mips_avatar

GLM is really good, but sonnet outclasses it. Anthropic really cooked with sonnet 5+

I wish someone would make a decision model that’s as good as sonnet but like 10-100x cheaper. Because for this project which is a pretty small dataset you’re spending ~$300 on sonnet usage. like jev can probably do this task but it’s not smart enough to reliably be used

Seems like micron is pulling out all the stops to get their new line open. Source is someone who lives near the micron plant and talked about their insane hiring and work right now.

It's basically the batmobile from Dark Knight. A small bridge jumper. Still pretty cool

Do you think that vendors like nvidia will be able to keep their bios proprietary and encrypted? Would be cool if we could enable things like nvlink on consumer cores

Author here: If you want to try it without downloading I set up a huggingface space with the model + environment you can try it for yourself: https://huggingface.co/spaces/topoloop/finch (type any place question; it streams the tool calls and cites what it read)

Please don't show this comment to anyone involved in educating me

It was apparently something derogatory about the profit Muhammad according to other reporting. Though I guess news outlets are skeptical to mention the profit Muhammad.

The problem with being a safety focused AI lab, is you're also a danger focused AI lab. I don't think being danger focused leads you to build inspiring things.

We really need some open source project to manage all the shims across the world. Like in Seattle the WSDOT API has live ferry locations but they never made a GTFS feed for, so i made my own shim for it. I feel like there needs to be some way to make the data access easier for developers.

As far as I can tell is these standards (MCP/MHS/etc) are just semi obvious tool inferfaces that Anthropic uses as training scenarios.

Surprised Amazon managed to fumble mechanical turk at the same time Mercor/Scale and all these other companies started hiring humans to do data labeling tasks.

I spent multiple hours every week in various risk management meetings one semester for our drone building student team. There was no risk management done, they just made it clear they would try and expell and sue us if we tested our drone.

Thanks! It was really cool to work a layer deeper in the stack

Direct link to the repo on github: https://github.com/jonready/vllm-ios

I wouldn’t include YouTube given how only Google is allowed to index it

I hope it will force institutions to simplify entitlements. Right now the institutions can pretend like their kafkaesque system is fine because nobody manages to navigate the maze.

A single 3090 will train qwen 0.8B just fine. While it’s not a very capable model any training technique you would want to master can be used to make real progress. And all the skills you need to learn how to do this can be learned watching Andrej Karpathy’s zero to hero series (shame he quit educational content and went to anthropic)

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.

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

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

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.

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 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.

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

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.