Research
We study the work we teach.128 pieces, written to be checked.
Every programme is shaped by what employers ask for and how the field is changing. We publish what we find, with the sources and the limits, so you can check it. Then you learn to do the same.
128papers and reports
13hours of reading in all
Research is the fifth step
- Course
- AI Tutor
- Competitions
- Community
- Research
Latest research · Industry & Society
Can a non-technical founder build an AI product in 2026?
Yes, to a working prototype and first customers with AI coding tools; not safely alone once sign-in, payments, sensitive data, security and scale are at stake. Where vibe-coded products break, when to bring in an engineer, and why to validate demand first. Startups do hire for it: 23 of 315 AI job ads we counted were titled founding engineer.
All research · 128 piecesLatest first
Claude Code or Codex: which coding agent should an engineering team use? Use both and choose per task: Claude Code suits interactive work in your repository, Codex suits well-specified tasks you delegate and review later. Of 416 software engineering postings collected on 29 September 2026, 27 named Claude Code and 12 named Codex, usually alongside other tools. How the working styles differ and the team practices that matter more than the tool. 6 min read Do software engineering jobs now expect you to use AI coding tools? About one in six say so in writing: 67 of 416 software engineering postings named an AI coding tool, led by Claude Code, Cursor, Codex and Copilot. Those ads ask even more often for production experience, system design and ownership. How interviews are changing and what employers mean by good. 5 min read How can an operations manager use AI without writing code? Give chat assistants the reading and drafting work in an operations week: first drafts of SOPs, triage of inbound requests, reporting summaries, vendor comparisons and forecast stress-tests, with a person checking every output. Then add one no-code automation with an approval step. What never to paste into an AI tool, who approves what, and how to measure hours returned. 6 min read How do you build a production app on the Claude API? Put the API behind your own server, version the system prompt, use structured output and tool use, stream responses, retry the right errors, cache stable prompt content, log cost per request and run evals before every change. Of 315 AI job ads collected on 19 September 2026, 49 (16%) named Anthropic or Claude. 6 min read How do you build a voice AI agent that survives a real phone call? A voice AI agent is a streaming loop of telephony, speech-to-text, a language model with tools and text-to-speech; the hard parts are latency, interruptions, turn-taking, testing on recorded calls and handoff to a human. Voice or speech appeared in 8 of 417 AI and ML job postings we collected on 29 September 2026, about 2%: a niche specialism. 6 min read How do you delegate coding tasks to an AI agent like Codex without losing control? Decide what done means before the agent starts: a five-part spec, a failing test or reproduction it must turn green, strict sandbox and approval settings, an AGENTS.md, and a review of every diff against the spec. How to run several tasks in parallel, and what never to delegate. Of 416 software engineering postings collected on 29 September 2026, 67 named an AI coding tool and 12 named Codex. 6 min read How do you get a job in AI evaluation (evals) in 2026? Evals are a skill employers hire for inside other roles, not yet a job title: 73 of 315 AI job ads (23%) asked for evals, but only 2 of 3,610 postings had evaluation in the title. Which roles ask for it most, the skills involved, and how to prove you can do it. 6 min read How do you get into robotics and physical AI as a software engineer? A software engineer moves into robotics by adding perception, control and testing from simulation to real hardware. In 33 postings at robotics and physical-AI employers (29 September 2026, including non-software roles), 58% said on-site, 42% mentioned a degree, 18% a PhD, and C++ (27%) sat almost level with Python (30%). 6 min read How do you start using Claude Code on a real codebase? Start small and supervised: one scoped change in code you know, a plan before any edit, a CLAUDE.md with your test command and conventions, tight permissions, tests as the contract and a review of every diff. A practical first two weeks for one developer, and when to add hooks, skills and MCP. Claude Code was named in 27 of 416 software engineering postings collected on 29 September 2026. 6 min read How is an AI product manager different from a regular product manager? An AI product manager does the core PM job for features that are sometimes wrong: defining quality with evals, trading off cost, latency and accuracy, and owning guardrails. Of 109 product manager postings collected on 29 September 2026, 71% mentioned AI, 62% asked for ownership and 27% mentioned agents, but only 11% said evals. 6 min read Is data science still a good career in 2026, now that AI writes the code? Yes, for the part of the job that decides things: AI has made the code cheap, not the judgement. But startups building AI hire far fewer data scientists than AI engineers: 9 of 315 AI job ads were titled data scientist, against 48 titled AI engineer. What the 17 data scientist postings we counted on 29 September ask for, and how to become hireable. 6 min read Is prompt engineering still worth learning for developers in 2026? Yes, as an engineering skill rather than a job title: prompt work now means system prompts, structured output, examples, tool descriptions and evals that check each change. Of 315 AI job ads collected on 19 September 2026, 17 (5%) named prompt engineering, against 73 (23%) for evals and 123 (39%) for agents. 6 min read What can you actually build with multimodal AI in 2026? Document and form extraction, screenshot and UI understanding, image inspection, chart reading, audio transcription feeding a text model, and video summarisation. What each looks like, where multimodal models fail (small text, counting, spatial reasoning, confident wrong extraction), how to test on your own images, and the cost and privacy questions. 6 min read What does a forward deployed engineer do, and how do you become one? A forward deployed engineer builds software inside a customer's organisation until it works for them, then hands it over running. What the job involves day to day, what 29 FDE job ads ask for (customers in 76%, against 28% of AI ads), the pay the ads print, and the route in. 6 min read What does an AI sales engineer do, and how do you become one? An AI sales engineer runs technical discovery, builds the demo and proof of concept on the buyer's data, and wins the security and architecture review before the deal closes. Of 49 sales and solutions engineer postings collected on 29 September 2026, 39 (80%) mentioned AI, and 59% asked for customer or product sense. How the role differs from a forward deployed engineer, and the route in. 6 min read What is a GTM engineer, and how do you become one? A GTM engineer builds the systems a sales and marketing team runs on: account data, enrichment, signals, AI-drafted outbound, CRM routing and attribution. Of 3,610 postings on engineering-heavy job boards, exactly one used the title, because GTM roles are posted on LinkedIn and sales boards instead. What the job involves, the skills, and the route in. 6 min read What is an AI automation engineer, and is it a real job in 2026? An AI automation engineer replaces manual business processes with automations that include AI steps, in n8n, Make, Zapier and Python. The work is real; the title is rare: about ten of 2,610 job postings used it. Where the work is actually posted, the skills, and how to prove you can do it. 5 min read What is an analytics engineer, and how is it different from a data analyst? An analytics engineer builds the tested, documented data models and metric definitions that everyone else's numbers come from; a data analyst uses them to answer questions. All 13 analytics engineer postings we counted on 29 September 2026 asked for SQL, dbt and a cloud warehouse, against 3 of 10 data analyst postings for dbt. What the job involves and how to move into it. 6 min read What is the Model Context Protocol (MCP), and should developers learn it? MCP is an open protocol, introduced by Anthropic in 2024, that lets AI applications connect to tools and data through standard servers exposing tools, resources and prompts. What it is, when to build a server rather than call an API directly, and the security risks. Of 315 AI job ads collected on 19 September 2026, 28 (9%) named MCP. 6 min read When should you fine-tune an open-weights model instead of using an API? Fine-tune an open-weights model only after prompting and retrieval are measured and still short, and when the gap is format consistency, a narrow high-volume task, latency, or data that must stay in-house. Fine-tuning appeared in 24 of 315 AI job ads (8%) in our 19 September 2026 sample, against evals in 23%. 6 min read AI engineer, ML engineer or data scientist: which should you train for in 2026? The AI engineer, unless you have a specific reason not to: 48 ads titled it and 62 more were software roles at AI companies, against 13 for ML engineer and 9 for data scientist. What each does all day, which suits you, what each takes to become hireable, and how people move between them. 5 min read Can a small business owner use AI without hiring anyone? Yes, for customer replies, admin, marketing and one automation, each pays back in its first month. What not to attempt alone, the low-cost monthly stack, the Privacy Act reality for businesses under A$3m, the order to do it in, and when it is time to bring in help. 5 min read Can you get an AI engineering job in Australia without a computer science degree? Yes, and it is the normal route: a degree is mentioned in 12% of AI job ads, a portfolio in 17%, 'shipped' in 39%. The three Australian doors that do need a degree, what the role is actually tested on, how long it takes, and what the portfolio must contain. 5 min read How do you build an AI portfolio that hiring managers actually read? Three deployed projects rather than thirty repositories, a live URL and a number on each, and a paragraph per project written for someone who will not run the code. What is wrong with most AI portfolios, which three projects, and the 150-word write-up that earns the next five minutes. 5 min read
24 of 127 shown
Research groupsComing next
Read it, then do it.Small groups, one open question.
You learn research by doing it. A research group is a small, time-boxed team working on one open question, with the same rule as everything on this page: show the sources and the limits.
Opening soon. Tell us the question you want to work on.
Propose a question- 1One open questionSomething the field has not settled yet, small enough to answer in a few weeks.
- 2A small groupA few people, time-boxed, with a clear end date.
- 3Sources and limitsYou collect the evidence, count it, and write down what it cannot tell you.
- 4Published hereWork that holds up can join this library.
Want to test what you read?
Competitions put it to work against a rubric. The community is where you argue about it.