Tim Dettmers' lab says the research unit is now the ecosystem, and Open Source Week ships an agent harness, auto-compaction it claims beats Claude Code and Codex
- At CMU, Tim Dettmers asked his class of 150 who feared not getting a job after graduating, and about 80 percent raised their hands, which he reads as roughly 120 students concluding there is no place for them in the future.
- Dettmers argues that agents cut project work from a year to weeks or days, so the unit of research is now the ecosystem: components that build on each other rather than self-contained papers, and Open Source Week ships that ecosystem as code.
- The release promises three things: frontier autonomous research, the most efficient test-time scaling he knows of, and auto-compaction that is far more efficient than what Claude Code or Codex implement.
- The lab's work sits at the intersection of inference-serving frameworks, agent harnesses, and autonomous research systems, and targets hardware as small as a couple of GPUs or a MacBook, with the harness abstracting away technical details a user does not need.
- Commenters dispute the framing: a current mlsys grad student says academia's problem is not GPU scarcity but that the important innovations come from industry, and others cite New York Fed data showing computer engineering and computer science majors have the #2 and #4 highest unemployment rates among recent graduates.
Hacker News opinions
Absent any evidence yet, this reads like AI psychosis. Big claims need the goods, but he's an assistant prof at CMU, not the usual suspect, and Open Source Week is supposedly on the way. I'll wait and see what actually ships.
The leap from 'I'm scared I won't get a job' to 'I don't believe there is a place for me in the future' is enormous. Plenty of students worry about graduating and still get hired. I stopped reading once it was obvious the whole thing was LLM-written.
CMU is the #2 CS program in the country. If their students think they can't find work, things really are bleak.
Yeah, 'I believe both stories are wrong, and wrong for the same reason.' Anyone who writes their own essays would never write that sentence.
I'm in an mlsys lab and the sentiment is real, but GPUs aren't the whole reason. The important work happens in industry now: shaving 1 percent off TPOT saves a hyperscaler serious money, and you only get the inside context if you're there. Papers make the gap worse.
You can build an inference serving stack with hardware anyone can buy. The problem is incentives, not GPU count.
Same story in chip design. Making real chips costs too much for a university, so useful research is hard. I've basically stopped reading papers in my own area. And every funding agency wants AI work while handing out almost no GPU time.
'Demand for software engineers is higher than ever' is a wild claim to make. The market has gotten worse every year since 2022, especially at entry and mid level. Computer engineering and CS majors are #2 and #4 for unemployment among recent grads per New York Fed data.
Graduate counts grew even faster than the market did, so both claims could be true at once. Also, students being cautious about this is not irrational.
The market for reasonably experienced software people in London is hot right now. Lots of hiring, salaries going up.
The rebound is mostly in senior roles. Indeed's posting data shows everyone else, new grads included, falling out of favor. We're nowhere near where we were even a few years before the pandemic.
Entry level demand is down, experienced demand is up. That tracks with what LLMs are good at and what they aren't.
The bigger problem is that the IT industry may no longer need as many white-collar people at all. If ordinary consumers get poorer they cut discretionary spending, including IT services. Memorization matters less now; framing problems and picking tools matters more.
So you have to know the answer is 4 before you learn to add 2 + 2? That's what 'stop acquiring skills first, then solve problems' sounds like.