Loading the job postings and counting them…
Loading the job postings and counting them…
AI Career Compass
Which AI roles do employers advertise, what do their postings ask for, and what should you learn next? Every number comes from the job postings we track, and every claim links to its evidence.
We started reading these employers’ feeds on 6 Oct 2026. The 30-day window therefore holds one reading so far, and changes over time appear once two full windows exist.
Measured In our tracked sample, forward-deployed and solutions engineers, who help customers put AI to work, are the most advertised role family, in 386 of 1694 AI postings, ahead of AI engineers in 223 and machine learning infrastructure and MLOps engineers in 217.123
Measured The postings we track ask most often for machine learning (asked for in 682 of 1694), Python (657) and communication skills (557).123
Written by an AI analyst (claude-opus-5-5) and checked by code against the numbers it cites. If a number changes, this text is replaced.
Each posting belongs to one role family, chosen from its title and its duties, so the families add up to all 1,694 postings. 582 postings also fit a second family; those overlap and are not counted twice. Open a family for its skills, pay and what to learn next.
| Rank | Role family | Postings | Share | Employers |
|---|---|---|---|---|
| 01 | Forward-deployed and solutions engineers | 386 | 23% | 32 |
| 02 | AI engineers | 223 | 13% | 38 |
| 03 | ML infrastructure and MLOps engineers | 217 | 13% | 41 |
| 04 | Machine learning engineers | 149 | 9% | 30 |
| 05 | AI product, design and program roles | 108 | 6% | 30 |
| 06 | Research engineers | 98 | 6% | 26 |
| 07 | AI safety, security and policy roles |
Search by the name you know. We match role families and count the postings whose titles contain your words.
Out of all 1,694 AI postings. “Asked for” adds three kinds of request: required on its own, required as one of several options (“Python or Java”), and listed as a qualification without saying required or preferred. Preferred skills are counted apart. No skill is required everywhere.
Programming, data, maths and statistics, and domain knowledge
| Skill | Asked for | Required | One of several | Listed | Preferred |
|---|---|---|---|---|---|
| Python | 657 of 1,694 | 157 | 135 | 365 | 95 |
| Software engineering practice | 446 of 1,694 | 119 | 22 | 305 | 71 |
| TypeScript or JavaScript | 157 of 1,694 | 28 | 53 | 76 | 68 |
| C++ | 149 of 1,694 | 20 |
What our reporting covered, next to what employers ask for. Where the two agree or differ is a question to look into, not a verdict.
| Topic | Our stories, 30 days | Postings that ask for it | Only prefer it |
|---|---|---|---|
| AI agentsLatest: Claude Code adds mods that can rewrite how the agent works; GitLab fixes a 9.9-rated flaw in its self-hosted AI gateway | 36 | 305 of 1,694 (18%) | 148 |
| Testing and evaluating modelsLatest: Microsoft’s live transcription model tops an outside test; Published AI accord asks for auditors and board oversight | 5 | 185 of 1,694 (11%) | 131 |
| Training and post-trainingLatest: OpenAI blames Moonshot-linked users for a copying campaign; Reddit will end RSS feeds and its public API over scraping | 7 | 196 of 1,694 (12%) | 101 |
| Inference, chips and computeLatest: Amazon says it has dropped NDAs with data centre officials; US charges a Californian over $300M of AI servers to China | 21 | 196 of 1,694 (12%) | 148 |
| AI safety and securityLatest: Google pauses a bug bounty, flooded by automated reports; AI labs testify under oath in NYC; SpaceXAI stays away |
Pick the role you are aiming for. Its page asks which skills you already have (optional, never saved) and suggests up to three next steps and one portfolio project, each with the requirement behind it.
Measured We began tracking on 6 October 2026, so the window holds one reading so far and shows no trend; many listings are not new, as only 400 of 1694 were first published in the 30 days up to that date.1
Measured The postings we track come from 46 employers that publish their jobs through public feeds, not the whole job market; some of the largest employers run their own job systems and are not included.1
Role families are our best judgement from each posting's title and duties: for 287 of 1694 postings these point two ways, and 150 fit no family.
| 90 |
| 5% |
| 28 |
| 08 | Robotics and autonomy rolesMostly one employer | 69 | 4% | 11 |
|---|
| 09 | AI chips, kernels and compilers roles | 57 | 3% | 12 |
|---|
| 10 | Research scientists | 55 | 3% | 18 |
|---|
| 11 | Data scientists and applied scientists | 54 | 3% | 20 |
|---|
| 12 | Data engineers for AI | 20 | 1% | 10 |
|---|
| 13 | AI tutors and data specialistsLimited sampleMostly one employer | 18 | 1% | 5 |
|---|
| AI postings that fit no familyCounted, not ranked | 150 | 9% | 38 |
|---|
Change over time: not shown. We started tracking on 6 October 2026. A trend needs two full 30-day windows of our own observations, so the first comparison is possible on 4 December 2026.
| 58 |
| 71 |
| 44 |
| Data analysis | 137 of 1,694 | 23 | 22 | 92 | 38 |
|---|
Modelling, training, retrieval, agents, evaluation and other methods
| Skill | Asked for | Required | One of several | Listed | Preferred |
|---|---|---|---|---|---|
| Machine learning | 682 of 1,694 | 154 | 9 | 519 | 185 |
| Large language models | 536 of 1,694 | 127 | 24 | 385 | 257 |
| AI agents | 284 of 1,694 | 65 | 26 | 193 | 140 |
| PyTorch | 178 of 1,694 | 30 | 63 | 85 | 57 |
| Evaluation (evals) | 175 of 1,694 | 33 | 3 | 139 | 126 |
Deployment, infrastructure, reliability, security, monitoring and cost
| Skill | Asked for | Required | One of several | Listed | Preferred |
|---|---|---|---|---|---|
| Backend services and APIs | 370 of 1,694 | 99 | 8 | 263 | 52 |
| Distributed systems | 269 of 1,694 | 59 | 13 | 197 | 38 |
| Cloud platforms (AWS, GCP, Azure) | 223 of 1,694 | 56 | 17 | 150 | 94 |
| System design and architecture | 210 of 1,694 | 46 | 0 | 164 | 23 |
| MLOps and ML platforms | 182 of 1,694 | 51 | 3 | 128 | 90 |
Publishing and leading research
| Skill | Asked for | Required | One of several | Listed | Preferred |
|---|---|---|---|---|---|
| Research publications | 65 of 1,694 | 14 | 2 | 49 | 85 |
| Leading research | 46 of 1,694 | 10 | 1 | 35 | 18 |
| Open-source contributions | 42 of 1,694 | 13 | 1 | 28 | 91 |
Customers, communication, product thinking and leadership
| Skill | Asked for | Required | One of several | Listed | Preferred |
|---|---|---|---|---|---|
| Communication | 557 of 1,694 | 134 | 1 | 422 | 51 |
| Working with customers | 288 of 1,694 | 51 | 0 | 237 | 55 |
| Leading and mentoring | 206 of 1,694 | 64 | 1 | 141 | 51 |
| Product thinking | 126 of 1,694 | 28 | 0 | 98 | 29 |
| Technical writing | 116 of 1,694 | 36 | 1 | 79 | 12 |
AI rules, policy and trust and safety
| Skill | Asked for | Required | One of several | Listed | Preferred |
|---|---|---|---|---|---|
| Risk and compliance | 52 of 1,694 | 15 | 0 | 37 | 52 |
| AI governance and regulation | 24 of 1,694 | 6 | 0 | 18 | 5 |
| Trust and safety | 20 of 1,694 | 4 | 0 | 16 | 13 |
| Public policy | 4 of 1,694 | 2 | 0 | 2 | 3 |
| 51 |
| 87 of 1,694 (5%) |
| 29 |
| AI rules and governanceLatest: OpenAI to watermark ChatGPT text in the EU, with limits; AI labs testify under oath in NYC; SpaceXAI stays away | 29 | 27 of 1,694 (2%) | 8 |
|---|
| Open models and open sourceLatest: Google pauses a bug bounty, flooded by automated reports; Reflection unveils Beam, an open model to rival China’s labs | 17 | 42 of 1,694 (2%) | 91 |
|---|
| Robotics and self-drivingLatest: US Army sets up a command to field autonomous systems; Image AI maker Black Forest Labs released a robot model | 9 | 57 of 1,694 (3%) | 57 |
|---|
| Voice, image and video AILatest: AI child abuse images already top all of 2025, IWF says; Appeals court pauses Minnesota’s AI nude-image ban for xAI | 10 | 83 of 1,694 (5%) | 74 |
|---|
| On-device and local AILatest: Nvidia adds a 64GB DGX Spark, from $4,999 at partners; Analog Devices will buy Alif for $1.35 billion | 4 | 56 of 1,694 (3%) | 45 |
|---|
AI agents Agree
Our reporting We published 36 stories on AI agents in the 30 days to 6 October 2026, such as one on add-ons that can reshape how a coding agent works and one on a serious flaw in self-hosted software for AI workflows.123
Measured In the postings we track, 305 of 1694 ask for AI agent skills, and another 148 list them only as preferred.12
Still uncertain: We have one reading of the postings so far, so we cannot yet tell whether this match will last.
Inference, chips and compute Agree
Our reporting We published 21 stories on inference (running AI models), chips and compute in the 30 days to 6 October 2026, such as one on charges that AI servers were smuggled past export rules and one on secrecy deals over planned data centres.123
Measured In the postings we track, 196 of 1694 ask for inference, chips and compute skills, and another 148 list them only as preferred.12
Still uncertain: Our examples are about chips, trade rules and data centres, while the postings ask for hands-on skills, so the two may not measure the same thing.
Testing and evaluating models Differ
Our reporting We published 5 stories on testing and evaluating models in the 30 days to 6 October 2026, such as one on a live transcription model that topped an outside test and one on a published AI accord that asks for outside auditors.123
Measured In the postings we track, 185 of 1694 ask for skills in testing and evaluating models, and another 131 list them only as preferred.12
Still uncertain: Testing work rarely makes news on its own, so the small number of stories may say more about the news than about what employers ask for.
AI safety and security Differ
Our reporting We published 51 stories on AI safety and security in the 30 days to 6 October 2026, such as one on a reward program for software flaws that paused after a flood of automated reports, and one on AI labs questioned under oath at a city hearing.123
Measured In the postings we track, 87 of 1694 ask for AI safety and security skills, and another 29 list them only as preferred.12
Still uncertain: Safety work can appear in the duties rather than as a named skill, so these posting counts may miss some of it.
Stories are our own checked reporting from 7 Sept 2026 to 6 Oct 2026, counted by their tags. Postings count once each, when they ask for at least one of the topic’s skills. A topic in the news is not proof of hype, and a skill in many postings is not proof of hiring: the two measure different things, so we never divide one by the other.
Developers can now reshape a coding agent from the inside, and share the result with others. The same power is a risk: a mod from an untrusted author can read files and secrets. That makes installing a mod like running any other program.
AI agent platforms add new ways in: here, the text template behind an AI workflow became a route to run code on a server. Only companies that run the gateway themselves need to update. No attacks using the flaw have been reported.
The US bans sales of its most powerful AI chips to China, and cases like this test how well those rules hold. The government says the servers were routed through other countries to hide where they were going. Lui has not been convicted, and no defence lawyer was listed when Bloomberg reported.
Secrecy deals that stopped local officials from telling residents about planned data centres have fuelled the backlash against them. Garman wrote that over 100 data centre moratoriums are being considered across the US. The change is Amazon’s own description of its practice; it gave no date or details.
Voice agents need to hear users quickly and correctly before they can answer. On Artificial Analysis’s streaming leaderboard, which we checked, the model has the lowest error rate of 32 models shown. Microsoft’s speed figures are its own.
The public text shows what was promised: four layers of internal and outside checks, but no deadlines, no reports to the government and no penalties. Senator Mark Warner said the president was telling the companies to regulate themselves. The accord itself says its steps may one day become law.
AI tools make it cheap to send bug reports, but each report still takes an expert’s time to check. When most are wrong, a program that pays for real flaws stops working. Google is one of several programs to pull back this year.
For the first time, a city made AI labs answer questions in public and under oath. The labs backed safety in general words, while the Speaker pushed them to put numbers on the worst risks. The 10 bills are still proposals, and nothing was decided on Monday.
How every number is counted, the groups of employers and the safeguards: our method. Career data is refreshed by a script and goes live only after the editor reviews it in a pull request.