Posts

Some thoughts on how the criteria for high-value scientific work are likely to evolve toward fundamental innovation and physical implementation. Generative AI has significantly lowered the barriers to writing and publishing. As the volume of daily publications accelerates, individual papers are increasingly difficult to distinguish within the literature. In this environment, in my view, two categories of work are likely to retain exceptional value: truly original, paradigm-shifting research, and work that translates scientific concepts into the physical world. ...

I came to control and reinforcement learning from chemical engineering, and in that passage the general formulation of the sequential decision problem was valuable to me, as a means of reading across fields in which I held no native vocabulary (thanks to legends like Dimitri Bertsekas and Warren Powell). However, for most researchers and practitioners, the object of interest is not the general formulation but the reduced problem that a particular field actually studies. ...

Interestingly, the chemical industry — and ethylene production in particular — provides a significant historical precedent related to the ongoing discussions on AI and labor. By the mid-20th century, ethylene plants transitioned to automated flow systems operating continuously with minimal human oversight. Labor productivity improvements within the industry reflect virtually constant manning levels independent of plant size. Empirical analysis demonstrates that the scale coefficient for labor is not significantly different from zero [1]. Consequently, once a fixed level of labor is reached, plant output can be increased to any level by merely increasing other inputs while holding labor almost constant. ...

TEP Studio is a new implementation of the Tennessee Eastman challenge problem. The Tennessee Eastman process has been a standard testbed for process control and fault diagnosis for three decades because it models a real industrial plant, with its full nonlinearity, coupling, and disturbance structure. TEP Studio brings that model into a form that ML and reasoning systems can use directly. Three properties distinguish it: It is a self-contained Python implementation, and it is able to simulate one process-hour in 8 ms — more than five times faster than the Fortran reference and about two orders of magnitude faster than an existing Python implementation. It allows for systematic LLM access. The process description is exposed through a Model Context Protocol (MCP) server. It is built from the bottom up to serve ML and data-intensive applications. It is also a Gymnasium environment, so reinforcement learning and control methods can train on it directly. Code is on GitHub, and you can try the simulator live in the TEP Studio demo. I’ve also written up the design principles behind it in more detail on the projects page.

Anthropic has reportedly raised funding at a valuation close to $1 trillion, and SpaceX may eventually seek a valuation near $2 trillion in its IPO. Obviously, one way to think about where things are going is to look carefully at where we are now. About a year ago, I bought an AMD Ryzen AI Max+ 395 system with 128 GB of quad-channel LPDDR5x-8000 memory. Today, I can reliably run one of NVIDIA’s most capable open models, Nemotron 3 Super 120B A12B, locally. That is an extremely impressive model to be running on a personal machine. ...

I admire what the EIA has done by publishing what they called a Retrospective on its Annual Energy Outlook reports. It is an excellent step, as it adds an important feedback loop — comparing past projections against what actually happened is exactly the kind of accountability that makes forecasting a discipline rather than a ritual. Link: https://lnkd.in/dGb_ejsb

In this post, I would like to think about agents that already run our world. They are not the fancy agents that have come to dominate the technology discourse. They are time-tested workhorses. To understand the true challenge of creating a trustworthy autonomous system, I propose to think about the most successful and widely deployed agent in history: the Proportional-Integral-Derivative (PID) controller. I choose the PID controller precisely because it is simple, real, and tested time again in the real world. It is a decision-making policy in its purest form. While the control algorithm itself is well-studied, the enormous amount of engineering required to make it operate reliably highlights the critical gap between an abstract algorithm and a functioning agent. Analyzing these necessary steps provides a crucial reality check for any credible claim about autonomy. ...

Apple’s recent research article, The Illusion of Thinking, made waves. But the reaction quickly moved beyond the technical. For some, it confirmed LLM limitations; for others, it was inconsequential and human-level intelligence is inevitable. And so the cycle continues: one side dismisses, the other inflates. This misses the point. The real question isn’t who’s right about the future, but whether these technologies are useful now, and how they might become more useful tomorrow. As an engineer, I think we should spend less energy predicting what these systems will become and more time evaluating what they allow us to do today. ...

While working on algorithms, I often thought about how to demonstrate their utility. A bench-scale setup is unlikely to capture many aspects of the complexity of industrial systems. It is too idealized and overlooks important interactions among units. Mathematical models, by contrast, can simulate much more. They can incorporate nonlinear behavior, multivariable dynamics, feedback, and uncertainty. In theory, a well-constructed simulation that captures these complexities should be more informative than an idealized lab setup. ...

The Core Problem: Optimization is Hard Plants are Dynamic: Chemical processes rarely sit perfectly still. Raw materials change, equipment ages, customer demand shifts, and unexpected disturbances occur. Trying to find the single “best” steady operating point isn’t always effective because the plant is always changing. Traditional Models are Complex: For decades, engineers have used mathematical models to understand and optimize these processes. Methods like Real-Time Optimization (RTO) try to calculate the best settings based on these models. However, creating accurate dynamic models (models that capture how things change over time) for complex, interconnected plants is incredibly difficult, time-consuming, and expensive. Keeping these models up-to-date is also a major challenge. Resistance to Change: Implementing complex new optimization systems based on these models faces resistance due to cost, complexity, and uncertainty about whether they’ll actually work reliably in the real world. The Proposed Solution: Learn Directly from Experience (Data) ...