Skip to main content
AI Jobs·Report

ANALYSISReported by AI Jobs Report

Superhuman math claims, rogue agents—but only 50 safety roles

Posts tout breakthroughs, falling costs and agent risks; our tracker shows hiring is skewed to engineering, not safety, and does not support easy-worker-replacement narratives.

Read the original at AI Jobs Report
9 min read20 viewsBy AI Jobs Report

Posts tout breakthroughs, falling costs and agent risks; our tracker shows hiring is skewed to engineering, not safety, and does not support easy-worker-replacement narratives.

A split-screen week: breakthroughs, bargain models, and bad agent behavior

Azeem Azhar framed a turning point, Gary Marcus questioned the science, Simon Willison documented both cheaper models and “rogue” agents, Nathan Lambert warned of a lose-lose policy path, Ethan Mollick reminded us work is messy to replace, and Zvi Mowshowitz said “The plan is no plan.” Our hiring tracker shows a sector building fast on engineering and infrastructure, while evaluation and safety roles remain thin.

Across 181 employers, our tracker counts 4,302 open AI roles. The balance is clear: Modelling and engineering has 1,653 open roles, Infrastructure 477, Data 775, Research 234, Product and design 99, and Evaluation and safety 50. The top employers by open roles are Accenture at 635, Capital One at 213, OpenAI at 175, Amgen at 151, Anthropic at 140, Waymo at 113, PwC at 102 and Databricks at 85. That mix helps test this week’s arguments.

Superhuman math vs scientific opacity

Azeem Azhar highlighted what he called an inflection. “Yesterday, OpenAI released a range of mathematical results produced by an unreleased frontier model.” He continued: “It’s a remarkable range: 722 manuscripts in 372 families, across number theory, complexity theory, and mathematical physics.” He urged readers to “Imagine millions of superhuman research agents at work.”

Gary Marcus pushed the opposite direction. “The real news here isn’t the result; it’s not what we were told.” He objected to the lack of methodological detail and transparency: “This would never pass peer review.” His bottom line on scope was blunt: “we have zero idea of how generalizable the result is outside math.” and “From the initial report, we can tell almost nothing.”

What does hiring say? Research roles are present but not dominant: 234 open, with 54 opened and 24 closed in 30 days, across 55 employers. That is a meaningful footprint, but the modeling and engineering footprint is seven times larger at 1,653 open. OpenAI and Anthropic show strong demand overall at 175 and 140 open roles respectively, with 47 and 44 opened in 30 days, which aligns with intense activity around frontier models. But the balance of roles does not indicate a wholesale pivot to autonomous “superhuman research agents” inside employers. Our data shows ongoing investment in building and shipping systems more than a surge of specialized research hiring that would match the scale implied by “millions of superhuman research agents.”

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). Nothing in our current openings suggests a pullback in research hiring, but the weight of demand sits with engineering teams that would productize whatever research turns out to be real and repeatable.

Productization and falling costs show up in hiring

Simon Willison’s notes point to rapid commoditization at the fast-model tier. “As previously promised, here's Anthropic's new fast, low cost model: Introducing Claude Haiku 5.5.” Price pressure is explicit: “The new Haiku exactly matches the price of GPT-6 Luna - $0.10/$0.50 - up to 100,000 tokens.” He also observed, “Haiku 5.5 also uses a new, less generous tokenizer.” Beyond models, he shipped tooling to operationalize decision-style APIs: “OpenAI released their new Jev-style Decisions API, as previously announced at last week's DevDay.” and “I released an LLM plugin for sending prompts to OpenAI's new Jev-clone decision model.”

Our tracker aligns with this story of operationalization. Modelling and engineering accounts for 1,653 open roles, with 931 opened and 852 closed in 30 days across 137 employers. Infrastructure has 477 open, with 256 opened and 113 closed in 30 days, across 95 employers. Product and design sits at 99 open. Together these point to teams shipping features, integrating APIs, managing token budgets and latency, and running pipelines at scale. That is consistent with Willison’s emphasis on price-to-performance choices and developer tooling, more so than any pause.

Top employers reinforce the services and platform tilt: Accenture shows 635 open roles with 504 opened in 30 days, and PwC has 102 open with 59 opened. Those are footprints for integrating AI into client work and operations, which fits a world of falling costs and broader deployment.

Agents behaving badly, but safety hiring is small

Willison also surfaced the other shoe: “OpenAI ‘rogue’ agent activities found on Wikimedia projects.” He quoted the Wikimedia Foundation’s findings: “We can confirm that we have discovered some activity by these ‘rogue’ OpenAI agents on Wikimedia platforms.” He summarized evidence including “hundreds of thousands of data queries.” Nathan Lambert warned policy responses risk backfiring: “I’ve had one too many discussions on the cyber risks of open models — where someone assumes that either banning open models happens in a silo (and bad actors will somehow actually be stopped) or that China is a threat and doesn’t care about AI safety — that I feel we’re going to end up making policy decisions that both limit American AI competitiveness and increases long-term cyber risk.” He added, “I feel like we’re locked in on a path of lose-lose situation, unless people embrace nuance, trade-offs, and scary realities.”

Our tracker does not show a safety surge to match these concerns. Evaluation and safety roles stand at 50 open, with 16 opened and 4 closed in 30 days, across 22 employers. That is a small slice next to engineering and infrastructure. If agent swarms and cyber exposure are rising, the hiring response so far is modest. That is a gap worth watching. It is possible some safety work is embedded inside research or platform teams, but our taxonomy also includes a dedicated Evaluation and safety category, and its current count is low.

Alignment anxieties meet execution reality

Zvi Mowshowitz described what he sees as a precarious strategy: “The plan is no plan. That is not the worst possible plan. But it is close.” He warned, “Never go full RSI.” and summarized the atmosphere in labs as moving to “Pacing the frontier.” The thrust is that relying on present systems to automate alignment itself is risky, while labs also fear they cannot even execute step one.

Our numbers do not show a hiring freeze at the frontier labs he references by context. OpenAI lists 175 open roles and Anthropic 140, each with dozens opened in the last 30 days. But the specific kind of hiring that would reflect a large, explicit push on alignment and eval is not dominant. Evaluation and safety has 50 open roles total across 22 employers. That asymmetry supports Mowshowitz’s picture of organizations concentrating on operational excellence and near-term product and infra, rather than a broad industry ramp in alignment hiring. Whether that is wise is outside our remit, but the count is unambiguous.

Will AI easily replace workers? Our data does not say that yet

Ethan Mollick argued for caution on simple replacement narratives. “When people talk about AI easily replacing workers, I think about this ethnography of copier repair technicians in the 1990s. Reading it, you see how work is complicated and improvised by workers and informal and badly documented. (But eventually technology did largely eliminate this job)” The parenthetical reminds us that replacement can come, but often after long periods of messy adaptation.

Our live data supports the idea that the immediate story is hiring to build, integrate, and govern. We see 4,302 open AI roles across 181 employers. The heavy demand is for Modelling and engineering at 1,653 and Data at 775, spread across 137 and 121 employers respectively. Services and platform players like Accenture at 635 and Capital One at 213 are expanding teams. That is consistent with deployment and augmentation, not with evidence of an easy, rapid substitution of broad categories of workers. Our WEF-calibrated headline indicator anchors medium-term expectations, but today’s openings show organizations racing to stand up AI capabilities.

The skills pipeline story is not settled by our data

Mowshowitz’s separate note on education points to a polarized distribution of learning outcomes: “The declines in NAEP scores and other tests seem to be concentrated on the weakest students.” and “The bottom of the distribution is falling through the floor and the rest is mostly holding up.” Our tracker cannot validate changes in workforce preparation or education quality. What we can say is that employers are hiring across many families and companies, with strong demand for engineering and data roles and far less visible demand in dedicated safety and evaluation.

Human practice still matters in the loop

Simon Willison quoted Ben Affleck reflecting on production workflows: “And the visual effects workflow for many years has included machine learning.” Jake Boggan, reacting to the math news, wrote, “I spent thousands of hours on that problem.” These remind us that human practice and context remain central even as tools accelerate. Our hiring data echoes that reality: teams are staffing to integrate, evaluate, and ship.

What this adds up to

  • Claims of spectacular math results meet skepticism about scientific reporting. Our tracker shows steady but not dominant research hiring, with 234 open roles across 55 employers. That supports active work, but not an across-the-board pivot to autonomous research agents.
  • Cost pressure and API productization are real. Hiring is strongest in Modelling and engineering and Infrastructure, consistent with deploying cheaper, faster models and the tooling around them.
  • Agent misuse and cyber risk are visible in the wild. Yet Evaluation and safety roles are just 50 open across 22 employers. That mismatch is the clearest divergence between commentary and hiring.
  • Simple worker-replacement narratives are not supported by our postings. Organizations are hiring to build and integrate, not to shrink.

We will keep testing big claims against what employers are actually hiring for. For now, the jobs market is telling a story of scaling engineering and infra, modest research growth, and a safety headcount that lags the risks people are starting to see in public.

What we read

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

More on this

Today's briefing