DeepSeek Hires 150: What Are AI Companies Really Missing?

I compared the ~150 positions DeepSeek released this round with hiring pages from several AI companies I organized over the past two weeks using Excel, Grok, and Kimi K3, actually running through job categorization. Conclusion first: it's worth viewing this expansion as a signal, but not everyone is suitable to apply. Changes in job structure are more informative than the numbers themselves. In these public roles, I didn't see many pure research titles; server-side, pre-training data, AI search, Agent Harness, Agent Infra, and frontend/client-side ranked very high. People interested in backend, data, evaluation, permissions, and toolchain directions should pay attention; those just wanting an "AI talent" label might need to reconsider.

I split the jobs into three piles. I copied job names, departments, and requirements from recruitment info into Google Sheets. I just started trying Google Sheets these few days, unfamiliar with filtering, initially splitting job names into multiple rows, making it look like positions were missing. After switching to one row per job, I backed up in Excel, then had Claude Fable 5 classify them by distance from the core model. I wrote the prompt three rounds; the first version classified all cross-disciplinary AI technical talents as product, which I felt was wrong. The second version added factors like data governance, permissions, evaluation, and deployment involvement, resulting in more stable outcomes.

Roughly three piles emerged. Data & Search includes Pre-training Data Engineers, AI Search Algorithms, and Architecture. Operations & Tools includes Agent Harness, Agent Infra, and Server-side Development. Delivery & Experience includes Frontend, Client-side, and Cross-disciplinary Technical Talents. Agent Harness is closer to an agent execution framework, not responsible for the model itself, but for connecting the model to tools, data, permissions, logs, evaluation, and rollback. The model says a sentence; the Harness must make it actually query databases, call interfaces, retry on failure, and leave auditable traces.

Most surprising was Pre-training Data Engineers being placed so prominently. Many beginners think data roles aren't sexy, but from my view, this looks like building long-term barriers. Model capabilities diffuse, but data cleaning, retrieval quality, evaluation sets, and user feedback loops are hard to replicate. I had a similar feeling writing a data center supply chain dashboard before; abstract risks ultimately fall down to electricity, cooling, fields, and on-site processes. Job titles don't lie.

The more automated, the more system workers are needed. Does stronger AI mean companies need more people? Look at this in two layers: headcount may not increase, but types change. Previously, a few geniuses could drive model breakthroughs; now, to build products, search, Agents, and platforms, you need engineering pipelines. I saw in recruitment info that Cui Tianyi mentioned his department is still severely understaffed, interviewing every day, recruiting for over a month without filling enough spots. This detail is more real than the number of positions. Hard-to-fill roles are often those requiring simultaneous understanding of models, systems, data, and boundaries.

I also tried pushing table notifications to Slack via viaSocket. viaSocket is new to me, good for quickly chaining daily apps, but far from a universal socket. Inconsistent fields and classification drift meant manual verification was still needed at the end. Without someone sorting complex logic, the more tools connect, the messier it gets.

My judgment is that for some time, AI company hiring will shift from "who got the model running" to "who can integrate the model into real business." Models will become cheaper and more general; scarce assets are data moats, search quality, Agent execution frameworks, evaluation audits, and edge delivery. Looking forward, people who can handle models, permissions, logs, rollback, and costs simultaneously will be more sought after than those who can only write prompts.