Frontier AI on Your Own Hardware

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...

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.

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.

^ this

It's quite a leap to go from "I'm afraid of not getting a job after graduating" to "I do not believe there is a place for me in the future". Maybe it hasn't been the case for computer scientists in the past decade, but it's pretty normal to be worried about getting a job after graduating from one's college problem, even if one knows that most people get jobs when they graduate.

I stopped reading soon after that, when it became obvious that the article was LLM-written.

> I believe both stories are wrong, and wrong for the same reason. They assume the future of research belongs to whoever has the most GPUs.

As a current grad student in an mlsys lab, the sentiment is definitely true. However, I think the root cause is almost certainly not the lack of GPUs in academia, at least it's not the complete reason. The problem is the most important innovations really come from the industry right now. If you want to work on LLM inference serving, it is the frontier labs or hyperscalers that have the most incentives to solve the problems because improving the TPOT by 1% can save them tons of money. It is also much easier to catch up with the fast-growing field if you are in the industry because you get to talk to so many insiders (at least that has been my experience during the summer internship).

Papers only amplify this problem. Traditionally, academia is supposed to work on radical ideas that industries don't want to try. In recent years, these ideas are harder to get in as papers because the quality of peer review at top conferences is awful nowadays.

I'm not even going to talk about the AI slops in research papers and their artifacts. Guess why I'm posting on HN right now instead of working?

Finally, it's very disappointing to see that a professor at top school is so careless about editing stuff created by LLM. He might be busy, but the number of people that get discouraged by the LLM writing style will hurt his purpose of promoting the open source week. And apparently some people from industry had a better sense of that [1]. Yet another example of why some people prefer industry to academia these days.

[1] https://rfd.shared.oxide.computer/rfd/0576

I need it yesterday!

(At the moment, I'm using the OSS version of headroom. That was a real PITA to configure for local-only, but I think it is working reasonably well now.)

Cliff Compaction though???

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.

This also follows from how current LLMs work. In practice, when I use them in domains I already understand, they can produce very high-quality results. But in domains I do not know well, the results can be poor, and the bigger problem is that I may not even be able to judge how poor they are.

So my conclusion is that AI will reduce the number of jobs, but it will not eliminate the need for people.

In education, the value of memorization may decline in the AI era. We may instead place more emphasis on domain modeling, problem framing, or the ability to choose and use tools effectively. But the more fundamental issue is that the IT industry may simply lose the capacity to employ as many people as it once did.

More precisely, I mean white-collar labor.

I think the deeper cause is a K-shaped economy in which the lower and middle classes become poorer. When ordinary consumers become poorer, one of the first things they tend to cut back on is discretionary spending, including spending on many kinds of IT services.

The core infrastructure layer is different. Large incumbents such as Microsoft and Google already dominate much of it, and they are likely to be more resilient. Search, video consumption, and a few other essential digital services will also remain strong. But many other IT services are, in practice, discretionary goods. Those companies may be hit much harder if consumers have less purchasing power.

People talk constantly about productivity these days, but we were already living in an age of overproduction before AI. AI is moving us from overproduction into an era of explosive production. The problem is that production can expand far faster than people’s ability to consume.

The cycle is supposed to be:

*products → revenue → employment*

But if the consumers who are supposed to support that revenue become poorer, the cycle weakens. Productivity alone cannot solve that.

I agree with the author that academics need to move beyond treating papers as the primary unit of achievement. Much of what the article argues is reasonable.

But there is another difficulty. Most academics built their reputations through papers. They use that reputation to obtain speaking opportunities, consulting work, grants, and other forms of income and status. Even if one person decides to move beyond the paper-centered system, it is difficult to change much unless the larger incentive structure changes as well.

My view is that IT workers have, in a sense, been working to reduce their own jobs since long before AI. The more infrastructure becomes centralized, the more peripheral and smaller companies are squeezed first. AI is simply another example of that process.

Until recently, people often said that highly skilled IT professionals were difficult to replace. AI changes that perception. Even when it does not fully replace knowledge workers, it can put significant downward pressure on the wage premium attached to specialized knowledge.

I do think AI will raise productivity. But companies will also reduce headcount accordingly. And if purchasing power becomes increasingly concentrated among a smaller group of people, product development itself may become more biased toward the preferences of those few consumers. That can create another negative feedback loop.

The claim that universities can simply choose important problems that are cheap to validate is also more difficult than it sounds.

If validation itself increasingly depends on AI, and universities cannot afford to own enough GPUs, then they remain dependent on large AI companies. That dependency will inevitably influence which research problems are practical to pursue.

Any research program is constrained by the institutions and funding sources that make the research possible. Always.

At the same time, I actually agree with the author that universities will become more important.

People often talk about “skill” as though it were some pure and independent quantity, but in my experience hiring rarely works that way. If one candidate is highly capable without a degree and another is equally capable with a degree, employers will usually prefer the credentialed candidate.

More broadly, people tend to hire those with whom they feel cultural familiarity and trust. University networks provide exactly that. Alumni often help other alumni, directly or indirectly.

So I think universities may increasingly become both social institutions and stronger elite-training clubs.

For someone like me, coming from a poorer country and without much money, there may not be many choices in that system anyway.

Still, I think the author’s argument is far too optimistic.