Practitioners say smaller, cheaper, good-enough AI is winning; our tracker shows 671 data roles opened in 30 days and infrastructure close behind, while research hiring is smaller.
The emerging split: frontier breakthroughs vs good-enough deployments
Several practitioners this week argue that practical, cheaper AI is overtaking the need for constant frontier leaps, while others focus on organizational turbulence and safety. Azeem Azhar frames the new reality in costs and deployment: “Amazon spent some $1.8 million on a Claude project that ran for five months. A senior employee said: “It’s difficult to figure out how much anything [AI-related] costs”.” He also reports he “called out an outrageous few days when the bot blew through $500 a day.” Nathan Lambert highlights China’s small-model strategy: “This puts the model more or less at the frontier of agentic coding benchmarks, with only ~750B parameters – a third of Kimi K3!” And David Ha narrows the focus of what businesses actually need: “most current baseline models are actually perfectly fine for 99% of everyday work.”
Our hiring tracker shows employers weighting toward applied buildout rather than pure frontier research. Our tracker counts 3,256 open AI roles across 102 employers, taken from their own job feeds. By role family, Data leads: 643 open, 671 opened and 28 closed in 30 days, across 74 employers. Infrastructure is also heavy: 302 open, 343 opened and 41 closed in 30 days, across 55 employers. Research remains active but smaller: 185 open, 198 opened and 13 closed in 30 days, across 38 employers. Product and design sits at 71 open, 79 opened and 8 closed in 30 days, across 34 employers. Evaluation and safety has 39 open, 44 opened and 5 closed in 30 days, across 15 employers.
Cost and complexity pressures are hiring drivers, not brakes
Azhar’s account emphasizes that cost control is a capability in its own right. He notes “The price war that has broken out between Anthropic and OpenAI in response to Chinese advances has helped even more.” He also tightened his own stack: “By default, RMA now uses my token allowance on OpenAI Codex, which is already paid for in my $200-a-month subscription. It will fall back to DeepSeek v4 Flash or Pro if OpenAI is unavailable.”
Our numbers suggest employers are not waiting for clean answers on cost; they are hiring the people who will produce them. The surge in Data roles (671 opened in 30 days) and Infrastructure (343 opened) is consistent with organizations that need telemetry, routing, model selection, data pipeline reliability, and spend optimization. That weighting also fits with integrators and platform-heavy firms leading the open requisition counts: Accenture has 601 open, 2 opened in 30 days; OpenAI has 155 open, 50 opened in 30 days; Anthropic has 102 open, 15 opened in 30 days; Databricks has 94 open, 16 opened in 30 days; PwC has 85 open, 0 opened in 30 days.
Where Azhar’s cost story implies a growing need for disciplined model operations, our tracker supports it. The data do not show hiring freezing in the face of messy bills. Instead they show employers pushing forward on the plumbing that makes the bills smaller and predictable.
Smaller models and post-training: real, and echoed in hiring
Lambert argues that China is keeping pace on performance without massive parameter counts. “Today, Z.ai announced their GLM-5.3 model, currently only available in the coding plan, coming soon to their API and in two weeks’ time to Hugging Face (open weights).” He adds: “This model looks exceptional, with a somewhat astounding increase in scores.” He highlights the method: “Scaling post-training is all we did for GLM-5.3.”
Ha’s practical lens complements this: “most current baseline models are actually perfectly fine for 99% of everyday work. not everyone needs a state-of-the-art coding model to get things done.” Taken together, these views say the near-term productivity opportunity lies in tailoring and operating models people already have, not in waiting for the next model plateau.
Our tracker is directionally consistent. Research has 185 open, 198 opened and 13 closed in 30 days, across 38 employers. That is real demand. But it is smaller than the combined engine room of Data and Infrastructure. If the most valuable marginal work for many firms is fine-tuning, post-training, prompt and retrieval engineering, routing, and observability, you would expect more hiring in data pipelines and infra. That is what we see. The numbers back the “good enough and well configured” thesis more than a “frontier at all costs” one.
Organizational turbulence vs the jobs signal
Azhar also argues the market misread leadership shifts at one of the largest labs. “Jeff Dean and Sanjay Ghemawat are leaving Google after more than a quarter-century, as you know.” He writes: “This is what the market believed, and Alphabet’s share price dropped 4% in a day. But in our view, this is as much a signal about capital and compute allocation as it is about talent.”
We cannot speak to Alphabet from our data, but the broader research hiring picture does not show a retrenchment. Research roles across employers are 185 open, 198 opened in 30 days. Meanwhile, companies at the center of recent safety debates are still adding roles. Zvi Mowshowitz situates a podcast discussion in a specific context: “This podcast exists in light of recent misalignment and hacking events at OpenAI, Anthropic and UK AISI.” Our tracker shows OpenAI with 155 open, 50 opened in 30 days, and Anthropic with 102 open, 15 opened in 30 days. That is not a collapse in supply or demand for talent at these labs; if anything, it suggests continued investment while the governance and alignment debates continue.
Bottom-up tooling and verification are getting built
Sebastian Raschka shows how quickly verification layers appear when needed: “Substack recently launched its AI detector feature in the UI, which is super interesting.” He adds, “Putting one and one together, I thought it would be interesting to show how an AI detector can be implemented.” He goes further: “I will also use it as a verifier to train a small language model to produce text that avoids detection.” And on use cases: “In practice, such a detector can be used to filter out spammy content, but also to potentially improve your personal writing without turning it into AI-generated text.”
Simon Willison demonstrates the developer tooling wave: “Tool: CORS Chat I built this today (with GPT-5.6-Sol xhigh) to help test Qwen 3.8 27B running in LM Studio on both my M5 MacBook Pro and an NVIDIA DGX Spark.” He notes, “It provides a web UI for exercising an OpenAI-Responses-compatible chat endpoint.” And the UX details are evolving fast: “One fun detail is that it notices SVG images that are being generated and progressively renders them in the chat while the tokens are still streaming in.” In another post he advises, “Don't classify. Hallucinate!” and credits a workflow: “Doug Turnbull has a neat solution.” The method: “Tell the model to output tags without any details of the existing vocabulary, then use vector embeddings against the existing corpus to find the concrete tags that are closest to the ones the model imagined might fit!”
Our figures show small but growing hiring for these layers. Evaluation and safety roles are 39 open, 44 opened and 5 closed in 30 days, across 15 employers. Product and design roles are 71 open, 79 opened and 8 closed in 30 days, across 34 employers. That is a modest share of total openings, but the activity aligns with the kind of verification and tooling Raschka and Willison are building.
Adoption headwinds are sector-specific, and hiring skews big
Ethan Mollick warns that not all sectors welcome AI even when it could help economically: “For those who don’t follow video games, there is constant policing of any AI use among small, indie developers. They are the most resource constrained firms in a field where profits are rare & artistic vision is often compromised, but they are punished more harshly than big devs by their audience.” Our tracker likely undercounts such indie dynamics because it captures roles “across 102 employers” from employer job feeds, and top employers are large enterprises and integrators. Accenture has 601 open, 2 opened in 30 days; Capital One has 177 open, 0 opened in 30 days; Amgen has 173 open, 0 opened in 30 days; Waymo has 106 open, 13 opened in 30 days. The concentration in capital-rich firms helps explain why we see strong data and infra hiring despite pockets of cultural resistance in smaller creative industries.
Mollick also offers a glimpse of applied augmentation in a legacy domain: “As Infocom games are now open source, I had Codex update 1987’s word game “Nord and Bert Couldn't Make Head or Tail of It” for playing now.” That is a microcosm of what many enterprises are attempting with their own legacy assets.
What our tracker says overall
Summing up: the voices this week emphasize cost control, post-training, and practical deployment. Our figures match that emphasis. Data and Infrastructure dominate openings. Research is steady but smaller. Safety and product roles are present and growing from a low base. The net effect is an implementation-heavy market that prizes making existing models useful, observable, and affordable over chasing a single frontier leap.
Our headline jobs-created figure is an indicator 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). Against that long-run calibration, the current snapshot of 3,256 open AI roles across 102 employers, and the tilt toward Data and Infrastructure, point to an economy moving through the build-out phase of good-enough AI. The commentary this week mostly agrees with that picture. Where it does not — for example, concerns about leadership churn or cultural resistance — our data show those forces are not throttling hiring at the employers we track.
What we read
Every quote above is taken verbatim from one of these posts.
- Zvi Mowshowitz: On Dwarkesh Patel's Podcast With Ryan Greenblatt
- Azeem Azhar: 🔮 The curious economics of a $6 AI agent #597, 🔮 The market misread Google’s AI exodus
- Nathan Lambert: GLM-5.3: How Chinese labs keep stride with the frontier
- Sebastian Raschka: Building an AI Text Detector From Scratch
- Simon Willison: CORS Chat, Don't classify. Hallucinate!
- Ethan Mollick: For those who don’t follow video games, there is constant policing of , As Infocom games are now open source, I had Codex update 1987’s word g
- Emily M. Bender: Finally got my author copies of the UK paperback of The AI Con, and th
- David Ha: [unpopular opinion: gemini is a pretty great model
most current baseli](https://bsky.app/profile/hardmaru.bsky.social/post/3mt5xqqg2n224)