An AMD AI max+ 395 - I use the one from frame.work (https://frame.work/de/en/desktop) with 128GB unified RAM and it can run a 120b model (gpt-oss:120b) just fine.

See Wendel's review here - https://www.youtube.com/watch?v=L-xgMQ-7lW0

There are other mini-pc manufacturers, the mainboard is the important part.

Wow, that's quite beefy.

minio is an S3-compatable object store, the linked s3mini is just a client for s3-compatable stores.

It'd look like this: https://www.smbc-comics.com/comic/copyright

"Something went wrong. Please try again".

Do you suggest not doing it then?

I suggesting doing it faster and acknowledging it was a mistake not doing it sooner.

While you might be able to continuously update the model, are you able to continuously update the moderation of it? As the article says, it takes time to tune it and filter it; if you allow any content in without some filtering of outputs you might end up with another Tay. You'd have to think the liability would slow down the ability to simply update on the fly.

Also, if the proportion of training data available is larger for more established frameworks, then the ability of the model to answer usefully are necessarily dictated by the volume of content which is biased towards older frameworks.

It might be possible with live updating to get something about NewLibX but it probably would be a less useful answer compared to asking about 10YearOldLibY

Moderation is the real reason it will be difficult to have online learning models in production. I think the technical side of how to do it will not be the biggest issue. The biggest one will be liability for the output.

That reminds me of the story of the 500 mile email (https://www.ibiblio.org/harris/500milemail.html)

I read this yeaaaars ago. I'm about to re-read this, but before I do, I think this was the article that installed a little goblin in my brain that screams "TTS" in instances like this. I will edit this if the article confirms/denies this goblin.

EDIT: mostly, probably, sort of.

Funny story. He must thank the department of statistics for the quick turn around.

It takes too long to iterate on a character design.

For more explanation: I've been playing around with stable diffusion on my laptop recently; I have a gtx 4070 with 8GB dedicated VRAM so it's not nothing.

The main problem I have is that it takes a lot of iteration on a prompt to get what I want, at lower resolution and sampling steps, before I know that I'll get roughly what I want.

I tried making a character in Eggnog, and before I could be sure what I was getting, it told me it'd take 15-20 minutes to be ready. I worry that this will just make me wait a long time for a character that isn't what I want, and starting again too many times will put me off.

The iteration and feedback loop needs to be tighter in my opinion, or people will get unsatisfactory results and be unwilling to go back and fine tune.

Thanks, this is helpful feedback. We're definitely frustrated with how long it takes to load a character. We'll see what we can do to give a better sense of what the character will look like before the training job kicks off. We should be able to show some intermediate results.

I think you're talking about spinning up a temporary environment running the code and connecting via a local IDE to inspect it, whereas OP is talking about hosting the IDE remotely.

No, Google actually runs a remote web IDE called Cider. The latest version is derived from VSCode.

At Google, people can use "Cider" which is a web browser based IDE, and they can use a "Cloudtop" which is a desktop virtual machine provisioned via Google's cloud infrastructure, as alternatives to a dedicate physical workstation.

Nope, the VM has the git repo + IDE binaries + everything else. You can code against it using just a web browser if you want.

I use InfluxDB for this, it comes with a frontend UI and you can configure Telefraf as a statsd listener, so the same metric ingestion as datadog pretty much. There are docker containers for these, which I have added to my docker-compose for local dev.

I think it does log ingestion too, I haven't ever used that, I mostly use it just for the metrics and graphing.

Do you mind sharing any publicly available examples of this set up on github or somewhere? TIA

That sounds very promising indeed! It might be enough for what I’m after for my projects!