Loading the job postings and counting them…
Loading the job postings and counting them…
AI Career Compass · method
Every rule behind the numbers on the careers pages: where the postings come from, what counts, how roles, skills and pay are read, and what we refuse to claim.
We read the public job feeds that employers use for their own careers pages (Greenhouse, Lever and Ashby), the same lists their pages show. We read no job boards that need an account, and no copies of other sites’ lists. We send one request per second, and our name and address with each one: SiliconAINews-CareerCompass/1.0 (+https://siliconainews.com/careers/method).
The 47 employers were chosen by hand in five groups: AI companies we report on, and other well-known AI employers whose careers pages use one of these feeds. Some of the largest technology companies run their own job systems, so they are not in the sample. The sample leans towards AI labs and fast-growing companies in the United States. It is a tracked sample, not the whole job market.
These pages give an overview of the market. They name no employer and show no single job ad: ads close and change, and one company’s ad is not the market. Every number counts many postings together.
10 employers · 575 open AI postings, counted once
13 employers · 301 open AI postings, counted once
9 employers · 426 open AI postings, counted once
5 employers · 176 open AI postings, counted once
10 employers · 216 open AI postings, counted once
Titles mean different things at different employers, so each posting is placed by its title and its duties together. Every posting has one primary family, so the families add up to the total; a second family it also fits is recorded but never counted twice. When two families fit almost equally, the posting is marked uncertain. Postings that fit no family are counted and shown, not hidden. Taxonomy version 1.
How exact is this? On 7 October 2026 a second reader, an AI agent that saw 60 randomly chosen postings and the family definitions but not our answers, placed each one itself. It chose our primary family for 38 of the 60 (63%), and a family we also recorded for 47 (78%). Families are useful groupings, not exact facts: read the counts as approximate.
| Family | What it does | It also covers |
|---|---|---|
| AI engineers | Builds products and systems on top of AI models. | Language-model features, agents, retrieval, evaluation, and the software around them; not only "wrappers" or chatbots. |
| Machine learning engineers | Builds, trains and ships machine learning models. | Ranking, recommendation, perception, forecasting and language models alike; not only training foundation models. |
| ML infrastructure and MLOps engineers | Builds the systems that train, serve and run AI models. | Training clusters, inference and serving, ML platforms, GPUs and the tooling other engineers use; not only deploying models. |
| Research scientists | Develops new methods and studies how models behave. | Model capabilities, training methods, interpretability, safety and science applications. |
| Research engineers | Turns research ideas into experiments and working code at scale. | Training runs, experiment tooling, evaluations and the engineering behind research results. |
| Data scientists and applied scientists | Uses data, statistics and models to answer questions and improve products. | Experiments, analysis, forecasting and applied modelling. |
| Data engineers for AI |
Next to every count we show how many employers it comes from. When one employer has 50% or more of a group’s postings, the page says so, because the numbers then mostly describe that employer.
Each skill has a domain and, separately, a status in each posting. The status comes from the heading it sits under (“Requirements”, “Preferred qualifications”) and from its own words (“is a plus”, “must have”).
Every count is written as “X of Y postings”. No skill is required everywhere. Domains: Foundations, AI and application methods, Production and operations, Research, Working with people and products, Safety, policy and governance.
Duties are counted from the responsibilities lists. Years of experience are the smallest number of years a requirement states. Degrees are the lowest degree a requirement states, with a note when the posting accepts equivalent experience.
Level comes from the title only (for example “Senior” or “Staff”). A posting whose title states no level stays “not stated”: we never guess a level from years of experience.
The default window is the 30 days up to the last refresh. Each page shows the window’s start, its end and the data cutoff. The main count is postings listed at least once in the window. We also keep each posting’s own publication date and the day we first saw it: an old posting we read for the first time is not a new opening, so the first reading of each feed is marked as such.
Data older than 14 days is labelled stale on every careers page. A feed that fails keeps its earlier data, marked as not re-checked, and is named on the pages.
A change compares two equal windows with the same filters, the same role definitions and the same feeds, and both windows must pass the safeguards. We show the counts, the change in share in percentage points, and the relative change; when the earlier count is zero, we say “newly observed” instead of a percentage. We started tracking on 6 Oct 2026, so the first comparison is possible on 4 Dec 2026.
These are rules we set for the product, not tests of statistical significance. We show no confidence scores, because we have none we could defend.
Remote never means “from anywhere” unless the posting says so. We keep the places a remote posting allows (for example “Remote, United States”), and say when a posting does not name them.
We read pay ranges exactly as advertised, from the feed’s own fields or the posting’s text. We keep the currency, the period (a year, an hour…), the pay type (base pay, on-target earnings, total pay), the place and the employment type. A range whose period the posting does not state is shown but never pooled. A bare “$” counts as US dollars (or Canadian, Australian…) only when every place in the posting is in that one country.
Ranges are pooled only when all of these match, and only with at least 10 of them. We report the median of the advertised ranges’ midpoints and the middle half of those midpoints, with the number of ranges and the share of postings that show pay. This is advertised pay: it is not what people earn, and employers that publish ranges may differ from those that do not. Nothing is converted between currencies.
For each topic we count our own checked stories in the 30 days up to the data cutoff, by their tags, and the postings that ask for the topic’s skills. The two numbers measure different things. We never divide one by the other, and never call a topic hype: a topic in the news may matter without many postings, and many postings do not prove hiring.
Code counts everything. An AI analyst then writes the market snapshot and the notes on news and postings, from a list of the measured numbers and our own stories. Its instructions begin:
You are AI Career Compass, the evidence-led career analyst for Silicon AI News. Explain what the supplied dataset supports in clear, practical language. Use general knowledge only to explain concepts or offer explicitly labelled advice. Never use it to invent current market findings.
You tick the skills you have; nothing is sent or saved. A posting needs a skill from you when it asks for it and you have neither it nor an alternative the posting accepts (“Python or Java”). We suggest up to 3 skills that the most postings need, fewer if you have little time, each with the requirement behind it, a way to show it, and our background lessons where we have them. The portfolio project is a template for the role. This is advice: it cannot promise interviews, jobs or pay.
| Builds the datasets and pipelines that models are trained and tested on. |
| Data collection, cleaning, labelling pipelines, data quality and training-data infrastructure. |
| AI product, design and program roles | Decides what AI products should do, designs them and runs the programs that ship them. | Product managers, designers and technical program managers working on AI products or research. |
|---|
| Forward-deployed and solutions engineers | Works with customers to build and deploy AI systems in real settings. | Forward-deployed, solutions, customer and deployment engineers, and applied AI roles that face customers. |
|---|
| AI safety, security and policy roles | Studies and reduces the risks of AI systems, and works on the rules around them. | Alignment and safety research, red teaming, AI security, trust and safety, policy and governance. |
|---|
| AI chips, kernels and compilers roles | Makes AI run fast on hardware: chips, GPU kernels and compilers. | Chip design and verification, GPU and accelerator software, kernels, compilers and performance engineering. |
|---|
| Robotics and autonomy roles | Builds AI that senses and acts in the physical world. | Robot learning, perception, planning and controls, simulation, and self-driving systems. |
|---|
| AI tutors and data specialists | Writes, labels and checks the data that trains and tests AI models. | AI tutors, annotators and raters, and subject experts who write and grade model answers. |
|---|
Before anything is shown, code checks the analyst’s report: every number must come from a metric the sentence cites, every cited number and story must exist, no employer may be named, statements from sources must cite them, a note about a topic may cite only that topic’s numbers and stories, and promises of jobs or pay, links and markup are refused. A report is shown only for the exact view it was written for, and only while the numbers it cites are unchanged; otherwise the page shows a summary written by code. Text from postings, news and searches is treated as data, never as instructions.
The website never calls a model and holds no key. The refresh job runs the analyst; the current reports were written by claude-opus-5-5. Prompt version 2. Like our stories, the analyst’s text is generated by AI under the editorial rules of Silicon AI News and checked against the data it cites.