Methodology

Where the data comes from

Every night we read the public job boards of the 80 companies listed below, straight from their applicant-tracking systems (Greenhouse, Ashby or Lever). We store only derived facts — title, location, remote status, the posted base-pay range, a role family and skill tags — and link every role to the company's own posting. We do not copy job descriptions.

This measures posted demand: what companies advertise. It is not a count of hires, and some postings are evergreen or duplicated. A posting that disappears from a board that loaded successfully is recorded as closed that day; a board that fails to load closes nothing, and neither does one that comes back empty while we still list roles from it (an empty board is far more often a moved or changed feed than a company closing every role). A company whose board has failed for more than 3 days drops out of every number until it loads again.

Reposts and general applications

A company that lists the same title in the same location more than once has one opening, so every number counts it once and keeps the date we first saw it (44 postings folded today). Titles are compared ignoring case and punctuation only, so “Software Engineer, Data Pipeline” and “Software Engineer, Onboard Maps” stay two openings. “Don’t see your role? Apply here” and other general-application postings are not openings and are left out entirely (13 today). The open-roles list shows every posting as the company lists it.

How long postings stay open, and how long they take to close, are published once there are 30 days of history (6 so far).

Pay

Pay is the annual base range a posting publishes (most US postings must, under pay-transparency laws). It is not total compensation: equity and bonus are not in these ranges. A band is the median of the published lower bounds and the median of the published upper bounds, shown only when at least 10 postings in the group publish pay; every band is shown with the share of postings that disclose pay. For total compensation by level, Levels.fyi is the reference.

Role families and skills

Role families are assigned from the job title by an ordered set of rules (first match wins), so “Research Engineer, Evaluations” is a research role and “Sales Engineer” is go-to-market. Skill shares count technical postings only, from a fixed vocabulary. A language model (Claude Haiku 4.5) read 4,964 of the 4,982 technical postings and tagged them from that vocabulary only; the other 18, new or edited since the last nightly batch, carry keyword-rule tags until the model reads them. The keyword rules still tag every posting as a baseline, and the monthly audit starts where the two disagree. A company's own name and product are never counted as skills, and ambiguous words (such as “Go”) are only counted in a language context.

How precise the tags are

Every month we draw 50 technical postings at random and check each tag and role family against the posting. A tag is right when the posting's responsibilities or requirements mention the skill; a mention only in the company's description of itself does not count. A skill whose tags are right less than 90% of the time (once at least 5 were checked) is not reported anywhere on the site until the next audit. A skill with a wrong tag in the sample and fewer than 30 checks gets more, so that one tag cannot swing the verdict: postings carrying it are drawn at random to bring it up to 30 checks, each checked for that tag alone. Audit of 2026-09: role families right 44 of 50; extra checks, 18 more for LLMs and 19 more for Agents.

Companies covered

Boards are read as the company publishes them. xAI's board also lists roles at X (the social network) and at its data centers; they are counted, under their own role families.

How we pick learning resources

The learning map is a short list, not a directory. Every pick has to meet all of these:

• A primary source or a practitioner. The author built the thing, runs it in production, or teaches it from that experience.
• At least one independent endorsement, linked and dated in our content files: a Hacker News discussion with 100 points or more, a course syllabus that assigns it, a GitHub repository with 1,000 stars or more (shown with the date we counted), or repeated recommendations in the community the role belongs to. We never show third-party ratings: no star scores, review counts or view counts.
• Fresh. A resource about a tool or a library is at most 18 months old; one about principles has no age limit.
• Reviewed one by one. A pick reaches this site only after it has been read and approved, on a date the content records.
• “I used this” only when it is true. The badge marks a resource the site's author actually worked through.

A step has at most 3 links: the best free resource, a paid one only when it is clearly better, and a project or checkpoint. A video is its author's own upload, played from the recommended segment; a book names its chapters.

Affiliate links

Some picks may link through an affiliate program: if you buy through one, we earn a commission, at no extra cost to you. Every such link says so right beside it — “We earn a commission if you buy through this link.”, or for Amazon, “As an Amazon Associate I earn from qualifying purchases.” — and is marked as sponsored for search engines.

Ranking never looks at commissions. The best pick comes first whether or not it has an affiliate program; whether it has one is looked at only after it has been picked. The link text is always the resource's own title, a platform's name is plain text beside it, and we show no platform logos, book covers, prices, star ratings or review counts. The monthly email links only to this site's pages, never to an affiliate link.

Citing these numbers

Cite as “AI Career Atlas (aicareeratlas.com), AI Hiring Index, <month>” and link the page you quote.