[{"section":"Mapping the Landscape","title":"Mapping the Landscape of AI Agents","description":"The loop behind every agent Dominik Lukeš closes his Oxford Mapping the Landscape series with a definition: an agent is a model using tools in a loop to achieve complex tasks that require multiple steps. He shows the difference with an expenses claim that an agent completed in about thirty minutes, and traces how models took over more of the orchestration, from tool calling in 2023 to long loops that became reliable in late November 2025. He then covers what an agent needs, how to choose a model for it and what can go wrong. His advice for getting ready is to collect more context.\n\n You'll come away with\n\n- A working definition of an agent, with tools and loop as the keywords.\n- The loop in order: the harness sends instructions to the model, the model plans and sends commands, the harness passes them to the computer, and the results return to the model.\n- What an agent has to work with: a model writes from its training and its context, and nothing else, which is why desktop agents work on files in a folder, with AGENTS.md and skills.\n- The risks on your own files, unintended edits or deletions and changes made in bulk, and the advice to keep backups, ask for a changelog and read each permission request.","url":"/talks/2026-11-10/"},{"section":"Mapping the Landscape","title":"Mapping the Landscape of Large Language Models","description":"Choose your next AI model Dominik Lukeš's lecture maps the large language models of autumn 2026. It starts from one idea, that all AI is a model plus orchestration, and covers the models; agents follow the week after. He argues that what large language models add is meaning: they take instructions in plain language and handle messy input, and are less exact, so their output has to be checked. His short version is that bigger models are better, reasoning models are better than non-reasoning ones, and semantic models are better than non-semantic ones.\n\n You'll come away with\n\n- The three kinds of model, computational, predictive and semantic, and the step each does best when a customer's email is processed.\n- How two training phases make a model: pre-training gives general knowledge and skills such as language, and post-training gives the chat, the instructions and the reasoning.\n- How to read the model families and their generations, and how reasoning level affects score and cost, with medium to high as the sweet spot.\n- Why benchmarks saturate and why some labs train for them, and what else to compare: speed, latency and price.","url":"/talks/2026-11-03/"},{"section":"More talks","title":"AI as Tool User: Desktop agents for more productive study and research","description":"Let AI work with your files Dominik Lukeš explains how desktop AI agents can help with study and research. He starts from why people study at Oxford, to change their brain and to change the world. He thinks AI is fine to use for that, as a spotter: nobody builds muscle by asking someone else to go to the gym. He then shows what he does with Codex and Claude Code, and how files such as AGENTS.md and skills give an agent the context it needs.\n\n You'll come away with\n\n- A line on using AI in study: students still do the heavy lifting, exam and thesis rules differ so they should ask, and AI's work is never claimed as their own.\n- The move from AI as oracle, to tool maker, to tool user, and what a desktop agent is: a model using tools in a loop.\n- Research tasks an agent can do, among them replicating a paper in place of summarising it, running mini experiments and filing literature searches into Zotero.\n- How to set an agent up and keep it honest: an AGENTS.md file in every folder, skills for repeated tasks, a smarter model rather than a smaller one, and always checking what it does.","url":"/talks/2026-10-30/"},{"section":"Mapping the Landscape","title":"Mapping the Landscape of AI and Research","description":"Put agents to research work Dominik Lukeš maps how AI in research has changed. When ChatGPT arrived in 2023 the concern was that AI would write or edit papers; by 2026 AI makes independent discoveries and runs full research workflows, much of it true only in the last ten to twelve months. Autonomous discovery gets the headlines but is still the exception, so the lecture concentrates on assisted discovery. He shows what a desktop agent did for a literature review and for a weekend project cataloguing his old community radio recordings, and what this means for research projects, journals and supervision.\n\n You'll come away with\n\n- Three roles of AI, oracle, tool maker and tool user, and why people have moved from chatbots to desktop agents.\n- How the agent loop runs through a harness, a model and your computer, and that it can run for minutes, hours, even days.\n- Two worked cases: searching Elicit through its API and importing tagged papers into Zotero, and a published radio archive that the agent estimated at three to six hundred hours of work.\n- What changes for research: who is on the team, what aims are realistic and what it costs, plus journals overwhelmed with submissions and supervisors who do not know how to supervise new methods.","url":"/talks/2026-10-27/"},{"section":"More talks","title":"AI as Tool User: Coding agents for more productive study and research","description":"Make AI do the work In this lecture to MSc students, Dominik Lukeš describes his switch from ChatGPT to a desktop agent. On his computer he opens Codex or Claude Code, with the same models as before. He defines an agent as a model using tools in a loop to achieve complex tasks that require multiple steps, and shows what it did for him: a catalogue of 207 hours of radio recordings, a security audit of his machines and a completed expenses claim. The key to making it work over time is context, supplied through files, and his advice is to build in checking and double-checking, depending on how much the result matters.\n\n You'll come away with\n\n- How a chatbot differs from an agent, shown with a downloads-folder example in which the model asks for file sizes and the harness runs the command.\n- What agents can do for study and research, from a Zotero collection built from an Elicit search to a quick experiment that gave evidence instead of a guess.\n- How to give an agent context: an AGENTS.md file in the top folder and skills as folders with a SKILL.md file, and what data leaves your computer for each task.","url":"/talks/2026-10-26/"},{"section":"Mapping the Landscape","title":"Mapping the Landscape of AI Applications and Industry Developments","description":"Who really pays for AI Dominik Lukeš opens the lecture with a map of the AI industry in three layers: models, chatbots and apps, each of them a model plus orchestration. He places the labs, from the three established leaders to the returning ones and China's, and the companies that build on their models or add them as features. He then turns to what people use and pay for. His view is that agents are the new category, that full agent use now costs $100 to $500 a month, and that agents will become the normal way of using AI, whether in months or in a year.\n\n You'll come away with\n\n- A way to place any AI tool as a model, a chatbot or an app, and to say what a model plus orchestration means in each.\n- The package every lab now sells: a model family, a chatbot and a coding agent, free use with limits, and a subscription plan with credits, with virtually no unlimited use left.\n- What people use and pay for: by monthly visits ChatGPT is first, then Gemini, then Claude; in YipitData's US panel paid use rose from 2.1 per cent to 4.5 per cent of eligible consumers, and the top 1 per cent of payers account for more of the spending than the bottom half.\n- Why Dominik thinks apps in design, coding, research and productivity may matter less as chatbots absorb them, though he does not say they will go away.","url":"/talks/2026-10-20/"},{"section":"More talks","title":"AI agents for research workflows","description":"Put agents to work Dominik Lukeš shows researchers and professional services staff what AI agents can do in research workflows. A chatbot is no longer his main interface to language models: he uses Codex, and he says what changed in 2026 is better models and better harnesses. He works through an expenses claim, a paper replicated from its reported numbers and a corpus query in Sketch Engine, and explains the loop behind them and what the model sees. He ends on building in verification and on getting used to reaching for an agent automatically.\n\n You'll come away with\n\n- An agent as a model plus a harness plus a loop: the model plans and sends a command, the harness passes it to the computer, and the result returns to the model.\n- Two ordinary delegations: a Git problem solved in about fifteen minutes, and an expenses claim taken from email to money in the bank.\n- What the model sees and where data goes: a sandbox with folders that are never sent, and the point that a local folder does not mean a local model.\n- How to check an agent's work: a good-looking total of 97.5 per cent can hide one group at 35 per cent, so checks are needed, and a saved script that reproduces the same counts does not show that the method is right.","url":"/talks/2026-10-12/"},{"section":"More talks","title":"Developments in AI capabilities and industry trends","description":"See how fast AI changes Dominik Lukeš gave this talk on 30 September 2026 to a group concerned with admissions. He picked five trends from the many: capability, agents, voice, price and detection. Models keep improving rapidly, so everything about capability is a dated observation, and most policies still assume a chatbot. His answer was to raise the standards, and he had no answer yet on what the new baseline for competence should be.\n\n You'll come away with\n\n- A sense of the pace: a new model version every few weeks, and general benchmark scores up by 57 to 76 per cent.\n- The move from oracle to toolmaker to tool user, with the always-on assistants launched since February.\n- Two consequences he names: if you pay more, you can do more, and assumptions about equity and equality go back to the starting point.\n- A way to approach AI detectors: test them on existing data before using them, since false positives are worse than false negatives. His own small experiment gave no false positives, which he says will not hold for everything.","url":"/talks/2026-09-30/"},{"section":"More talks","title":"The State of AI in September 2026","description":"Let AI handle the details In September 2026, Dominik Lukeš says, the state of AI is defined by agents and by constant change. He shows what an agent does with his own trip to York: it put every journey on his calendar and completed the expenses claim, which cost an estimated $39.94 at API list prices. He covers what else agents are used for and how fast the products that offer them are changing. His main recommendation is the ChatGPT desktop app, which works on local files in a sandbox.\n\n You'll come away with\n\n- What an agent does differently from a chat: it makes a plan, uses tools, reads the results and revises, and not everything called an agent does this.\n- One worked example in full, from calendar to expense claim, with what it cost in tokens and in subscription use.\n- Where the products are heading: always-on assistants such as OpenClaw and Meta's Muse, custom GPTs retiring in favour of skills, and Chat and Work expected to merge.\n- What the free tier gives you, a taste of agents, and why the full value costs more.","url":"/talks/2026-09-24/"},{"section":"More talks","title":"Intern to Toolmaker: Five Roles AI can Play in Productive Study","description":"Turn AI into study tools Dominik Lukeš told Oxford students they are at Oxford to change their brain and to change the world, and that AI is a good use when it makes both better. He set out five roles for AI in study: intern or assistant, tutor, consultant, translator and toolmaker. Each has a use and a danger, and the answer to errors is built-in verification, since people make errors too. The toolmaker role leads to desktop agents that use tools on your own computer, which is new in 2026. He called the lecture a menu of options and advised against trying everything at once.\n\n You'll come away with\n\n- Three rules for using AI in study: do not say you did something when AI did it, do not say you worked unaided when you had help, and do not use AI when told not to.\n- What each of the five roles does, with his examples: a background search, a calendar file made from a description of his week, feedback on a tweet, a song and a mind map, and a Czech exercise website.\n- Why asking for a summary of a hard passage fails, and the steps he uses instead, from bullets and propositions to definitions, sense checks and your own writing with feedback.\n- A first task for a desktop agent, the downloads folder, and a way to use it on a paper: replicate one analysis rather than summarise.","url":"/talks/2026-09-11/"},{"section":"More talks","title":"Strategies & tools for productive reading and writing","description":"Read strategically, write with purpose Dominik Lukeš gives first-year students at Opportunity Oxford a morning session on productive reading and writing; the following session covers AI. Academic reading has a purpose, so it is strategic and cumulative: the reader's job is to identify, summarise and debate arguments. Writing is the core of academic practice, and he quotes Thomas Basbøll that scholars write to open their ideas to the criticism of their peers. Both improve with practice.\n\n You'll come away with\n\n- How to read strategically: read, skim and scan, how to skim, how to bootstrap from reference lists, and why not to go into a reading cold.\n- A consistent system for notes, with a warning against keeping all of them in Word, and what Zotero does for readings.\n- Habits for writing: write for your peers, Helen Sword's BASE model, deliberate practice, an eight-week writing boot-up and writing with others.\n- Two working rules: back up every statement, and always have an outline.","url":"/talks/2026-09-10/"},{"section":"About","title":"How these were made","description":"The process behind the videos and learning pages","url":"/how/"}]