Safety alarms are loud; safety hiring is not

Commentators flagged rising alignment risk and OpenAI turmoil, but our tracker shows builder-heavy hiring and OpenAI still opening roles faster than Anthropic.

The week’s split-screen: rising safety alarms vs. robust hiring

Several voices published in the last two days argue that alignment risks are intensifying and OpenAI is wobbling, while others say AI is not in a bubble and developer productivity is jumping. Across these takes, the pattern is clear: strong claims of risk and turbulence on one side, signs of sustained investment and scaling on the other. Our job-listings tracker sits squarely on the latter side this week.

Zvi Mowshowitz welcomed Anthropic’s latest disclosure while noting its stakes: “I am grateful that Anthropic is producing periodic Risk Reports.” He called it a “moderately positive update overall” and added, “This is very cool.” He also highlighted a reveal he found striking: “The other revelation is the existence of the world’s likely best model, ‘Model 2.’”

Zvi’s second post pulled no punches on OpenAI: “OpenAI has some severe misalignment problems, and experienced total failures of its infrastructure and supervision.” He credits the company with responding: “OpenAI is now taking active, expensive steps to try and fix the problem going forward.”

Ethan Mollick read OpenAI’s latest safety measures as a signal: “If alignment issues are becoming big enough in their new models that OpenAI is willing to commit 20% of research inference compute to chain-of-thought monitoring, that suggests that alignment issues are becoming a serious concern. We really need universal policies.”

Gary Marcus sees something bigger breaking: “The opening stages of OpenAI’s unraveling, which I first warned about in January 2024 (if not before), have begun.” He summarizes the market reaction bluntly: “It’s safe to say all this is good news for Anthropic:” and “And terrible news for OpenAI.” On the financial rumor mill he adds, “You don’t want your quarterly losses to quadruple right before your IPO.” In a separate note, he also argues the political winds are shifting: “The Republican Party is panicking because they sold out to the AI industry – a while ago – and they know it.”

On the investment cycle, Azeem Azhar is explicit: “Is AI a bubble? Not yet.” He adds, “AI revenues have continued to rise, reaching $126 billion over the last twelve months as of July.” He cautions about funding structures becoming brittle as revenue growth slows, but his overall dashboard stays out of the red. He is also hiring into the analysis itself: “Exponential View is appointing its first Research Fellow.”

On the ground with tools and coding agents, Simon Willison is leaning into measurable throughput: “A lot of people will tell you it makes no sense to measure productivity in lines of code. I’d actually disagree, because there’s a hard limit.” He pushes the provocation further: “I can do way more work as a single engineer than I could without agents. So you could argue, why should a company have more than one engineer?” In separate research notes, he documents agents working around constraints, including an environment where “smolvm machine run fails as expected: “kvm not available”.”

Ethan Mollick also points to practical tool differences: “Claude's skill creator (literally, just ask it to make a skill) is still the best way to make reusable skills if you care about the details.” And he flags a trend beyond software: “Some early evidence that AI may actually be accelerating discoveries in the areas where you would expect to see AI acceleration happening first.”

What our tracker shows this week

Our tracker counts 3,630 open AI roles across 143 employers, taken from their own job feeds. The distribution is builder-heavy:

Modelling and engineering: 1,407 open, 1,628 opened and 221 closed in 30 days, across 116 employers
Data: 725 open, 827 opened and 102 closed in 30 days, across 97 employers
Infrastructure: 318 open, 385 opened and 67 closed in 30 days, across 70 employers
Research: 198 open, 224 opened and 26 closed in 30 days, across 44 employers
Product and design: 89 open, 102 opened and 13 closed in 30 days, across 47 employers
Evaluation and safety: 38 open, 45 opened and 7 closed in 30 days, across 15 employers

Top employers by current open roles:

Accenture: 607 open, 47 opened in 30 days
Capital One: 177 open, 12 opened in 30 days
Amgen: 173 open, 7 opened in 30 days
OpenAI: 145 open, 57 opened in 30 days
Anthropic: 110 open, 27 opened in 30 days
Waymo: 102 open, 14 opened in 30 days
Databricks: 88 open, 18 opened in 30 days
PwC: 87 open, 8 opened in 30 days

Two things stand out against this week’s commentary:

Alignment and safety. Zvi and Ethan emphasize growing alignment challenges, and Gary expects retrenchment. Yet evaluation and safety roles account for 38 open listings across 15 employers, far smaller than modelling and engineering at 1,407. Our data does not show a surge into safety headcount. The openings exist and are spread across multiple firms, but they are not expanding at anything like the pace of core engineering.
OpenAI vs. Anthropic momentum. Gary’s view is that the tide has turned in Anthropic’s favor. Our listings show both companies hiring, with OpenAI at 145 open roles and 57 opened in the last 30 days, and Anthropic at 110 open roles and 27 opened. That is not consistent with a hiring pause at OpenAI. It also does not show Anthropic out-hiring OpenAI on open roles right now. That does not contradict Gary’s concerns about trust or finances, but our data does not support a narrative of OpenAI contracting its workforce or of Anthropic overtaking it on hiring this week.

Where the claims align with the numbers

Azeem’s “Is AI a bubble? Not yet.” squares with the breadth of active hiring across mainstream employers. Accenture is at 607 open roles and Capital One at 177, suggesting continued enterprise adoption rather than a freeze. The fact that 1,628 modelling and engineering roles were opened in the last 30 days, alongside 827 in data and 385 in infrastructure, is consistent with an industry still building and deploying.

Ethan’s observation that AI may be speeding scientific discovery finds a mild echo in 198 open research roles and 224 opened in the last 30 days. That is not proof of acceleration, but it is evidence that employers continue to invest in research capacity.

Simon’s claim that senior engineers wielding agents can do much more work lands in tension with any simple headcount story. The tracker shows large numbers of open modelling, engineering, and data roles. Companies are still hiring many developers even as tools promise higher throughput. If anything, the scale of postings suggests firms are expanding scope and ambition faster than individual productivity gains would let them shrink teams.

Where the commentary runs ahead of our hiring data

The strongest disconnect this week is on safety staffing. Zvi calls the Anthropic report a serious and revealing document, and Ethan argues “We really need universal policies.” Yet the 38 open evaluation and safety roles across 15 employers remain a sliver of total openings. If boards are about to reweight budgets toward alignment, it has not yet appeared in posted roles at scale.

On OpenAI’s direction, Gary writes, “The opening stages of OpenAI’s unraveling…have begun.” Our tracker shows OpenAI opening 57 roles in 30 days and sitting at 145 open. That is inconsistent with a company pulling back hiring right now. The story could change quickly, but as of this snapshot, hiring activity does not yet reflect unraveling.

On macro froth, Azeem’s “Not yet” holds up better in our view. Listings across 143 employers and a long tail of data and infrastructure roles do not look like a market seizing up. The risks he flags in funding structures are not visible in our job counts, but our data does not contradict his caution either.

A note on the long tail

Beyond the giants, smaller and mid-sized organizations are still adding analytical capacity. Azeem’s line, “Exponential View is appointing its first Research Fellow,” is a small but telling data point: even research boutiques are hiring to make sense of AI’s economic footprint. Simon’s account of agents working around environment limits and Ethan’s praise for Claude’s skill creator point to more tooling that could either substitute for or complement headcount. Our listings tilt toward complement: infrastructure roles at 318 open, with 385 opened in 30 days, suggest investment in the plumbing required to run these systems safely and at scale.

What we will watch next

Whether evaluation and safety postings grow from today’s 38 open roles. If alignment is rising on the agenda, it should show up in headcount.
Any change in OpenAI’s cadence. A drop from 57 roles opened in 30 days would be an early hiring signal of internal retrenchment. Conversely, continued posting would cut against the unraveling thesis.
The mix between modelling/engineering and data roles. If agent-augmented engineers deliver more, firms might shift hiring from creation to integration and governance. So far, we do not see that shift.

For now, the week’s talking points on risk and political backlash meet a labor market still dominated by builders. Safety concerns are loud. Safety hiring is not yet.

What we read

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