yiyingzhang

This approach only works for small context requests. For large context and relatively smaller output (say understanding a huge code base), the cost will mainly be on prefill, and sending the large context to multiple models will only increase the cost, possibly by some factor.

With AI, the speed of finishing a project and publishing a paper is very much accelerated. But you are right that the industry role in the program diverges from the original goal of independent academic research. But my guess is that there will be students liking this program.

What you described sounds like this: https://github.com/GenseeAI/gensee-crate. It offers whole workspace fork and merge. But the merge is still somewhat cumbersome and require manual overseeing.

Many AI startups open source part of their work and write blog posts without publishing papers, including the startup I work for. Sometimes, startups just have for more pressing things to do and don't have time to write up scientific papers for publishing.

Also, there are simply too many AI papers that make peer-review publishing quite meaningless these days (e.g., AAAI this year got more than 50K submissions).

The U.S. National Science Foundation announced a $47 million investment over five years, along with additional financial commitments from nearly three dozen universities and private industry, to launch a pilot of a new four-year Ph.D. model that will provide more than 250 doctoral students nationwide with hands-on experience in industry research and development as part of their graduate education.

Cool idea! How do you guarantee privacy?

What stunned me is actually OpenAI's response to the incident. They team up with Hugging Face to put in more guardrails. Is this really the answer? Self-evolving AI can always find a way out, given more time, in a sense.

> My list is like 200 items now

Do human developers check for 200 items when they do code review? How long would that take? It's quite clear that AI code review could be better even with some error.

What's easy for human to review is also easy for AI (small code base, small PRs). What's hard for AI is also hard for human, if not more. But the time cost is so different that it's almost a nobrainer to choose AI to code and review, especially considering most software out there is not so critical :)

For academia people who care about publication, arxiv is more like a place to claim a spot before others do and before a paper gets accepted somewhere. It's also easier to get open-access publications from arxiv, as some journals are still behind paywalls.

Very good point! Accountability is very hard if the ecosystem is monopolized by a few giants.

The uncomfortable part is not that Anthropic wants to detect resellers or distillation pipelines. That is normal adversarial business.

The uncomfortable part is that a “safety” company put a covert classification channel into the system prompt of a developer tool that now routinely gets filesystem, shell, git, and browser access.

If a random npm package changed invisible-ish punctuation based on your timezone and API host so its backend could classify you, we would call it malware-adjacent telemetry. I don't see how Anthropic is different in this case.

Unfortunately, most students today just want to find the easiest way to get a good grade. The percentage of students truly want to learn is very low. For the most, they'd prefer instructors who feed them with exam problems. This is very sad, but true.

Another issue is with the curriculum and course structure, which should long been updated. But that's another rabbit hole on its own, especially in public universities with a big hierarchy of system. The sad truth is professors have no passion in teaching outdated curriculum, and students have no desire to learn.

We've been discussing this in our department for some time. No one gets fired yet :)

As a university professor, I honestly don't understand the point of grading. Who will look at and care about grades? Likely company HR. But then why should we (professors) do the screening for companies for free? Also, grades have long been inflated to a point we might as well just give everyone an A and let companies figure out how to select people.

Curious what's the typical switching frequency in your experiments and experience. How do you control the tradeoff of cache matching and model efficiency?

It all depends on the context window size. A small context size with fast performance won't be very useful today, as most workloads (like requests behind codex) usually have very long context.

How's this different from Firecracker?

This is cool! I may introduce it to the undergrad OS course I teach at UCSD. Does it have memory hierarchy?

This a a step away from university-driven research towards giving money to industry. FWIW, NSF's prior industry investment has rarely yield any true impact. Most successful startups spun from universities directly go for VC fund, esp. in CS.

Isn't this true since the beginning of software development? AI hasn't changed that yet

"Safety evals are an exception I believe eval startups can work when they're targeting safety benchmarks specifically. Researchers who want to work on safety evals tend to be ideologically opposed to working on capabilities, which means they don't migrate to post-training or applications due to monetary incentives."

This is quite interesting. Seems more relevant in 2026.

This is another Cerebras? fwiw, it took Cerebras many years to finally get a handle on the yield and the cooling problem. Wondering if they just hired a bunch of people from Cerebras.