Two very different rooms have started asking me the same question. Faculty want to know what to do with courses built around homework a language model finishes in under a minute. Companies want to know whether any of this is worth buying, and whether the data they already keep could support it. I teach graduate artificial intelligence and machine learning, I review for machine learning journals, and my current research is about what these models can and cannot be trusted to do with technical writing — so I answer from evidence rather than enthusiasm. I ask who is in the room before I prepare anything.
A student can get a clean solution to almost anything I used to assign as homework. That retires some of my habits, not my job. This is the talk about what I changed: what moved into class time, what I stopped collecting, and which things I still make students do with a pencil — read the problem, sketch the system, estimate the answer before computing it. Faculty usually leave with a short list of assignments to stop giving.
Fluent output is easy to mistake for understanding, and students make that mistake before anyone else does. I talk about what shows up in office hours and in oral checks: where the reasoning stops when I ask a second question, which students the tools genuinely help, which ones they quietly hurt. I keep what the research supports separate from what is still faculty-lounge anecdote, including my own.
A take-home problem set is no longer evidence of much. In-class problem solving, short oral defenses, projects built on data the student had to go out and collect, drafts that show their own history — these hold up, and I have run all of them in required engineering courses at full enrollment. I also cover what failed: detection software, and honor-code language that produced arguments instead of learning.
Whether to allow these tools is settled; students already use them. The open question is what they are actually good for in an engineering course. A patient explainer at midnight. A first pass at code. A translator for students working in their second language. A reader for the paragraph a student cannot fix alone. I show the assignment structures that made those work, and the three rules I ask for in return: check the units, check the source, and be ready to defend it out loud.
My current research asks a narrow version of the same question: what does a language model recover from an engineering paper — the parameters, the units, the limits the authors were careful to state — as against what its summary sounds like? For anyone who has watched a confident summary leave out the one assumption that mattered. Related work is on the Research page.
Most of these projects fail earlier than anyone expects. The data exists, but it is spread across twelve spreadsheets, three field formats, and one person’s memory. We go through what you have, how it was recorded, how much of it is consistent enough to learn from, and what it would cost to start collecting the part that is missing. Now and then the honest finding is that a query and a chart answer the question and no model is needed.
Sort the work by what a wrong answer costs. Drafting, summarizing records, sorting inspection photos, a first-pass takeoff: mistakes there are cheap and get caught. Anything headed for a stamped drawing, a permit, or a client deliverable needs a person who can be held responsible and a check that person can actually perform. Firms that draw that line deliberately do better than firms whose staff draw it for them.
Your people are already pasting company material into chatbots. That is a policy question before it is an IT question, and a ban is rarely the answer. I cover what to write down, what to log, what a vendor’s accuracy claim is worth once you test it yourself, and what firms in your sector are really doing with this — which is usually less impressive and more useful than the press releases suggest.
How to run a pilot that ends in a decision instead of a demo: one problem worth money, a measure agreed on before the work starts, a small evaluation set of real cases including the awkward ones, and a named owner for the thing once it works. Buy against build, when to stop, and how to train the people who will use it every day.
One subject treated properly for the room in front of me. The teaching and business subjects above, and the technical research talks as well: physics-informed models for tide levels, beach profile classification, sediment mechanics, and the conditions under which a learned model beats established practice.
A working session on a problem the group already has — a course being redesigned, an assessment policy being written, a process someone wants to automate. People bring the real material and leave with something written down, not a lecture with exercises attached.
An independent read on method choice, data adequacy, and how a result is being evaluated, for groups partway into something — before it has to be defended in front of a reviewer, a client, or a board.
Co-authored work on constrained modeling, environmental prediction, and the evaluation of language models on technical literature. A visit is often the sensible way to find out whether there is a paper in it.
Fluent is not the same as correct. Teaching students to tell the difference is most of the job now.
Booking
Tell me who is in the room and what you want them arguing about on the way out.