I think the biggest problem with the models is they don’t actually have any decent lookups except chunked document embedding search

Depends on your pcie connection. If they're both x16 then it's pretty low overhead, x8 is ok, but x4 is too slow. Also it's a bit tricky getting an optimal setups with mismatched vram, I think you could probably still make use of the full vram if you're clever but it's trickier.

for layer parallelism (e.g. to get more vram) the bandwidth between layers is essentially nothing (like 16kb per token I think), so I don't think x4 would even be a problem!

Good point. It's much more of an issue when running dense models with tensor parallelism. In that case, I'd look for an MoE model instead.

Unfortunately AMD bought them, so I don't think we will get to see another release from them.

Aha, thanks, that’s fresh; press release is from Aug 6

https://ir.amd.com/news-events/press-releases/detail/1296/am...

Qwen3.5 was awesome: fairly open and fully featured. 3.8 lacking vision, nerfing thinking modes, and low context length feels pointless.

It was inspiring watching the Gemini 3 pro team launch the model and bike away into the sunset never to launch anything again

Its been inspiring to use it and know that I dont really need much more for all my intents and purposes.

I appreciate OSM for maintaining a higher data quality bar than other projects (Overture places are mostly junk outside of USA), but it's also just artificially limiting itself by not allowing streamlined paths to data contributions.

The streamlined path has always been to verify your contributions in the field. Everything else is of dubious quality.

It seems like the path is streamlined once they know who you are.

This sort of feels like a generational thing but I would have flown over.

Like I have a list of a few hundred osm places websites that are clearly scam sites. I should be going one by one and filing them manually but I found this via a spam filter and it’s very robust. I should have a way of getting these scam sites reported to osm.

You could share it here and let more people remove the spam sites. If everybody removes a couple in no-time all will be removed.

Which is probably what might happen if you would have the option to "report them to osm". Someone should check things, because maybe you are a new person to them, maybe the spam filter might not be as robust as you might think, etc. Even if you are perfect, other submission might not be, so someone needs to check/review/etc.

It is streamlined for individuals, and I have no interest in it being streamlined for anything else.

I'd rather no data than bad data. There is no seperate baby & bath water. If a bulk source of data contains an unknown mix of good and bad data, that is all one big single item of bad data that is of no use to anyone.

I’m grateful that a principled group of people run OSM. Much like I’m grateful that a principled group run Wikipedia. But the rigidness has costs that I don’t think are being appreciated.

I try to create such a thing with https://mapcomplete.org ; StreetComplete.app is also beginner friendly (but a different approach)

Yes, you'll always need a user account

Would be interesting to see how fast it would be on 4x mac studio 512gb machines.

The problem i've had with finetuning models is that most of the time better prompting beats finetuning

Better prompting doesn't improve response time or price!

Ok but a task that works fine on qwen 397b can be finetuned on qwen9b. But in every case so far when building the eval for evaluating the traces I’ve discovered a better prompt that closes the gap better than the finetuning.

Yes, it can.

Problem is right now the biggest GPU boxes they have is single rtx pro 6000s.

I’m still kind of shocked that Dean Ball can tweet such incendiary stuff about OpenAI policy. Like presumably OpenAI would prefer it if their staff don’t pick fights with Trump administration officials.

designs on the time scale I guess. and openai might feel secure in their favor with the admin?