Our course runs on open-weight models hosted by a university consortium. We made that choice deliberately, and these readings lay out the arguments behind it.
The choice is part of a public debate about energy, data centres, and who controls the tools that increasingly shape scientific work. The three readings below come from authors with very different interests, and you should read them with those interests in mind.
Assigned readings¶
1. The open-weights letter
In July 2026, Nvidia CEO Jensen Huang published an industry coalition letter titled “Open Weights and American AI Leadership.” The letter argues that the US should not restrict publicly downloadable model weights. It launched with 25 signatories and passed 150 within days. Anthropic did not sign.
The letter (mirrored by Microsoft; PDF at
aka.ms/openletter)
2. A skeptical response
Jensen puts his thumb on the scales against open-weights fearmongering, Tobias Mann, The Register, 27 July 2026. Free to read.
Read the letter first and then The Register’s response. The letter makes a public-interest case for open models. The Register notes that the person who convened the letter sells the hardware that every model runs on, whether open or closed. Both observations can be true at once, and the exercise is to work out what you believe given both of them.
3. The scientific case
Buhler, Pérez & Boettiger, Why scientists should lead the shift away from AI mega data centres, Nature, August 2026.
If you hit a paywall, Berkeley News covers the article openly.
This article makes a different argument from the first two and does not concern American competitiveness. It argues that open models small enough to run on ordinary hardware can use far less energy than their commercial counterparts. It also argues that scientists have specific reasons to prefer tools they control, including reproducibility, autonomy and cost.
Relevance to data science¶
Much of the current campus discussion of AI focuses on opposing new data centres, and that concern is legitimate. It often rests on an assumption that deserves examination: that AI requires a hyperscale data centre, so that any use of these tools contributes to building them.
This course is a counterexample to that assumption. The models you use are open-weight, hosted by a research consortium, and small enough that several of them would run on your laptop. Nothing you do this semester requires a commercial data centre.
The concern about energy remains, but the counterexample turns it into more precise questions:
Efficiency arguments cut both ways. Processors have become orders of magnitude more energy-efficient over the decades, yet the total electricity spent on computing has gone up. Why would more efficient models reduce total energy use instead of expanding what people attempt? (This is the same argument from the module introduction about why faster coding has never led to less coding.)
If open models are cheap enough to run anywhere, who benefits, and who loses the ability to charge for access?
The letter’s signatories argue that openness improves safety through transparency. Anthropic, which did not sign, argues roughly the opposite. What would count as evidence for either position?
Discussion questions¶
Come to class prepared to say which of the three readings you found most persuasive. Then answer the harder question of what the author of that piece stood to gain.