We read 315 AI and machine-learning job postings published between July and September 2026 and counted what each one asked for. The short version: the job is now called AI engineer or plain software engineer far more often than ML engineer; "agents" and "evals" appear in more ads than PyTorch; and the words that show up most are not tools at all — they are production and ownership. Degrees are mentioned in one ad in eight. Certifications in one in fourteen.
This is our own count, not a quoted survey. The method, the sample and the numbers are all below, and the counts file the article reads from is published with it.
Where did the 315 postings come from?
From three open sources, collected on 19 September 2026: the Hacker News "Who is hiring?" threads for July, August and September (753 postings), Remotive's public remote-jobs feed, and Arbeitnow's public job-board API (1,700 postings, largely European). That is 2,470 postings in total. We kept the 315 where the title named an AI, ML, LLM, NLP, computer-vision, data-science or MLOps role, or the body named such a role explicitly. A fintech that mentions "AI" in its boilerplate did not count.
Each skill was counted once per posting if it was mentioned at all. A posting that says "Python" five times counts once. So every percentage below means the share of the 315 ads that mention it.
Two honest limits. Hacker News postings are short — often a paragraph — so they under-report tools compared with a full job description. And Australia is nearly absent: three of the 315 ads name it. The Australian market lives on Seek and LinkedIn, which we did not sample. The skills transfer; the salary and location patterns may not.
What are the roles actually called?
"Software engineer" and "AI engineer" — not "machine learning engineer". Of the 315 ads, 62 were titled some form of software engineer (full-stack, backend) at a company building an AI product, 48 said AI engineer, 23 said founding engineer, and 16 said forward deployed engineer. Only 13 said ML engineer and 9 said data scientist. Two asked for a research scientist.
That is the first finding, and it changes what to study. Most AI hiring in this sample is hiring for software engineering with models in the loop, not for model building.
Which programming languages do AI job ads ask for?
Python, then TypeScript — and then a long gap.
| Language | Ads mentioning it |
|---|---|
| Python | 119 (38%) |
| TypeScript / JavaScript | 60 (19%) |
| SQL | 32 (10%) |
| Java | 14 (4%) |
| C++ | 12 (4%) |
| Rust | 8 (3%) |
TypeScript being second is not an accident. Agent products are web products: a chat surface, a tool-calling backend, a dashboard. An "AI engineer" in this sample is as likely to be writing a Next.js route as a training loop.
Do employers still ask for PyTorch and TensorFlow?
Rarely, in these ads. PyTorch appears in 16 postings (5%), TensorFlow in 9 (3%), Hugging Face or transformers in 8 (3%), scikit-learn in 3. Compare that with the LLM stack:
| LLM-era skill | Ads mentioning it |
|---|---|
| Agents / agentic systems | 123 (39%) |
| LLMs, any mention | 108 (34%) |
| Evals / evaluation | 73 (23%) |
| Anthropic or Claude | 49 (16%) |
| RAG / retrieval | 34 (11%) |
| OpenAI API | 28 (9%) |
| Model Context Protocol (MCP) | 28 (9%) |
| LangChain / LlamaIndex / LangGraph | 27 (9%) |
| Fine-tuning | 24 (8%) |
| Prompt engineering | 17 (5%) |
| Vector databases | 16 (5%) |
"Agents" is mentioned in eight times as many ads as PyTorch. "Evals" — the discipline of measuring whether a model-backed system actually works — is in nearly a quarter of ads and was barely a job-ad word two years ago. MCP, a protocol that did not exist before late 2024, is now named as often as OpenAI's API.
One vendor note, with the caveat that Hacker News skews towards startups: Anthropic or Claude was named in 49 ads, OpenAI in 28, Gemini in 11.
Which cloud and infrastructure skills come up?
AWS first, and monitoring more than any single cloud.
| Infrastructure skill | Ads mentioning it |
|---|---|
| Monitoring / observability | 54 (17%) |
| AWS | 51 (16%) |
| Docker / containers | 32 (10%) |
| CI/CD | 31 (10%) |
| Postgres | 32 (10%) |
| MLOps tooling (MLflow, SageMaker, Vertex) | 26 (8%) |
| Azure | 24 (8%) |
| Kubernetes | 24 (8%) |
| Terraform / infrastructure as code | 21 (7%) |
| GCP | 20 (6%) |
The data-platform names — Spark and Databricks (4%), Snowflake and BigQuery (5%), Airflow (1%), dbt (1%) — are thin here because this sample is AI-product hiring, not data-platform hiring. Postgres beats all of them, which again says: these are application jobs.
What do the ads ask for that isn't a tool?
Shipping. Production or shipped appears in 124 ads (39%) — more than Python. Ownership or autonomy is in 117 (37%). Customers or product sense in 89 (28%). System design or architecture in 72 (23%). Communication in 68 (22%). Security in 51 (16%). Testing in 38 (12%).
Put the two halves together and the sentence most ads are writing is: you have shipped something to real users and you can be left alone to do it again.
How much do degrees, certifications and portfolios matter?
Less than a public body of work, on the evidence of what employers bother to write down. A degree of any kind is mentioned in 37 ads (12%); a PhD in 9 (3%); a certification in 21 (7%). A portfolio, GitHub profile, open-source work or side projects: 54 ads (17%).
Only 67 ads (21%) state a number of years of experience at all. Where they do, the median asked is five years; 26% ask for three or fewer, 58% for five or more. The other four ads in five are silent, which usually means the bar is "show us what you built".
Where are the jobs, and are they remote?
Remote in 144 ads (46%), on-site in 101 (32%), hybrid in 72 (23%) — some ads say two of these. Equity is mentioned in 81 (26%), a salary or range in 46 (15%), visa sponsorship in 27 (9%).
By region named: Europe 137, the United States 98, the United Kingdom 64, Canada 18, India 9, Australia 3. The European weight is Arbeitnow's; the US weight is Hacker News's. Neither is the world.
What should you study, if this is the market?
Build one deployed, model-backed application end to end, and be able to explain how you know it works. That single project covers the six most-mentioned things in the sample: Python or TypeScript, an LLM with tool use (the "agents" ask), an eval set that measures it, a Postgres table under it, a container it ships in, and monitoring after it ships.
Every twelve-week bootcamp at Square 1 is built as six of those, one every fortnight, deployed with a live URL. The Agentic AI, LLM Evals and Reliability, RAG and Search Engineering and LLMOps bootcamps map directly onto the four biggest asks above. If you only want the shortest bridge, the on-demand Evaluating AI Systems and MCP and Tool-Using Apps courses cover the two skills that grew fastest.
Can I check the numbers?
Yes. The counts file this article was written from is published at /research/data/ai-job-ads-2026-09.json: every figure above, the sources, the date collected and the classification rule. The collection and analysis scripts live in the site's repository under tools/job-ad-sample/; run them and you get the current month's threads. We will repeat the count each quarter and publish the change.