porridgeraisin

Large programming assignments also work decent (for now at least), where codex is encouraged, but we make the assignment research oriented and obscenely large with lots of messy and dependent steps, enough that you have to steer enough to do context management and not output slop reports/inferences. It is almost like a research project, where AI's "taste" is still extremely lacking (we often see totally circular/trivial-but-dressed-up investigative results from sloppers, very much like the slop neurips/iclr papers these days). Just one of these per course with a 10 minute in person viva at the end. Previously this used to be a project part of the course, but the old way with reports and presentations just ended up with way too much slop to use as a signal. Also, we just made it individual instead of a group project.

For this type of course, grading on a curve also made less sense, so we instead do a kind of "cluster" based grading. We start from the top giving S, see where the first big gap is, that starts A, the next big gap starts B, and so on. These clusters form really well in the above type of courses (empirically, undergrad lin alg, undergrad databases, and grad RL courses) and matches in-class judgement.

As forecasted

Well... obviously. What the fuck is the point of export controlling an already open model.

> for future ones

Nope. >2T models (i.e todays frontier) is explicitly not allowed.

Also see: https://www.reuters.com/world/asia-pacific/china-proposes-ad...

> China isn't the US. They don't share the same ideology, strategy or conclusions.

Money and power know no ideology. They do pretty much the exact same things as the US as well (from a 3rd country's POV). They expand territory, they capture industries, etc, etc, etc,.

> Keynote speech by Chinese President Xi Jinping

Do you take the same literal interpretations of other politicians' speeches? Surely you know thats politicking, the same way every G2/G7 leader says they want a collaborative world blah blah but it's naive to not know that it is cutthroat competition in reality.

> they would give it away to destroy the capitalist mode of production

Well, I won't take that dream away from you, but you're gonna be disappointed in this particular case, in our lifetime, with this particular country, atleast :)

Oh come on, this post sounds like one of those twitter bots. No one seriously expected them to just give away billions in R&D for free, and they arent. The chinese model companies are actively moving away from the open source setup. Even state export controls.

Yeah, for that, distribution via SaaS-ish products is important right? My thinking for now is that 10 years down the line, the responses API is not going to be what any of these guys will be selling. Rather it will be like whatever salesforce is doing, or that recently announced GPT-Synopsys product. Or stripe, AWS, type companies will sell it integrated here and there. Chinese software doesn't have that same reach right.

Within china of course they can scale out if thats what you meant

Makes sense. A lot of people I have seen that use chinese models, just download it and host it on US or their country's infra, so no money goes to alibaba. If they are OK with API access, they just use anthropic or openai. This move will basically cut that tap and make it so if you want the cheaper chinese models, you have to pay deepseek or alibaba themselves.

I have bad news for you...

Yep. After this has been done, the later text to sql outputs works half decent. The major issue I heard was that whenever a small assumption changes few months down the line, while humans are able to "JIT" apply it to each of their queries, these models sometimes lose it in all of the context and so it becomes hard to progressively update anything - you end up having to redo the whole workflow.

And in my experience, even the more complex dashboards requested by product or upper mgmt were stood up within the day by humans. Or the next day (but that was cos some ETL would be scheduled overnight rather than immediately). And I am not sure if speeding up this part would even matter to the guy who requested the dashboard. So I am not sure if this is a major bottleneck people should try to innovate in. If openai does the work for you then great but otherwise. Humans aided by copilot-of-old style autocomplete is probably more than enough.

Needless to say, text to sql is pointless for application queries made during API calls.

These guys also love text to sql. I know 4 people each implementing text to sql for their separate companies. Or text-to-redash dashboard in some cases. Funnily enough, text to sql is one of the things where LLMs are clearly sub-human. Mostly due to lack of easy verifiability.