It feels downstream of CMU's "reasonable person principle". They know that people are going to use AI on their homework, but they trust that they want to learn and improve their skills -- and this is good advice for doing so.
I'm somewhat biased because I was involved in a previous, related course. The important takeaways aren't really about gritty debugging of (possibly) large homework assignments, but the high-level overview you get in the process. AI assistance means you could cover more content and build larger, more realistic systems.
An issue in the first iteration of Deep Learning Systems was that every homework built on the previous one, and errors could accumulate in subtle ways that we didn't anticipate. I spent a lot of time bisecting code to find these errors in office hours. It would have been just as educational to diagnose those errors with an LLM. Then students could spend more time implementing cool stuff in CUDA instead of hunting down a subtle bug in their 2d conv backwards pass under time pressure... But I think the breadth and depth of the course was phenomenal, and if courses can go further with AI assistance then it's great.
This new class looks really cool, and Zico is a great teacher.
Yeah of course, I've been thinking about this a lot and I'm updating my beliefs all the time, so it's good to hear some more perspectives
A) I see what you mean. But I'm more so thinking: companies consider their models an asset because they took so much compute and internal R&D effort to train. Consequently, they'll take measures to protect that investment -- and then what do the downstream consequences look like for users and the AI ecosystem more broadly? That is, it's less about what's right and wrong by conventional wisdom, and more about what consequences are downstream of various incentives.
B) I don't really care about AI safety in the traditional sense either, i.e., can you get an LLM to tell you to do some thing that has been ordained to be dangerous. There's lots of attacks and it's basically an insoluble problem until you veer into outright censorship. But now that people are actually using LLMs as agents to _do things_, and interact with the open web, and interact with their personal data and sensitive information, the safety and security concerns make a lot more sense to me. I don't want my agent to read an HN post with a social-engineering-themed prompt injection attack and mail my passwords to someone. (If this sounds absurd, my Clawbot defaulted to storing passwords in a markdown file... which could possibly be on me, but was also the default behavior.)
C) This is a completely fair point, there's amazing work coming out of these smaller labs, and the incentives definitely work out for them to do a distillation step to ship faster and more cheaply. I think the small labs can iterate fast and make big changes in a way that the monolithic companies cannot, and it'd be nice to see that effort routed into creating new data-efficient RL algorithms or something that pick up all the slack that distillation is currently carrying. Which is not to say they're doing none of that, GRPO for example is a fantastic idea.
One way you could have a change in perspective is not just in the architecture/data mix, but in the way you spend test-time compute. The current paradigm is chain-of-thought, and to my knowledge, this is what distillation attacks typically target. So at least, all models end up "reasoning" with the same sort of template, possibly just to interlock with the idea of distilling a frontier API.
D) Interesting to hear. In my research, I find these models to be quite a bit harder to work with, with significantly higher failure rates on simple instruction following. But my work also tends to be on the R&D side, so my usage patterns are likely in the long-tail of queries.
A) The "IP" they're concerned about isn't the same IP you speak of. It's the investment in RL training / GPU hours that it takes to go from a base model to a usable frontier model.
B) I don't think the story is so clean. The distilled models often have regressions in important areas like safety and security (see, for example, NIST's evaluation of DeepSeek models). This might be why we don't see larger companies releasing their own tiny reasoning models so much. And copying isn't exactly healthy competition. Of course, I do find it useful as a researcher to experiment with small reasoning models -- but I do worry that the findings don't generalize well beyond that setting.
C) Maybe because we want lots of different perspectives on building models, lots of independent innovation. I think it's bad if every model is downstream of a couple "frontier" models. It's an issue of monoculture, like in cybersecurity more generally.
D) Is it really 90% of the performance, or are they just extremely targeted to benchmarks? I'd be cautious about running said local models for, e.g., my agent with access to the open web.
Somebody could use this as a starting point. http://touchscale.co/ You'd have to collect new data on touch strength vs. weight to get the regression parameters.
(If you do this, let me know and I can add it to the site above, and then we can both delight in the surprisingly large amount of unmonetizable traffic it gets.)
The choice of activation function isn't entirely clear to me, but I think it's definitely possible to make a network that operates entirely in the frequency domain. It would probably be pretty easy to start experimenting with such a thing with the nice complex number and FFT support in PyTorch 1.8. :)
Like you said, there's already a significant connection between convolutional networks and the Fourier domain (the convolution theorem).
Tangentially, I've recently worked on a project that focused on implementing convolution in the Fourier domain, and how that allows one to control useful properties of convolutions (like "smoothness" and orthogonality).
I'm somewhat biased because I was involved in a previous, related course. The important takeaways aren't really about gritty debugging of (possibly) large homework assignments, but the high-level overview you get in the process. AI assistance means you could cover more content and build larger, more realistic systems.
An issue in the first iteration of Deep Learning Systems was that every homework built on the previous one, and errors could accumulate in subtle ways that we didn't anticipate. I spent a lot of time bisecting code to find these errors in office hours. It would have been just as educational to diagnose those errors with an LLM. Then students could spend more time implementing cool stuff in CUDA instead of hunting down a subtle bug in their 2d conv backwards pass under time pressure... But I think the breadth and depth of the course was phenomenal, and if courses can go further with AI assistance then it's great.
This new class looks really cool, and Zico is a great teacher.