What happened

In class he asked who was afraid of not getting a job after graduation. About eighty percent of 150 people raised a hand — roughly 120 students. At the same time PhD students email him eager to leave for a frontier lab, convinced academic research no longer matters. Both pictures, he writes, rest on one mistake: that the future belongs to whoever has the most GPUs.

The open-source week is one bundle, not a pile of notes. He promises three things: frontier autonomous research, the most efficient test-time scaling he knows, and context compaction he says is more efficient than what Claude Code or Codex implement. You can point the agent harness at a repository of CUDA kernels and leave it to keep improving them with no feedback along the way. On the Mac and Metal ports of their inference framework, that produced quantized inference of Qwen 3.6 35B-A3B at about 450 tokens per second and 1.5 bits per weight. Half precision uses 16 bits per weight; at 1.5 bits the same model uses about a tenth of the memory.

Then the hardware already on the desk. The popular local model has been Qwen 3.8 at 27 billion parameters. Their framework, he writes, runs the larger sibling — Qwen 3.8 Flash Next at 125 billion — on a single 24 GB GPU. DeepSeek V4.1, a 550B model, runs on AMD Strix, an NVIDIA DGX Spark, or a MacBook with 128 GB of memory. Context length, in his account, does not have to be managed by hand: compression and long-context handling are built in. He also describes combining these pieces into autonomous research that runs entirely locally and, by his estimate, beats deep-research systems from frontier labs, including Sakana AI and Google’s ScientistOne. On a fresh problem the agent spent about two hours on a lab machine, set a new lower bound on heuristic methods, built and tested a strong heuristic, moved closer to expensive methods trained with models, and found problems in the data the field uses for evaluation. They did not reach the overall state of the art. The same helper used to live in a Slack bot that fell over; students would write that the bot had gone quiet again.

Why it matters

If a strong model fits on a normal desktop card, “who owns more racks” stops being the only door into research. A university lab is strong exactly where a large company is thinner: its own problems, students who use the system every day, and the right to publish a harness, kernels, data, and a critique of the eval set together. Dettmers says the paper as the unit of credit is outdated, because readers should not have to stitch scattered work themselves.

The jobs point is about how work changes, not a promise that roles vanish. He does not pretend nothing shifts. He says the taught sequence — basics first, problems later — breaks, and demand for new products and hiring shows up after that break. The pessimistic reading stops at the turmoil. For an engineer this is more useful than “AI takes the jobs”: look at which models already run locally and which research loop you can assemble without someone else’s API.

In practice

Local inference does not replace a quality check. 1.5 bits per weight and hundreds of tokens per second are the author’s report for a specific Mac and Metal setup, not a guarantee for every card. An agent left alone on kernels is only useful if evaluation is cheap. Their constraint was that the eval had to fit on hardware they already had, and the problem had to be fresh, from the last few weeks.

Separate the open-source week’s promise from numbers already stated. He says two projects and four papers will ship together, a day late; this post does not name them yet. Until then the anchor is the figures he already gives, plus the honest limit that two hours did not win the whole problem, but did cast doubt on the evaluation data.

1. Before you rent a cluster, check whether the model fits in 24 GB or in laptop memory at a quantization you can accept.
2. Give a coding agent an automatic check. A multi-hour run without one is hard to tell from a polished empty report.
3. Watch token speed and whether the benchmark itself is broken — that is the failure he says they found.
4. Do not expect a chat bot to replace a harness. Their first Slack version was too brittle for daily student use.

Takeaway

Dettmers’s post is an argument that a small lab wins with local models, an agent, and open pieces people actually use, not with a GPU census. Treat the Qwen and DeepSeek figures as his report on his framework, not as an independent bake-off. The next thing to watch is whether the two projects and four papers arrive as one kit and whether they support those measurements.