The Expert’s Road
Where the field is being made: how research reaches the world, how to keep up without drowning, what it means to reproduce a paper, the landmark papers worth your weekends, and how much nobody yet knows.
The paper that introduced the Transformer was posted to a free website in June 2017, months before it was presented at a conference. Within a few years its architecture ran almost every major language model on Earth. That is how the frontier of AI moves: in public, at speed, through documents anyone can download. The barrier to entry is not access. It is the ability to read those documents critically, tell signal from noise, and check claims for yourself. That ability is what this track builds.
If the internals are new to you, start with The Advanced Road.
How the frontier works
AI research reaches the world through four channels, and an expert watches all of them.
Two cautions. First, an open-weight model is not the same as open source: you may get the parameters but not the data or training code. Second, a growing share of frontier results arrives as technical reports and blog posts from companies, with limited detail As of 2026. Treat claims you cannot check as claims, not facts.
Keeping up without drowning
Nobody reads everything. Experts read narrowly and deeply in their own area and skim widely outside it. A sustainable routine looks something like this:
- Pick one or two areasInterpretability, efficiency, evaluation, reasoning — choose where you want depth. Everything else is skimming.
- Skim abstracts weeklyFollow the arXiv listing for your subfield (for example cs.CL for language, cs.LG for machine learning) or a curated digest. Read titles, then abstracts, then figures. Most papers stop there.
- Read one paper properly each weekSeveral passes: abstract and figures; introduction and conclusion; then the method in full with a notebook beside you. Write a five-line summary in your own words.
- Follow people, not hypeResearchers whose work you have read and respected are a better filter than trending posts. Be suspicious of any result that only appears in a thread.
- Return to the conferencesAccepted papers, oral presentations and reviewer discussions tell you what the community thought held up.
This guide’s framing Missing a paper costs you little; reading ten shallowly costs you a week. Depth compounds.
What “reproducing a paper” means
To reproduce a paper is to take its description, rebuild the method, and check whether you get the claimed results. It is the most effective way to learn a technique and the best defence against over-claiming. In practice there are levels:
- Re-run — use the authors’ released code and checkpoints, and confirm the reported numbers.
- Re-implement — write the method yourself from the paper alone. This is where you discover the details the paper left out: learning-rate schedules, data filtering, an initialisation trick in an appendix.
- Replicate at small scale — test whether the core finding holds on a smaller model or dataset you can afford. Most independent researchers work here.
- Stress-test — change the data, seeds or benchmark and see whether the effect survives. Many do not.
Consensus Machine learning has a well-documented reproducibility problem: unreported tuning, single random seeds, test-set leakage and compute budgets outsiders cannot match. The community responded with code-release norms, reproducibility checklists at major conferences and an annual ML Reproducibility Challenge. Your honest small replication is a genuine contribution.
The research areas of this track
The pages after this one are still being written and will go live in turn. They cover:
READReading research papersA method for papers you cannot yet understand
Multi-pass reading, decoding notation, checking experiments, and keeping a reading log.
CoTReasoning modelsChain of thought and test-time compute
How chain-of-thought prompting became trained behaviour, and why spending more computation at answer time can improve results.
MIMechanistic interpretabilityFeatures, circuits and looking inside
Attempts to reverse-engineer what individual parts of a network compute.
ALAlignment and safety researchMaking capable systems do what we intend
Preference training, oversight, evaluations for dangerous capabilities, and the open problems of alignment.
RLReinforcement learningFrom bandits to RL on language models
The theory of reinforcement learning and its new role in training reasoning models.
MMMultimodal and world modelsVision, audio, video and action
Models that see, hear and act, and the debate over whether they build internal models of the world.
EFFEfficient modelsQuantisation, distillation, mixture-of-experts
Doing more with fewer bits, smaller students and sparse activation.
SCLTraining at scaleParallelism, hardware and the compute stack
How training is split across thousands of chips, and what limits it.
AGTAgentic systemsPlanning, memory, tools and evaluation
Research on agents that act over many steps, and how to measure them honestly.
?Open problems; contributingWhat nobody knows yet, and how to help
Two closing pages: a map of unsolved questions, and practical routes into research, open source and careers.
Landmark papers to read
No list is complete, but these ten are cited constantly and repay a careful read. Tap a spine for the reference and why it matters.
For the current frontier, add Wei et al., “Chain-of-thought prompting elicits reasoning in large language models” (2022) and DeepSeek-AI’s “DeepSeek-R1: Incentivizing reasoning capability in LLMs via reinforcement learning” (2025), which described in public a recipe for training reasoning behaviour with reinforcement learning. For the story behind several of these, see The Deep Learning Revolution and Attention Is All You Need.
Honest uncertainty at the frontier
At this level you will meet confident claims in every direction. Hold them lightly. Some of the biggest questions have no settled answer:
- Debated Does scaling keep paying? Scaling laws held remarkably well for pretraining loss, but whether more compute, data and test-time reasoning continue to buy proportional real-world capability is argued over, as is the supply of high-quality data.
- Debated Do benchmarks measure what we think? Scores saturate quickly and test items leak into training data, so headline numbers can overstate progress.
- Speculative When, if ever, will systems reach AGI? Forecasts by serious researchers range from a few years to many decades, to “the question is ill-posed”.
- Speculative Can we understand and control far more capable systems? Interpretability and alignment research are young; whether they will keep pace with capability is unknown.
- Scaling optimists: argue that larger models, more compute and reinforcement learning on reasoning will keep producing large capability gains, as they have since 2012.
- Architecture sceptics: researchers such as Yann LeCun argue that next-token prediction alone will not yield human-level understanding, and that new architectures — for example world models that learn from video and action — are needed.
- Measurement critics: argue that much reported progress reflects benchmark overfitting and contamination, and call for harder, held-out and real-world evaluations.
- Risk-focused researchers: argue that whatever the timeline, capabilities are outpacing our ability to verify safety, and that alignment and governance deserve far more effort now.
Your job as an expert is not to pick a team. It is to know which claims rest on evidence you have checked, which rest on evidence you have not, and which rest on none.
Choose one paper from the shelf. Read it in three passes, write a five-line summary, then find one claim you could test yourself on a laptop in a weekend — and test it.
// Pause and reflect
- Which one or two research areas will you go deep on, and why those?
- Think of a recent AI headline you believed. What evidence would you need to check it properly?
- Of the open questions above, which do you think is most likely to be resolved in the next five years — and what would change your mind?
Your notes are saved privately in this browser.
Key terms on this page
Sources & further reading
- Ashish Vaswani et al., “Attention Is All You Need” (2017).
- Jared Kaplan et al., “Scaling Laws for Neural Language Models” (2020); Jordan Hoffmann et al., “Training Compute-Optimal Large Language Models” (2022).
- Long Ouyang et al., “Training language models to follow instructions with human feedback” (2022).
- DeepSeek-AI, “DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning” (2025).
- Joelle Pineau et al., “Improving Reproducibility in Machine Learning Research (A Report from the NeurIPS 2019 Reproducibility Program)” (2020).
- S. Keshav, “How to Read a Paper”, ACM SIGCOMM Computer Communication Review (2007) — the classic three-pass method.
- arXiv and OpenReview — where to find preprints and public reviews.