The ten pathways at a glance
Index of all ten transitions with realistic timeline, difficulty, and entering salary band (US, illustrative 2026 figures). Difficulty is a rough composite of technical depth required, market saturation, and how much luck/timing weighs on the outcome.
| From → To | Timeline | Difficulty | US salary band (illustrative) |
|---|
| Software engineer → AI app builder | 3-6 months | Low | $110k - $180k |
| Writer → prompt engineer / AI content lead | 3-6 months | Low-medium | $70k - $140k |
| PM → AI PM | 3-6 months | Medium | $130k - $220k |
| Designer → AI UX designer | 6-12 months | Medium | $110k - $180k |
| Consultant → AI strategy consultant | 6-12 months | Medium | $120k - $250k |
| Educator → AI educator + tools dev | 6-9 months | Medium | $80k - $150k |
| Independent operator → AI solo founder | 6-12 months | High (capital-bound) | $0 - $1M+ (binary) |
| Software engineer → ML engineer | 12-18 months | Medium-high | $160k - $300k |
| Data analyst → ML engineer | 12-18 months | High | $140k - $260k |
| Researcher → AI researcher (lab) | 12-24 months | Very high | $250k - $900k+ |
From → ToSoftware engineer → AI app builder
Timeline3-6 months
DifficultyLow
US salary band (illustrative)$110k - $180k
From → ToWriter → prompt engineer / AI content lead
Timeline3-6 months
DifficultyLow-medium
US salary band (illustrative)$70k - $140k
From → ToPM → AI PM
Timeline3-6 months
DifficultyMedium
US salary band (illustrative)$130k - $220k
From → ToDesigner → AI UX designer
Timeline6-12 months
DifficultyMedium
US salary band (illustrative)$110k - $180k
From → ToConsultant → AI strategy consultant
Timeline6-12 months
DifficultyMedium
US salary band (illustrative)$120k - $250k
From → ToEducator → AI educator + tools dev
Timeline6-9 months
DifficultyMedium
US salary band (illustrative)$80k - $150k
From → ToIndependent operator → AI solo founder
Timeline6-12 months
DifficultyHigh (capital-bound)
US salary band (illustrative)$0 - $1M+ (binary)
From → ToSoftware engineer → ML engineer
Timeline12-18 months
DifficultyMedium-high
US salary band (illustrative)$160k - $300k
From → ToData analyst → ML engineer
Timeline12-18 months
DifficultyHigh
US salary band (illustrative)$140k - $260k
From → ToResearcher → AI researcher (lab)
Timeline12-24 months
DifficultyVery high
US salary band (illustrative)$250k - $900k+
Pathway 1 — Software engineer → AI app builder (3-6 months)
The fastest real transition. If you already ship production code in Python or TypeScript, you can be useful on an AI product team within a quarter. The job is wiring foundation models into existing products: retrieval-augmented generation, agent loops, evaluation harnesses, prompt management, cost and latency tuning. You do not need to train models. You need to be excellent at integration, observability, and API economics.
Starting skills required: comfortable with a modern web stack, REST/streaming APIs, async patterns, basic SQL, and at least one vector database concept. Target skills to add: the Anthropic Messages API or OpenAI Responses API end-to-end (streaming, tool use, structured outputs), one retrieval framework (LlamaIndex or LangChain — both have steep complexity costs, pick deliberately), one eval framework (Inspect or Promptfoo), one observability tool (Langfuse, Helicone, or built-in provider dashboards), and prompt-caching economics. Read the Anthropic prompt engineering guide and the OpenAI cookbook end-to-end; they are short and load-bearing.
Study path (specific): DeepLearning.AI's short courses on LangChain and on Building Systems with the OpenAI API (free on their platform); Hamel Husain and Shreya Shankar's evaluation course materials (publicly posted on Hamel's blog at hamel.dev); the Anthropic prompt engineering interactive tutorial on GitHub at anthropics/prompt-eng-interactive-tutorial.
Portfolio milestones: one RAG system over a non-trivial corpus (10k+ documents, real chunking decisions, an eval harness with at least 50 graded examples); one agentic workflow with tool use and a recovery loop; one cost/latency optimization writeup with before/after numbers. Open-source one of them with honest receipts including failure cases.
Entering salary band (US, illustrative 2026): $110k-$180k for a mid-level role at a non-FAANG company. FAANG and frontier labs pay substantially more but hire at higher bars.
Common traps: (1) Building demo apps that never see eval data; the field has matured past demo screenshots. (2) Chasing the newest framework every six weeks instead of mastering API fundamentals. (3) Underestimating how much production AI engineering is operations and evals, not prompts. (4) Believing your portfolio matters more than your shipping evidence — interviewers want to see deployed systems with users.
Pathway 2 — Writer → prompt engineer / AI content lead (3-6 months)
Real, but smaller than the 2023 hype suggested. The role of dedicated "prompt engineer" with no other skills has largely consolidated into broader content-and-AI hybrid roles: documentation engineers, AI-assisted content leads, technical writers embedded in AI product teams, and red-teamers on safety/policy work at labs. The career is real; the job title is in flux.
Starting skills required: strong writing, editorial judgment, comfort with structured documents and version control (Git basics), willingness to read papers and provider documentation as primary sources. Target skills to add: deep familiarity with one major model family's strengths and failure modes (Anthropic's Claude, OpenAI's GPT, or Google's Gemini); structured prompt design (XML tags, role separation, few-shot patterns); evaluation methodology for subjective outputs; basic Python or TypeScript to run prompts at scale (Jupyter is enough); content ops tooling.
Study path: Anthropic's prompt engineering documentation (claude.com/docs); the Anthropic prompt engineering interactive tutorial; Ethan Mollick's One Useful Thing newsletter and his book Co-Intelligence (Penguin, 2024) for working intuition; Lilian Weng's blog posts on prompt engineering (lilianweng.github.io) for the technical layer. For red-team and safety angles, read Anthropic's published Acceptable Use Policy and any current model cards on the official site.
Portfolio milestones: one published prompt library with version history and eval results; one writeup comparing model behavior on a meaningful task across at least two providers, with actual numbers; one piece of public technical writing on AI methodology that gets cited or linked by working practitioners.
Entering salary band (US, illustrative 2026): $70k-$140k depending on whether the role is editorial-leaning (lower) or engineering-leaning (higher). Pure "prompt engineer" titles are increasingly rare; the role inside applied AI teams pays better than the standalone job title suggests.
Common traps: (1) Believing prompt engineering alone is a durable career; it is a skill on a path, not a destination. (2) Building Twitter-style "prompt pack" portfolios with no evals; serious teams ignore these. (3) Missing the model-specific behavioral knowledge that separates working prompt engineers from generalists.
Pathway 3 — PM → AI PM (3-6 months)
If you have shipped real products, this is a fast and well-paid transition. AI PM work is product management with three added domains: model evaluation literacy, AI economics (token costs, latency budgets, throughput), and the safety/policy surface that ships with any LLM-touching product. You do not need to write training code. You need to be able to read an eval report and tell a senior engineer whether the launch is acceptable.
Starting skills required: senior PM track record at a software company, comfort with technical specifications, working relationship with engineering. Target skills to add: read the major model cards and capability evaluations critically (MMLU, GPQA, SWE-bench, HumanEval, MATH); design custom evals for your product surface; understand the difference between zero-shot, few-shot, fine-tuned, and RAG approaches at the architecture level; know the major providers' pricing and rate-limit structures; understand RLHF, constitutional AI, and the basics of alignment as it affects shipping decisions.
Study path: read the GPT-4 technical report (arxiv 2303.08774) and the Anthropic Claude technical reports as published; Chip Huyen's Designing Machine Learning Systems (O'Reilly, 2022) for production ML literacy; Andrej Karpathy's Intro to LLMs YouTube talk (1 hour, free) for clean fundamentals; Hamel Husain's writing on AI evals at hamel.dev; OpenAI and Anthropic's public docs sections on safety and use policies. For the policy-and-governance layer, read NIST's AI Risk Management Framework (AI RMF 1.0, January 2023, NIST.AI.100-1).
Portfolio milestones: ship one AI feature in your current role with documented evals and a launch retrospective; write one internal or public spec for an AI feature that includes an explicit evaluation plan; lead one go/no-go decision on a model-touching feature based on real test data.
Entering salary band (US, illustrative 2026): $130k-$220k base for mid-to-senior PM roles at non-FAANG; substantially higher at frontier labs and FAANG. Equity component can dominate cash at well-funded AI companies.
Common traps: (1) Talking about "AI features" without understanding eval methodology; senior engineers will dismiss you. (2) Treating model selection as a vibes decision instead of an eval-driven one. (3) Underestimating how much AI PM work is governance and policy, not just feature design.
Pathway 4 — Designer → AI UX designer (6-12 months)
Genuinely new design territory. AI products break most of the patterns design schools teach: outputs are non-deterministic, latency is variable, errors are linguistic instead of system-level, and the user's mental model of the agent matters more than the visual hierarchy. The roles are real and growing; the cost is that very few schools or courses are teaching this well yet.
Starting + target skills
Starting: strong product design portfolio, comfort with design systems and Figma, working sense of interaction patterns. Target: streaming output design, agentic workflow UX, citation and provenance patterns, refusal and uncertainty UX, eval-driven design (UX research with model outputs as variable), latency-tolerant interaction design.
Study path
No canonical curriculum yet — read working designers
Maggie Appleton's essays on AI design (maggieappleton.com); Linus Lee's writing at thesephist.com; Ben Hylak's design writeups on Raindrop and similar tools; Nielsen Norman Group's growing library on AI/LLM UX patterns at nngroup.com; case studies from Linear, Notion, and Cursor (all post public design retrospectives). No single canonical course exists yet — read primary practitioners.
Portfolio milestones
One shipped or convincingly prototyped AI surface with streaming, error states, and uncertainty handling. One case study showing iteration based on real model behavior (not mocked). One critique or redesign of a current AI product with a clear design thesis.
Salary + traps
US illustrative 2026: $110k-$180k mid-senior; higher at frontier labs and well-funded startups. Traps: (1) Designing for the demo, not the failure mode. (2) Ignoring streaming and latency as primary design variables. (3) Building Figma mocks that hide the non-deterministic reality of the output.
Pathway 5 — Researcher → AI researcher at a lab (12-24 months, hard)
The hardest pathway on this page, and the most credential-bound. Frontier labs (Anthropic, OpenAI, Google DeepMind, Meta FAIR, xAI, and a small set of well-funded startups) hire research engineers and research scientists from a narrow funnel: top-tier PhDs, published authors on highly-cited papers, and a small number of exceptional engineers with research output. If you do not already have a PhD or a serious publication track, this transition realistically takes 2-4 years, not 12-24 months — the timeline in our table is for people already adjacent (postdocs, strong applied ML engineers with publications, etc.).
Starting skills required: graduate-level math (linear algebra, probability, optimization), PyTorch or JAX fluency at the training-loop level, at least one peer-reviewed publication or strong open-source equivalent (the Hugging Face leaderboard isn't a substitute), demonstrated ability to read and reproduce papers. Target skills to add: deep familiarity with one research area (post-training, alignment, interpretability, scalable oversight, evaluation methodology, mechanistic interpretability, etc.); ability to design and run experiments at scale; ability to write a paper that meets ICLR/NeurIPS/ICML bar.
Study path: Karpathy's Neural Networks: Zero to Hero YouTube series (free, exceptional); Stanford CS 224N (NLP with deep learning, public lectures at web.stanford.edu/class/cs224n); the Goodfellow/Bengio/Courville Deep Learning textbook (free at deeplearningbook.org); Sebastian Raschka's LLMs from Scratch (Manning, 2024) for hands-on transformer building; primary papers — read Attention Is All You Need (arxiv 1706.03762), GPT-3 (arxiv 2005.14165), the InstructGPT/RLHF paper (arxiv 2203.02155), and the Constitutional AI paper from Anthropic (arxiv 2212.08073). Subscribe to Distill.pub archives and Anthropic's interpretability writeups at transformer-circuits.pub.
Portfolio milestones: one reproduction of a non-trivial paper with public code and a writeup of where you diverged; one original contribution (a new eval, a finding on an open question, a clean implementation of a recent technique with measurable improvements); ideally one accepted workshop paper at a major venue or a widely-cited blog post in the research community.
Entering salary band (US, illustrative 2026): $250k-$900k+ total compensation at frontier labs. Equity at private labs can dominate cash. Salaries here have been distorted by the funding environment and may compress over the next 24-36 months.
Common traps: (1) Believing online courses replace publication-quality work; they do not. (2) Spending years on "prep" without producing public output researchers can read. (3) Underestimating the role of advisor lineage and network in hiring at the very top labs.
Pathway 6 — Software engineer → ML engineer (12-18 months)
The classic transition, and still real. ML engineering is the production discipline of training, fine-tuning, deploying, and monitoring machine learning systems at scale. Distinct from research (you implement and operationalize, you don't propose new methods) and distinct from AI app building (you work at the model and infrastructure layer, not the API integration layer).
Starting skills required: solid software engineering background (3+ years production), strong Python, comfort with distributed systems and cloud infrastructure, basic statistics and linear algebra. Target skills to add: PyTorch fluency including training loops, optimizers, and mixed precision; experiment tracking (Weights & Biases or MLflow); data engineering for ML (feature stores, data versioning with DVC); model serving (Triton, vLLM, TGI, or one of the managed alternatives); evaluation methodology; one specialty (computer vision, NLP, recommender systems, etc.).
Study path: fast.ai's Practical Deep Learning for Coders (free, course.fast.ai) — start here, it is the highest-leverage entry point; Andrew Ng's Deep Learning Specialization on Coursera (paid, but the financial aid is real) for the math layer; Chip Huyen's Designing Machine Learning Systems (O'Reilly, 2022) and her follow-up writing on chiphuyen.com; the Made With ML production course materials (madewithml.com); for the deep transformer layer, Karpathy's Zero to Hero series; for serving, read the vLLM paper (arxiv 2309.06180) and the TGI documentation.
Portfolio milestones: one end-to-end project taking a model from training through production serving with evaluation, monitoring, and a writeup of cost and latency tradeoffs; one fine-tuning project with honest comparison to the base model on a held-out test set; ideally one contribution to a major open-source ML project (transformers, PyTorch, vLLM, etc.).
Entering salary band (US, illustrative 2026): $160k-$300k for mid-to-senior ML engineers at non-FAANG. FAANG and frontier labs significantly higher.
Common traps: (1) Confusing notebook proficiency with production ML engineering — the gap is enormous. (2) Skipping the math and getting stuck at the "can run a script, can't debug a divergent training run" plateau. (3) Building Kaggle-style portfolio projects that don't show production discipline (data drift, monitoring, rollback, A/B testing).
Pathway 7 — Data analyst → ML engineer (12-18 months)
Harder than the software-engineer transition because it requires building production software engineering muscle in parallel with ML skills. Many data analysts try this transition and stall at the "can do ML in a notebook, can't ship it" stage. The successful path is to treat the software engineering gap as the primary risk and close it deliberately.
Starting skills required: SQL fluency, Python including pandas and basic scikit-learn, business analytics experience, comfort communicating results to non-technical stakeholders. Target skills to add (in this order): production Python (typing, testing, packaging, async, real refactoring discipline); software engineering practice (Git workflow, code review, CI/CD); PyTorch and the rest of the ML engineering stack from Pathway 6; cloud infrastructure (one of AWS/GCP/Azure to working depth).
Study path: close the SWE gap first — work through Software Engineering at Google (free at abseil.io/resources/swe-book) and Real Python's intermediate tracks; then run the same curriculum as Pathway 6 (fast.ai, Huyen, Karpathy). Consider taking on a backend engineering role for 6-12 months as a stepping stone — many analysts who jumped to ML engineering directly stalled, while those who did data engineering or backend first succeeded.
Portfolio milestones: ship one production data pipeline with real engineering rigor (tests, monitoring, deploys); ship one ML system end-to-end as in Pathway 6; ideally land a data engineering or platform role first if direct ML engineering offers are not coming.
Entering salary band (US, illustrative 2026): $140k-$260k for mid-level ML engineering after successful transition; the analyst-to-engineer jump itself often involves a temporary compensation plateau as you rebuild.
Common traps: (1) Underestimating the software engineering gap; this is the killer. (2) Doing Coursera courses without writing production code. (3) Skipping the data engineering middle step when direct ML engineering isn't landing.
Pathway 8 — Educator → AI educator + tools developer (6-9 months)
A growing real category. Schools, universities, edtech companies, and corporate L&D departments need people who understand AI well enough to teach it responsibly and to build the tools that integrate it into curricula. The role spans curriculum designer, instructional technologist, applied researcher in learning science, and increasingly a tools developer position at edtech startups.
Starting skills required: teaching experience, curriculum design, comfort explaining technical concepts to non-experts, ideally some technical background (you do not need to code, but it accelerates this transition significantly). Target skills to add: working AI literacy from Pathway 2's reading list; one technical skill (Python or no-code AI tooling like Replit, Cursor, or Claude Code as a teaching surface); learning science fundamentals (spaced repetition, retrieval practice, deliberate practice); evaluation methodology for AI tutoring systems.
Study path: Ethan Mollick and Lilach Mollick's work on AI in education (Wharton, publicly published essays and papers — search Google Scholar for their AI tutoring work); the Khan Academy / Khanmigo public writeups; the AI Education community at AIEDU.org; Make It Stick by Brown, Roediger, McDaniel (Harvard, 2014) for learning science; ISTE's Standards for Educators on AI literacy at iste.org.
Portfolio milestones: one AI-augmented curriculum or course you have taught with documented student outcomes; one public writeup of a methodology or tool you built; ideally one piece of software (a custom GPT, a Claude project, a small web app) that other educators use.
Entering salary band (US, illustrative 2026): $80k-$150k depending on whether the role is in-school (lower), at an edtech company (mid), or as a consultant (variable). University-level instructional design with AI specialization pays better than K-12 generally.
Common traps: (1) Building hype-driven AI-in-education content; districts and serious edtech teams have become skeptical. (2) Ignoring the equity, privacy, and assessment-integrity issues that dominate real AI-in-ed deployments. (3) Underestimating how much this role requires you to push back on bad AI products, not adopt every new tool.
Pathway 9 — Consultant → AI strategy consultant (6-12 months)
Real and well-paid, but increasingly crowded. The market for "AI strategy" advisory work expanded fast in 2023-2025 and the bar to be taken seriously has risen with it. Successful consultants in this space combine prior domain expertise (finance, healthcare, legal, manufacturing, etc.) with genuine technical literacy and a track record of shipped engagements.
Starting skills required: existing consulting practice or in-house strategy role with a track record, deep domain expertise in at least one industry, executive communication. Target skills to add: AI literacy from Pathway 3's reading list; the economics of foundation models (training costs, inference costs, latency budgets, vendor lock-in considerations); the regulatory surface for AI in your domain (HIPAA, FERPA, GDPR, the EU AI Act); the difference between feasible and aspirational AI deployments at enterprise scale.
Study path: read McKinsey, BCG, and Bain's public AI reports critically — useful for the language, dangerous if taken at face value; Karpathy's Intro to LLMs talk; the EU AI Act primary text (eur-lex.europa.eu, search for Regulation 2024/1689); NIST AI RMF 1.0; the AI Index Report from Stanford HAI (annual, free, aiindex.stanford.edu) for cross-industry trends; provider economics docs (Anthropic, OpenAI, Google pricing and rate-limit pages).
Portfolio milestones: one shipped AI strategy engagement with measurable business outcome; one industry-specific writeup or whitepaper that establishes credibility; ideally one speaking engagement or conference talk in your industry where you say honest things about AI.
Entering salary band (US, illustrative 2026): $120k-$250k base for in-house strategy roles; consulting day rates highly variable, typically $2k-$15k/day for established consultants with credible track records. Big-three consulting AI practices pay senior rates with substantial bonus components.
Common traps: (1) Selling AI strategy without a track record of shipped AI work — clients have gotten much better at sniffing this out. (2) Recommending generative AI deployments where classical ML or process improvement is the right answer. (3) Underestimating the regulatory and procurement complexity at enterprise scale.
Pathway 10 — Independent operator → AI solo founder (6-12 months, capital-dependent)
Binary outcomes — read this section carefully. Most AI solo-founder attempts return zero. A small percentage return large multiples. The path is not a job transition; it is a deliberate bet with your time and capital that you should make with full information.
Starting position required: enough runway to survive 12-24 months without revenue (6 months minimum, 18 months ideal), prior product or technical experience, comfort with public failure, and ideally a specific problem you understand from the inside. The single highest predictor of success is domain-specific founder-market fit, not technical ability.
Target skills to add: enough technical depth from Pathway 1 to build credible product without a co-founder (or a co-founder); the economics of AI products (gross margin, COGS, churn dynamics specific to AI products where token costs scale with usage); distribution (most solo AI founder failures are distribution failures, not product failures); honest financial modeling.
Study path: Paul Graham's essays on startups (paulgraham.com), especially Do Things That Don't Scale and Default Alive or Default Dead; Patrick McKenzie (patio11) on independent software businesses at kalzumeus.com; The Mom Test by Rob Fitzpatrick for customer discovery; Stripe Atlas's guides for solo founders; the Y Combinator Startup School free curriculum at startupschool.org; for AI-specific economics, study public writeups from companies like Cursor, Linear, and Notion on AI feature COGS.
Portfolio / milestones: one product with paying customers (anything above $10 MRR proves the channel works); one honest financial model with realistic AI cost projections; one quarterly review of what you'd do differently with current information.
Outcome range (US, illustrative 2026): $0 (most common) to $1M+ revenue annual run rate (rare). Median solo AI founder outcome is closer to zero than to a successful exit. Compensation during the build phase is whatever your savings allow.
Common traps: (1) Believing the AI gold rush means any AI product will find a market — most won't. (2) Underestimating distribution; building is the easy part. (3) Running out of money before product-market fit; most failures here are timing and capital, not product. (4) Becoming a wrapper on a single foundation model with no defensible advantage.
Cross-cutting truths and a 12-month sample timeline
Things that apply to every pathway above. Print these and pin them somewhere visible. Below the list, a concrete 12-month sample timeline for the most common pathway (Software engineer → AI app builder), calibrated to a working engineer with ~10 hours per week of study and build time — adjust for your actual constraints. Months 1: API fundamentals (work through Anthropic and OpenAI quickstarts end-to-end; ship one small prototype to Vercel or Cloudflare; read the Anthropic prompt engineering tutorial cover to cover; start a public log). Months 2-3: RAG and evals layer (build one non-trivial RAG system over 10k+ documents; stand up an eval harness with at least 50 graded examples; read Hamel Husain's writing on evals end-to-end; publish one writeup with real numbers). Months 4-5: Tools, agents, observability (build one agentic workflow with tool use, retry logic, and observability; compare at least two provider implementations of structured outputs; start writing publicly about cost/latency tradeoffs). Month 6: Job search and open source (polish two portfolio projects with honest READMEs including failure cases; apply to AI-adjacent roles at companies whose products you actually use; start contributing to one open-source AI project to widen your network). Three questions to ask yourself before committing a year to any pathway: (1) What does the median outcome look like, not the survivorship-bias success story? (2) Does your runway match the realistic timeline rather than the marketing timeline? (3) What is your fallback if this transition takes 1.5x or 2x the expected time — the normal case?
- Luck and timing dominate single-instance outcomes. The same skills that landed someone a frontier-lab job in 2023 may not in 2026. Plan for the median, not the survivor.
- Public output beats private credentials almost everywhere except frontier research. A working portfolio with shipping evidence usually outperforms a fresh certificate.
- Evaluation literacy is the meta-skill. Across every pathway above, the practitioners who actually ship are the ones who can design and read evals.
- Foundation model providers move pricing and capabilities quarterly. Anything you read about "the best model for X" is dated within months. Re-verify before you commit architecture decisions.
- Most AI startup failures are distribution failures, not technical failures. Most AI career failures are not-shipping-publicly failures, not skill failures.
- The salary numbers in this document are 2026 best-effort estimates from public sources. Real offers have high variance. A single FAANG or frontier-lab offer can be 2-4x the median for the same title.
- This field selects against people who can't tolerate ambiguity, rapid change, and being wrong in public. If that's not you, look at adjacent careers where AI is a tool, not the core.
- If a pathway above takes longer than expected, you can usually compound — finish Pathway 1, then move toward Pathway 6 from a stronger position. The reverse (starting hard, bailing partway) leaves you with neither outcome.
