Where the gap is today
of technical postings at leading AI firms ask for AI agents. No USEK course description names them.
is the length of task the best AI model now completes on its own at a 50% success rate, up from 3.4 hours a year earlier and 20 minutes two years ago (METR, via Epoch AI).
leading-firm technical postings ask for quantum computing, blockchain or AR/VR. USEK teaches blockchain and AR/VR; most peers teach quantum.
measured for the local market: the public Gulf job boards reached held only 3 technical postings.
Which parts of the work AI already does
Best score by any model on public benchmarks, now and one and two years ago. Status: share of the benchmark's ceiling the best model reaches (80% or more: AI does most of it; 50 to 80%: a large part; under 50%: AI assists). Where AI does most or a large part, courses should shift from doing the task to specifying, directing and checking AI's work.
| Work measured | Two years ago | One year ago | Now | Status |
|---|---|---|---|---|
| Fixing bugs in real software projectsSoftware engineering | not measured | 0.742025-08 | 0.832026-04 | AI does most of it |
| Coding tasks inside an AI coding toolSoftware engineering | not measured | not measured | 0.602026-10 | AI does a large part |
| Writing scientific research codeSoftware engineering | 0.272024-10 | 0.462025-06 | 0.682026-11 | AI does a large part |
| Training and tuning AI models (ML engineering)AI and data | not measured | not measured | 0.452026-09 | AI assists |
| Finding and exploiting security flawsCybersecurity | not measured | 0.142025-10 | 0.742026-04 | AI does a large part |
| Producing CAD models from a descriptionEngineering design | 0.482024-10 | 0.742025-04 | 0.742025-04 | AI does a large part |
| Operating a computer to finish office tasksGeneral digital work | 0.052024-09 | 0.632025-09 | 0.722026-02 | AI does a large part |
| Operating a computer, harder 2026 tasksGeneral digital work | not measured | not measured | 0.312026-07 | AI assists |
| Professional deliverables, judged against industry experts (0.5 = parity)Knowledge work | not measured | 0.442025-08 | 0.502025-12 | AI matches experts in about half of comparisons |
| Completing real paid freelance projects end to endKnowledge work | not measured | 0.022025-09 | 0.412026-11 | AI assists |
| Answering expert-level science questionsScience | 0.622024-09 | 0.872025-07 | 0.962026-10 | AI does most of it |
Source: Epoch AI, Capabilities and benchmarking, CC BY 4.0. The pairing of each benchmark with a skill is ours. Benchmarks test set tasks; they do not mean a whole job is automated.
Each technology area, on every frontier
USEK: course entries in the 2025-26 catalogue naming the area. Peers teaching: of 9 universities with archived catalogues, how many named it in 2021 and in 2026. MOOCs and labs: of 8 course catalogues (DeepLearning.AI, Hugging Face, fast.ai, CS50 AI, Anthropic, OpenAI, Google, NVIDIA), how many name it today. Leading-firm demand: share of 1,356 technical postings at Anthropic (287), OpenAI (459), Databricks (532), Scale AI (78) naming it.
| Area | USEK | Peers teaching 2021 → 2026 | MOOCs and labs now | Leading-firm demand | Position |
|---|---|---|---|---|---|
| Cloud and DevOps | 13 courses | 6 → 8 | 3 of 8 | 38.0% |
Taught at USEK and in demand |
| Data engineering and big data | 9 courses | 5 → 6 | 2 of 8 | 30.5% |
Taught at USEK and in demand |
| Generative AI and LLMs | 3 courses | 2 → 8 | 7 of 8 | 27.4% |
Taught at USEK and in demand |
| AI agents | None | 0 → 3 | 3 of 8 | 22.3% |
In demand, not taught at USEK |
| Machine and deep learning | 18 courses | 9 → 9 | 6 of 8 | 19.0% |
Taught at USEK and in demand |
| Human-computer interaction and UX | 10 courses | 7 → 8 | 0 of 8 | 15.2% |
Taught at USEK and in demand |
| GPUs and high-performance computing | 3 courses | 7 → 8 | 2 of 8 | 9.9% |
Taught at USEK; some demand |
| Cybersecurity | 21 courses | 9 → 9 | 2 of 8 | 6.1% |
Taught at USEK; some demand |
| Responsible AI and tech ethics | 1 courses | 3 → 6 | 4 of 8 | 5.2% |
Taught at USEK; some demand |
| AI-assisted software work | None | 0 → 3 | 1 of 8 | 5.1% |
Some demand, not taught at USEK |
| Robotics and autonomous systems | 13 courses | 9 → 9 | 4 of 8 | 2.4% |
Taught at USEK; some demand |
| IoT, edge and embedded | 8 courses | 7 → 7 | 1 of 8 | 0.7% |
Taught at USEK; little leading-firm demand |
| Computer vision | 2 courses | 8 → 9 | 3 of 8 | 0.5% |
Taught at USEK; little leading-firm demand |
| Reasoning models | None | 0 → 1 | 1 of 8 | 0.2% |
Not taught; little leading-firm demand |
| Quantum computing | None | 5 → 9 | 1 of 8 | 0.0% |
Not taught; little leading-firm demand |
| Blockchain and Web3 | 7 courses | 3 → 6 | 0 of 8 | 0.0% |
Taught at USEK; little leading-firm demand |
| AR, VR and spatial computing | 10 courses | 7 → 7 | 1 of 8 | 0.0% |
Taught at USEK; little leading-firm demand |
| Green and sustainable computing | None | 1 → 2 | 1 of 8 | 0.0% |
Not taught; little leading-firm demand |
All 9 peers with archived catalogues now name quantum computing, up from 5 in 2021, yet none of the leading AI firm postings read asks for it. AI agents show the opposite: 3 of 9 peers name them, while about one in five leading-firm technical postings asks for them. The watch tracks both, so a program change can follow demand, not just what peers do.
Built to be re-read every day
The figures above are one reading. Each source below is meant to be re-read daily; scheduling those runs on USEK's own systems is the next step.
- AI capability: benchmark results, re-read daily as new models are scored.
- Leading-firm demand: public job boards, re-read daily, with every technical posting re-scored after each firm's repeated boilerplate is removed.
- MOOCs, labs and university catalogues: re-read daily for new or changed courses.
- A change that touches a USEK course is flagged the same day on the dean's sheet.
Local market demand. Public Gulf job boards hold too few technical postings to measure. The best source is USEK's own career services postings and employer partners, plus regional job sites with permission. Also: name-level matching undercounts lab courses (for example Anthropic's "subagents" and "agent skills"); the next version reads course content, not names.