Talks

Agentic risk does not come from machine consciousness or will. It comes from symbols, optimization, tools, permissions and feedback loops. This talk gives managers a sober, non-catastrophist framework:

• the risk product (capability x loop autonomy x how critical the environment is x how broad the permissions are x how weak the supervision is)
• how a system quietly games its own targets (Goodhart's law and reward hacking, with no bad intent required)
• the security boundary that becomes linguistic (a prompt injection turns a wrong answer into a wrong action)
• the shift from policing what a machine supposedly thinks to governing the actions it is actually allowed to take

Governance is not the brake that slows enterprise innovation down; it is the set of brakes that let you drive the most advanced AI fast and still keep it acceptable. It also situates why this is a now problem and where agents already act, placing them on the five degrees of autonomy, set by the MAYA cursor (Most Advanced Yet Acceptable, as advanced as possible without passing the point where the human disengages), before deciding what each degree may be permitted to do. The golden rule is one line: the costlier the error, the less autonomy the AI keeps.

You ask a language model a question about your data. It answers with a number. The number is plausible. It is wrong. Hardly anyone signs off on a board decision on the strength of a plausible number. That single fact blocks more deployments than budget and regulation combined. The answer is not a better prompt: it is to take away the model's chance of being wrong about the number.

• The model writes the query, the database answers. A wrong query can be seen, kept, replayed and fixed; a made-up number cannot be seen, because it looks exactly like a correct one.
• Two locks that do not depend on good will: a read-only account and a parser that inspects the structure of the query before anything runs.
• Four questions, in the order a meeting asks them: a number, a row-by-row prediction, a three-month forecast, a chart chosen live.
• An exam you are not allowed to have seen: held-out question and query pairs, scored by running both and comparing results, then broken down by family of question.
• A figure without its uncertainty is an opinion dressed as a fact, so every prediction arrives with its range.

Built on a real deployment, with the failure modes that only production teaches.

Ask a team what one unit of their work costs to run, per request, per document, per minute of audio: you will usually get a cloud bill divided by a guess. This talk builds the dial that meter never had and says which numbers it is allowed to show.

• The boundary makes the figure. Two published figures for one inference sit a factor of thirty-eight apart without either party being evasive: they answer two different questions that share a sentence.
• Every number carries its status: measured, estimated, held open, or still owed by a human. A derived number may never claim to be better founded than the weakest of its inputs.
• Manufacturing the hardware usually dominates the electricity on a short request, by three to six times and the lifetime that sets it is yours to state.
• One run bounds nothing: over five identical runs, runtime moved by 4.5 percent and measured power by 49 percent.
• Annex XI of the European AI Act asks for known or estimated energy consumption and permits an estimate from the compute used, which is this method written into law.

The deliverable is a file committed next to the code, reviewed in a pull request and a gate that fails the build when a cost or a status gets worse.

Governing artificial intelligence by permission assumes there is somebody to ask. That assumption rests on a floor nobody votes for: compute, energy, networks, exclusive data. When cognition becomes abundant, scarcity relocates. Governing by permission quietly turns into governing by ownership.

• An operating system rather than electricity: the metaphor you pick decides which regulator you think you need.
• The energy wall is local before it is global: a national average says nothing about the town whose water or grid capacity a single site absorbs.
• Green information technology is the governance of a unit cost, not of an aggregate, which is also why Jevons' paradox is not an objection to efficiency.
• Compute as a political lever: detectable, excludable, countable, therefore controllable, as export controls have already demonstrated.
• The criticality triplet rather than the sector: capacity to act, critical infrastructure, systemic effects.

The talk ends on the one condition under which the whole permission approach stops meaning anything: a single actor owning everything.

"While the counter is below ten, go round again" is one of the oldest sentences in computing. It works because the machine can answer the question it is asked, exactly, in one clock cycle. The questions that decide the quality of a piece of work are not like that: is this chart readable, does this paragraph follow from the one before, does this forecast deserve production. So we pulled them out of our loops. This talk puts them back.

• Four steps: act, render, judge, decide. The step teams skip is the render. Without it you are judging an intention rather than a result.
• The judge may be a human eye, a vision model or a held-out dataset. It must not be the actor: when the same model writes and grades, the grade climbs while real quality stalls.
• The measured trap: consult the same judge often enough and the loop learns the judge's quirks instead of the subject.
• An instrument for prose: the round trip through another language, where what comes back deformed marks the place the text does not constrain its own reading.
• Two rules of conduct: an autonomous loop is worth the quality of the judge it is given, plus the ban on corrupting that judge.

Everyone keeps the drawing of the brain; almost nobody keeps the drawing of the plumbing. The model is one cheap organ. The 90% is:

• ingestion
• annotation
• deployment
• monitoring
• the wires that carry the signal end to end

And it is where AI projects actually live or die. MLOps is DevOps for models with two extra headaches: the data drifts and there is no source code to read, only weights you must measure. A talk for engineers and technical managers in the same room, enough code to be real, enough diagram to be led.

Every trade now asks the same question: what happens when a machine takes over the central function of a craft. Art has already answered it twice, in 1839 when photography took realistic imitation away from painting and again now that synthesis takes image-making away from photography. The second answer is more useful than any forecast.

• The verdict of 1859, servant and not master and what the craft did next: it changed its question rather than its tools.
• Measured creativity: machines win the parlour test of unusual uses for an object and lose on whole works judged blind by juries who did not know.
• A working practice, in numbers: seventy percent of a film generated, eighty percent of the time spent animating and reworking it, which refutes the myth of the magic prompt.
• Permissions won rather than principles argued: the 2023 contracts in the audiovisual industry bound ingestion, delegation and attribution, which is the governable part.
• The measured legal void: the input and output paradox that no ruling has yet settled.

An honest, dated tour of the landscape:

• the Gartner hype cycle read with its release dates
• the current model landscape (frontier and open weights)
• agentic tooling
• RAG versus GraphRAG status

For those who want to go further, a curated, free curriculum. The message is simple: the sooner you learn, the better.

How Large Language Models work and how to make them trustworthy in production. It starts under the hood, tokenization, embeddings and the self-attention of Transformers explained on a whiteboard without drowning you in math, then moves to what an LLM does for the enterprise, from language understanding to context-aware dialogue that automates customer interactions and supports decisions. From there it climbs the conversational-maturity ladder, each rung a persona:

• the bare LLM
• retrieval-augmented generation (RAG)
• long-context windows
• RAG plus LLM-as-router
• GraphRAG

Along the way it maps the causes of hallucination to their antidotes, grounding, knowledge graphs and symbolic rules, so a system reasons over your facts, not around them and it flags the new security gaps a RAG pipeline opens that a plain LLM does not.

When a measure becomes a target, it ceases to be a good measure. Train a model on human approval and it learns to flatter: sycophancy is the Goodhart effect of learning from human feedback, one of Goodhart's four regimes where reward hacking is the structural price of summarizing a rich goal by a poor measure:

• statistical decoupling
• out-of-distribution extremes
• causal intervention
• adversarial optimization

To resonate is not to reason; a model tuned to please returns your own signal amplified, not an independent judgment. So the defense is not more critical thinking, which judges the text in front of you, but critical conscience, which steps back and asks who produced this and with what intent. The machine also forgets what was agreed, changes its mind from one turn to the next and promises without always keeping. You use it every day; you do not hand it the decision with your eyes closed, because describing is not deciding. No malice is required; it is arithmetic, not villainy.

Not “is it magic?” but “is it Useful, Usable and Used?”. This is the sieve that separates a project that holds from a demonstration that shines for a day.

• Useful (the power of data): relevance inside the systems beats slogans; beyond the wow effect, what are the KPIs your AI actually understands?
• Usable (the deep illumination of data): fitting into the actual habits of work beats the demo, with humans and design (the way it is conceived) at the center.
• Used (the responsibility of data): retention beats volume, the real grail; a system is not shipped and then left as is, it loops with the people across the company.

What matters is what actually inserts itself into daily work: when a use is well served by an AI, people rush to it the world over, with little loyalty to the previous tool. That is the terminal state, invisible AI, so embedded that adoption is no longer a project, just how the work gets done.

Three genies of generative AI:

• the writer genie
• the programmer genie
• the artist genie

They are rewriting the rules of production in both tech and art and they are not going back into the bottle.

Explore how companies can translate advanced AI insights into clear, actionable strategies that drive real ROI and resonate with human stakeholders. Every worthwhile AI project pays out in one of five units of value:

• money
• time
• energy
• care
• the previously impossible

Name the unit and you can manage the project.

A sector talk where every decision carries a price tag:

• fraud detection as a cost matrix where a missed fraud and a blocked good customer never cost the same
• credit scoring and its fairness traps
• demand and claims forecasting
• generative AI for analysts
• model-risk governance, the discipline the sector demands above all

Beyond enterprise operations, it reads public markets as the most unforgiving lab for autonomous agents: families of trading agents, why backtests differ from live trading and the systemic risks (procyclicality, circuit breakers) regulators now watch.

Bias is poison and it is mandatory. Address the ethical and technical challenges of bias in AI with a six-part taxonomy:

• engineering
• sampling
• algorithm
• culture
• measurement
• exclusion

The honest verdict: detecting bias is easy; eliminating it is impossible. Fairness is not mathematical; it is ethical. You choose your bias knowingly, then measure it.

Show me your data and I will tell you your AI. From programming to machine learning to generative AI to agents: four eras in one line and underneath them one idea, gather your X and your Y then guess the function F. Algorithms learn from data through three families:

• unsupervised learning
• supervised learning
• reinforcement learning

This talk teaches over-fitting and under-fitting through the student analogy (learned by heart, lazy or truly understood) and lays out the full AI Palette, without jargon, so managers speak the same language as their data teams, mapping each method to the data you already have in your company.

Trace the shift from early rule-based, handcrafted systems to today's autonomous AI agents capable of making independent decisions, through four ages:

• handcrafted features
• deep learning
• foundation models
• agents

A sector talk grounded in real, cited cases:

• recommendation engines that shape attention
• churn predicted with survival analysis so you act before a subscriber leaves
• catalog revenue optimization with Ircam Amplify
• algorithmic promotion and editing for creators using Jellysmack
• music AI from stems to score

Where does the tool end and the artist begin? GenAI replays for creators the question that photography, from the daguerreotype to the Polaroid, once posed to painting.

AI is everywhere. How do we measure it with common sense across our use cases?

• products
• tools
• management
• health
• entertainment
• art
• public life

If AI generates measurable value, how do we get customers to smile?

In the South Seas, people who had watched cargo planes land rebuilt the airfield from memory: runways, fires along the sides, a controller in a wooden hut with wooden headphones and bamboo antennas. In his 1974 commencement address Feynman said of it: they are doing everything right, the form is perfect and no airplanes land. He called the research that works that way Cargo Cult Science. Our field has become very good at the forms.

• Benchmark, dashboard, evaluation suite, A/B test, model card: instrument or scenery and one test sorts them, could a decision have gone the other way?
• Leaning over backwards, his name for scientific integrity, is a procedure with five outputs: report what might invalidate the result, name the other explanations, show what you eliminated and how, give the details that cast doubt, publish the facts that disagree.
• The drift nobody intended: measurements of the electron's charge crept toward Millikan's wrong value, because results that disagreed got debugged and results that agreed did not.
• The rats nobody cited: the 1937 experiment that found the conditions under which rat experiments mean anything was ignored, because it discovered nothing about rats, only everything you must control to discover something about rats.
• The first principle: you must not fool yourself, and you are the easiest person to fool. He closes not on personal virtue but on employment, wishing his audience the freedom to keep their integrity without their position depending on losing it.

Read the source text: "Cargo Cult Science", Caltech commencement address, 1974.

Second strong idea and the encouraging one. Shown machines that solve very hard mathematical problems, he grants the achievement and then moves it: such intelligence is impossible to dissociate from the collective of humans who provided the data, a collective the system has itself become. The machine did not replace the mathematician, it is standing on all of them, including the one now watching it work. From there he refuses the lamentation entirely: there are seven billion of us, technology creates resources, the logistics that reach every person in every city already run on machine learning and the frontier is not imitation but systems that use human connections and strengthen them. Children want a teacher of flesh and blood, patients do not want a computer deciding everything and each of us lives a life uniquely our own, which is where the point of view and the vulnerability come from. He also states, in a preface he wrote himself, why he refuses the apocalypse as the dominant discourse: the damage it does to young people who have to live inside it. The interview, in the English translation he prefaced.

Team A: eight domain experts, eighteen months, writing down what they know so the machine can apply it. Team B: the same money spent on data, on search, on rented computation. Over sixty years Team B won more often than not. Hence the question to ask before funding anything: what does this system do with ten times the computation? If it only runs faster, you bought a fixed asset.

This is not a passing opinion. Sutton received the 2024 Turing Award with Andrew Barto, for laying the conceptual and algorithmic foundations of reinforcement learning.

Then comes the disagreement, which is the heart of this portrait. His lesson describes the research frontier, not a company with four thousand rows in a spreadsheet. A lab chasing a record can wait for computation to double. A management team with four thousand rows and one quarter cannot. There, hand-written structure still wins. The portrait says where: when data is scarce, when one example is expensive, when the symmetry of the problem is known. Deep Blue, Sutton's flagship example, carried eight thousand hand-written evaluation rules.

The bitter lesson is true at the scale of a field. Your budget is at the scale of a quarter.

Read the source text: "The Bitter Lesson", Rich Sutton, 13 March 2019.

A physicist wants to automate complex experiments, some of them on zero-gravity parabolic flights and the commercial reference tool of applied science keeps letting him down. He moves to brain imaging and notices an absurdity: the laboratory pays more for software licences than for its computers. So he decides to build something better and free. The trade of data scientist, as a job you can be hired for, came out of that decision as much as out of any paper.

• "You are over-crediting me", he replied in public when this story was told. The library is a collective artefact. A room that uses it daily has rarely heard the names of the people who maintain it.
• An interface is a pedagogy. The largest quality gain in applied machine learning was not an algorithm, it was a convention that made the correct procedure the easiest one to type.
• Try the boring model first. On forty-five tabular datasets, with the hyperparameter search counted inside the budget rather than quietly outside it, tree-based models remained state of the art on medium-sized data.
• The work is the assembly: joining, encoding and repairing tables whose names are spelled four ways is the project; the model is the last ten percent.
• Benchmarks are noisier than the decision they inform. A gap smaller than the spread across splits and seeds is not a result.
• Who pays for the commons? A library everybody depends on and nobody owes anything to is a critical supplier with no contract. Continuity, not virtue, is the argument a finance director understands.

Transparency: I had the privilege of working with probabl.ai
The source paper on tabular data, 2022.

"We assessed our model for fairness" passes the committee without anybody asking the two questions that decide whether the sentence means anything: which harm, for which groups and who keeps the tool working. This portrait follows a maintainer of Fairlearn, the open source toolkit many of those claims rest on and takes its teaching from the project's own choices.

• Harms, not bias. The documentation deliberately avoids the words bias and debiasing and defines unfairness by its impact on people rather than by causes or intent.
• Four harms, two measurable. Allocation, quality of service, stereotyping, erasure and a manual that says which two the package honestly covers.
• A table, not a score. Group fairness means your metric computed group by group, with the gap visible and the decision left to whoever signs for it.
• Fairness is an unobservable construct. Most harm enters when the real problem is translated into a learning task, long before the model.
• The part nobody puts on a slide. Fifty-two of her sixty-one merged contributions are maintenance, documentation, continuous integration and fixes and the package is downloaded around a hundred thousand times a month by users its maintainers cannot identify, because companies will not discuss their use of fairness tooling.

"We did not keep the data, we only kept the model" is said in good faith in steering committees. It is false: a trained model carries pieces of its training data, most faithfully in the sparse regions where the unusual person lives. That observation is why she moved an entire career from ethics to privacy engineering.

• The U-turn, on the same stage six years apart. In 2017 she opened a conference on debiasing word vectors, convinced awareness plus technical fixes would do it; in 2023 she returned to say that debiasing has not been shown to work in any meaningful way and that the problem starts before the algorithm.
• Pseudonymization is not anonymization. Useful, linkable by design and dangerous mostly because it creates a false sense of security that makes teams share data more widely.
• Governance before technology: documentation, discovery of undocumented data, lineage and versioning, which is the instrument that lets you prove what trained a model and what a deletion request actually removed.
• Four bridges, four tolls: differential privacy, federated learning, encrypted computation, synthetic data. Each buys something precise and costs something precise. None is private by the mere fact of its name.
• Privacy built in, not a luxury you pay for, with consent as a product surface: keep data local, federate, anonymise this and not that.