lysecret

To encourage diversity between the different members in an ensemble. I think people are doing very similar things for MOE networks but im not that deep into that topic.

The thing is that regular weather forecasts are also not that great.

Im pretty deep into this topic and what might be interesting to an outsider is that the leading models like neuralgcm/weathernext 1 before as well as this model now are all trained with a "crps" objective which I haven't seen at all outside of ml weather prediction.

Essentially you add random noise to the inputs and train by minimizing the regular loss (like l1) and at the same time maximizing the difference between 2 members with different random noise initialisations. I wonder if this will be applied to more traditional genai at some point.

I wrote som truly awful code back in the day because of that but god it was glorious.

A usecase I see for this personally I have encountered a lot of “hot cache for some time then offload for historical queries” usecases which I have built by hand multiple times. This should be a great fit. E.g. write to Postgres then periodically offload to lakehouse and even query together (if needed). Very cool!

One usecase we have (we built it ourselves) is to periodically offload data from Postgres to lake house partitioned data on GCS. The way I see it this can now be done with a single query. Another one is the other way around to use posters as a query engine or to merge offloaded data with your live data.

Ong yes it works this would have made my past job so much easier.

Nice does this also allow me to write to parquet from my Postgres table?

Yea I use gcp vertex for this it already has Claude hope for OpenAI’s models eventually too.

I generally found 4o-transcribe to be more performant than gemini fyi.

Yea it failed on testcontainers for me. The pnpm install worked fine though.

Just played around with it the fact it’s on the phone is a big bonus.

I have setup a little workflow where given linear tags it sets up a work tree on my dev box installs deps and starts the implementation so I can take it over I prefer this workflow to the fully managed cloud based solutions.

This kind of fits in for issues where I’m basically sure I won’t have to take it over (and it can do it fully on its own). Which aren’t that many.

Very simple example there was a warning pop up on something where I thought there shouldn’t be now it’s done fully automatically from my phone in 5 mins. I quite like that these small changes become so easy.

He’s the reason I got into ml in 2016 I owe him the world. Great interview I was a bit surprised how little use he got out of agents but it makes sense I wonder how he feels about having such an essential role in creating the whole vibe coding idea.

I agree was about to comment that this article fits a domain model more than a data model.

This is very funny to me I rarely found this pushback to be useful and mostly very annoying. Claude code has the opposite problem sometimes though. To some extend this ends up being a pure preference of character. You want ur ai coder to be disagreeable or agreeable. Maybe this is a good way to differentiate.

This is a fascinating aspect. I also find much more usage out of Claude code and its willingness to go along with my ideas and be “steerable” seems to be part of it with codex many times it kept on overwriting changes I told it to make. It’s a clear kind of disagreeableness trait in action. Makes me think!

+1 for him he got me into agentic coding earlier this year.

It’s so funny to me but ever since they fixed that claud bug my experience has consistently been the exact opposite. Only thing I use codex now for are quite standard things it can solve end to end (like adding new features to my crud app) anything non standard iterating with Claude yields much better results.

I fully agree. My point was more a lot of commenters seem or implicitly compare the llm based approach with some “better” or “simpler” approach which really doesn’t exist from my estimation LLMs are sota for this kind of extractions (though they still have issues).

These kinds of errors have always existed and will always exist there is no perfect way to extract info from documents like this.

I see your point here but please take a look at the “standard” unstructured pdf extraction algos they have a lot of problems as well. Llm based extraction is still (on avergad) a big improvement.

Ok so after watching the latent space pod and going through their docs my understanding of what they are doing:

- solve it is mostly an approach less so an app they also build some Jp notebook style ide with branching and deep chat integration + a stateful vm which looks neat though

- my definition of the philosophical approach is: don’t let ai ever generate code you don’t understand be there every step of the way and build things “bottom up” in a very incremental way (basically exactly how Jp notebooks have always worked)

Now my view on the philosophy I actually don’t think I agree. With Claude code and codex I feel like my preferred development flow is “top down” to declare what you want on a high level and then let ai build some complete construct (making sure everything is type safe) and then dig in and understand this construct and iron out all of the details. I don’t think I need to understand fully each intermediate result.

I absolutely fundamentally agree that every line of the final output has to be fully understood and signed off by you.

Honestly, their videos and posts are very difficult to understand so if I misunderstood something I’m super happy to be corrected. I have high respect for the team and it’s nice to see theme excited.

Don’t forget they also have two versions for their genaisdk and you can also use their genaisdk through vertex great! Best part is all LLMs get horribly confused as well and mix different sdks etc.

Ok wow these models are great and fast! Tested it for pdf extraction tasks.

I agree this whole spec based approach is misguided. Code is the spec.

Generally use 2.5 flash for this, works incredibly well. So many traditionally hard things can now we solved by stuffing it into a pretty cheap llm haha.

Fully agree really enjoying TRPc for this.

What did you use instead?

Quite like the positioning of "this is the backend to your lovable ui", probably how chef (the vibe coding tool from the makers of convex) should have positioned it. (and kind of do).