8 Live Extended Thinking today - two new speech-to-speech models that are a similar shape to OpenAI's GPT-Live family."

Elsewhere, Azeem Azhar reaches for Adam Smith to frame the moment: "My wife asked me, “How would Adam Smith have made sense of AI?” I reckoned he’d love the technology as a productivity tool; he’d be skeptical about companies working together on any issue; and he’d be concerned about giving moral agency to machines." Ethan Mollick, meanwhile, balances capability notes with caution on measurement: "The state of public AI benchmarking is dire and is undermining our ability to understand how good AI is now," and he highlights a Google study claiming acceleration for scientists: "Study from Google about AI use in science, complex impacts: acceleration (7 hours saved per week)." On the labor narrative, Alex Hanna pushes back against a familiar storyline: "The second myth is that "AI" is the reason for recent layoffs, but a deeper look tells a different story."

These are sharp claims about risk, regulation, capability, productivity, and jobs. Our tracker lets us test where employer demand agrees and where it does not.

Where employers are putting headcount right now

Our tracker counts 4,300 open AI roles across 178 employers, taken from their own job feeds.

By role family today:

Modelling and engineering: 1,681 open, 1,046 opened and 567 closed in 30 days, across 142 employers
Data: 831 open, 545 opened and 249 closed in 30 days, across 123 employers
Infrastructure: 412 open, 186 opened and 74 closed in 30 days, across 98 employers
Research: 224 open, 66 opened and 22 closed in 30 days, across 50 employers
Product and design: 99 open, 49 opened and 13 closed in 30 days, across 50 employers
Evaluation and safety: 44 open, 14 opened and 5 closed in 30 days, across 20 employers

Top employers by open roles:

Accenture: 628 open, 817 opened in 30 days
Capital One: 186 open, 140 opened in 30 days
OpenAI: 175 open, 49 opened in 30 days
Amgen: 163 open, 54 opened in 30 days
Anthropic: 126 open, 43 opened in 30 days
PwC: 118 open, 99 opened in 30 days
Databricks: 90 open, 30 opened in 30 days
Waymo: 88 open, 16 opened in 30 days

The pattern is plain: employers are hiring most heavily in modelling and engineering, data, and infrastructure. Evaluation and safety roles exist, but at a much smaller scale.

Do rising risk alarms align with hiring?

Mowshowitz argues the center of gravity has shifted to existential risk: "It only took a few days for the conversation to fully pivot to the counteroffensive," and he frames this as a backlash against people warning "that we are in danger." He also highlights concrete misuse, noting that "A lot of bad guys try to use Claude to do bad things. Mostly they fail. We think." and that "Anthropic has disrupted a bunch of them, and offers an extensive report." He calls out a specific vector: "Indeed, distillation is by far the most important threat in this report, and the part of the report that will have the most impact."

Marcus worries about governance more than capability smoothness: "I do wish what Huang was saying were true, that giant corporations with trillions at stake would police themselves and not release products until they were safe. But I live in the real world. And in the real world OpenAI is burning cash, and needs more of it." He further claims a hidden screening regime: "Remember, even as Trump shouts from the rooftops of Truth Social that he is violently opposed to regulating AI … … he is in fact regulating AI, with a secret AI security framework." Yet, as he writes, "Almost nobody knows what’s actually in it."

Our hiring data does not show a large employer pivot toward evaluation and safety teams that would match a wholesale shift to “pacing the frontier.” Evaluation and safety account for 44 open roles across 20 employers. In contrast, modelling and engineering show 1,681 open roles across 142 employers, with 1,046 opened and 567 closed in the last 30 days. Research roles are also present but smaller at 224 open. If the private sector were reorienting staffing around evaluation and safeguards at scale, we would expect to see much more activity in that family than 14 roles opened and 5 closed in the last 30 days. On these numbers, the hiring emphasis is on building and deploying models rather than on the watchdog side.

That does not falsify Mowshowitz’s or Marcus’s claims about risk or policy opacity, but it does show that employer demand today does not reflect a broad rebalance toward safety hiring.

Productization is visible in commentary and in roles

Willison’s notes reflect an applied push: "Starting today, Claude Cowork and chat are merging into one Claude" and his aside that there are "Echoes of OpenAI renaming their Codex desktop app to ChatGPT a few weeks ago" captures a consolidation into general agent surfaces. His hands-on with real-time models underscores deployment:

"Tool: Gemini Live audio"
"The implementation uses no libraries."

Our tracker shows employers backing that kind of productization with headcount in adjacent families. Infrastructure has 412 open roles, with 186 opened and 74 closed in the last month. Product and design has 99 open, with 49 opened and 13 closed. Those are smaller than core modelling and data, but they are meaningful and consistent with the shift Willison describes toward agentic products and new interfaces like speech-to-speech.

OpenAI and Anthropic appear prominently in our top employers list, with 175 and 126 open roles, respectively, and dozens opened in the last 30 days. That is consistent with rapid product iteration from the labs, even as the public debate centers on risk and regulation.

Productivity narratives vs. the layoffs story

Azhar’s Smith-inspired take is that AI would be loved "as a productivity tool." Mollick gives a concrete, if anecdotal, datapoint about quality control and references in his own work: "when I had an AI go over my new book to check for errors, it found an error not in my reference itself, but in the original article my reference referred to." He also cites a study claiming "7 hours saved per week" for scientists.

Alex Hanna challenges a different storyline altogether: "The second myth is that "AI" is the reason for recent layoffs, but a deeper look tells a different story." Our tracker cannot adjudicate why any specific layoffs happened. What it can show is that employers are still opening AI roles at pace. Across the last 30 days, modelling and engineering saw 1,046 roles opened; data saw 545; infrastructure 186. Consulting and enterprise adopters feature in the top employers: Accenture lists 628 open roles with 817 opened in 30 days, PwC 118 open with 99 opened. That pattern is compatible with a productivity and deployment push and does not match a collapse in AI hiring.

Benchmarking worries vs. who is hired to measure

Mollick warns that "Most famous measures are maxed out," and that remaining benchmarks are "riddled with so many errors that they vastly underestimate AI abilities." If measurement is “dire,” you might expect more employers to staff evaluation and measurement-heavy roles. Yet evaluation and safety roles total 44 open across 20 employers, and only 14 opened in the last month. The numbers do not show a surge of hiring aimed at fixing or hardening evaluation. That gap between public worry about benchmarks and private hiring is one place where our data does not back the rhetoric.

Misuse and distillation: visible in discourse, not yet in demand signals

Mowshowitz treats model distillation without safeguards as the standout threat: "By exposing Chinese attempts at systematic fraudulent distillation of Claude, Anthropic has embarrassed and potentially antagonized the Chinese." Our tracker does not break out roles focused on anti-distillation or threat intelligence specifically. The closest proxy is the evaluation and safety family, which, as noted, remains small relative to builders. If companies are chasing defense-in-depth against misuse at scale, that is not yet evident in posted roles.

Bottom line

The last two days of posts swing between alarm and acceleration. Zvi Mowshowitz and Gary Marcus foreground existential risk, misuse, and opaque government screening, while Simon Willison chronicles rapid product consolidation and new real-time modalities, Azeem Azhar and Ethan Mollick point to productivity, and Alex Hanna challenges the “AI caused the layoffs” narrative. Our tracker shows employers are still hiring most heavily to build and deploy: 1,681 modelling and engineering roles and 831 data roles open, versus 44 in evaluation and safety. Where the commentary says the conversation has pivoted to risk, employer demand has not pivoted with it.

What we read

Every quote above is taken verbatim from one of these posts.