Tim Dettmers: Academia Is About to Have a Renaissance, Not a Funeral
Frontier AI on Your Own Hardware
Tim Dettmers argues that the future of AI research belongs to universities, not just big labs with the most GPUs. In his dlab Open Source Week, he releases an ecosystem of tools: an agent harness that autonomously optimizes CUDA kernels, runs 550B-parameter models on a single 24 GB GPU, and beats frontier deep research systems entirely offline. He also introduces CliffCompaction, an auto-compaction technique that cuts agent costs by 50% and enables sessions of over 100 million tokens.
Academia is probably about to have a renaissance, and the most exciting work of the next decade will happen in university labs — not in spite of their limited resources, but because of them.
- JSavageOne
> "The prediction was that software engineers would lose their jobs first, and the recent trend went the other way: demand for software engineers is higher than ever"
Hard to take anything the author says seriously making nonsensical claims like this. The software engineering job market has been getting worse every year since 2022 by virtually every metric. This is especially true at the entry and mid level. For example, computer engineering and computer science majors now have the #2 and #4 highest unemployment rates amongst recent graduates [1]
Students are smart to be cautious about the future, and it's annoying that adults with no skin in the game so flippantly dismiss these concerns without any data to back it up.
[1] https://www.newyorkfed.org/research/college-labor-market?utm...
- wrs
> What has to go is the paper as the unit of achievement, the thing that gets counted and compared. If the ecosystem is the unit of research, then building something that other people can build on has to count for more than the next increment.
The focus on building stuff as the measure of accomplishment is one reason I so enjoyed being at CMU a few decades ago, so I'm very happy to see this sentiment is still expressed by the new generation of faculty.
- ro_bit
Many people in the thread are calling out AI-isms and nonsensical parts of the article. Articles like this have gotten so tiring. Is there a single place left to go on the internet where people write their own work? Or at least read it before publishing.
- agosz
> The difficulty did not disappear. It moved. It is no longer hard to publish a paper. It is hard to publish a coherent ecosystem.
Terrible article. It's full of these; it's tell-tale LLM. I'm surprised people post these, especially after all the articles that got upvoted that people hate LLM written articles.
Nevertheless, my curiosity got the better of me, and I thought the author might have some valuable insight further down. I skimmed forward where he talks about pessimism and came across this:
> the “software engineer” job no longer exists — and both are now within reach: agent skills come with time, and deep specialization, which used to take years, is quick to acquire with agents.
This is just plain wrong with the experience we've had ramping up juniors on my team (compiler backend). It still takes a long time to get to know the domain. Agents can't replace context and knowledge about the whole system and how it fits together with other systems and the use cases it might be subjected to. Agents can help accelerate it, but it doesn't feel like it has been "quick to acquire" for the juniors that joined my team.
- blastingrock
Horrid to see a university professor publish an AI-generated article
- vedmakk
> you have to let go of the idea that you first acquire skills and basic knowledge and then solve problems. The order reverses.
^ this
- mark_l_watson
Love this article, right on.
re: “”Start with the harness, because it is what makes everything else usable.””
I have written three harnesses this year (in Common Lisp, Python, and Racket Scheme) and it has been a fantastic learning experience. I have been using neural network tech since I was on a DARPA neural network advisory panel in the 198os, and I am even more excited (by an order of magnitude) than I was back then.
I think Tim has it right, and the hyperscalers serve a function of bootstrapping smaller systems, but they are far from the whole story.
- aabajian
This article...does not explain how to run frontier AI on your own hardware. If that were possible, it would drop the floor out from under OpenAI and Anthropic.