mips_avatar

I'm not sure it would be more fun, but you could probably ingest openstreetmaps data so it's actually a real manhattan

So like oftentimes the picture will be of a church and there’s geographic coordinates for where the photo was taken. My qwen will use the geocoder to search for “church” at the coordinates of the photo and then read the Wikipedia articles about all the churches nearby and see if any of them could plausibly be the church. So far I have parsed about 2 million photos and have tagged about 800k places. My goal is to do the whole 40 million places to create a world map of open places with photos. The tool I’m using is topoloop for the geocoding

It's definitely being pushed more by the big labs recently, and it's interesting how they favor local hardware.

Maybe I'm slow but I didn't use them much until recently when Cursor and Claude Code made them a main part of the harness.

Very few people travel to visit the Amazon especially in Peru, Bolivia, and Ecuador. In the absence of tourist money these amazing ecosystems are being turned into agriculture and logged. In the protected areas supported by eco tourism the sound of the insects is deafening. But a few miles away in the banana plantations the area is completely silent. Even on “organic” banana plantations every insect is killed.

I’ve found it helps a lot with reconciliation tasks where tools like openrefine can’t handle it. Like I wanted to tag blog posts with links to Wikipedia articles that are relevant. But the thing is whats relevant changes a lot based on context in the blog post. So like a naive reconciler will tag the article for “sky” in the blogpost title “the sky above the Notre dame shone the morning we visited” when the element that should be tagged is the cathedral.

The coolest project I’ve got this running on is improving the depicts metadata for photos on Wikipedia. A lot of times they won’t have the landmarks tagged correctly in a photo. So I will load in all the metadata that exists from each photo and the pixels of those photos and give a small qwen agent access to Wikipedia search as well as a geocoder. It does a great job of figuring out what is depicted and tagging it with the correct depicts field. Im still early on but I have been able to double the number of places that have a photo attached to them on wikimedia

Qwen35ba3b can do a huge amount of data cleaning work on pretty modest hardware. Already have run about 100 billion tokens on it using 2x3090 gpus.

Well yeah Claude caching is frustrating. If implemented correctly subagents are really cheap.

One challenge/opportunity I've had is harnessing really wide running cheap agents. Any thoughts on how to move really cheap agents beyond basic summarization so we can go broader than the pricing of frontier llms allows?

even 8x rtx pro 6000 is only 768GB of VRAM. IDK how anyone is going to run k3

The more important question than subsidy is what is the tokenomics of running the model. If it's inefficient to run on an nvl72 cluster (or whatever the heck has enough vram to run a 3T parameter model), and k3 isn't very token efficient, then it might not be that compelling of an open weights model.

It would be really interesting to redo the public benchmarks for kimi k3 but token normalize the costs. Ok so maybe k3 beats fable on terminal bench, but how many tokens did it use?

I'm a creative person so my brain requires that I make something every day. Sometimes I make stuff that isn't very good. I've been told a lot by former bosses and random people that my desire to build stuff is frivolous. I could punch myself in the face with this reality, but then I would stop building.

I think we need some law that if you are above a certain scale you have to publish traffic data as a gtfs feed, ie basically apple and google have to.

The shared prompts are all cached so it's a cache read which is like 10x cheaper than a regular prefill

The problem is nobody can build a real business on models they don't control. Cursor focused on making a great AI coding experience but they didn't control Claude and got destroyed once Anthropic started explicitly training Claude to be better at using Claude code. And then there's the cost front, a good model is soon going to be served at 90+% margin, meaning the model makers have a ton of flexibility in how they choose to subsidize their first party products. Cursor had to eat the full API price of Claude while Anthropic could sell Claude Code at a massive subsidy. If you're an AI product then the model product synthesis is your product, and you don't get any of that from a proprietary closed model.

Thanks for making common crawl as good as it is. It’s a really important part of making the Internet better

I kind of wish the recent Google monopoly court ruling had forced Google to open up their index to anyone, not just Perplexity/other big players.

Qwen 35ba3b is ok, but it uses like 3x as many tokens as gemini 2.5 flash so you end up paying about the same to do an agentic run with gemini 2.5 flash and qwen 35ba3b.

I find that Gemini flash 2.5 performs about as well as Claude sonnet for non coding agentic flows except it’s actually fast enough

Yeah. Though I guess the point I thought of was like a deals site. That would have infinite pages and content

that's hard to do with rendered content, oftentimes the result depends on a backend service. Maybe you should make the service it's running public but that might be a line most aren't willing to cross.

It's such a good model for the price, for a lot of tasks it outperforms gpt5 at 3x the speed and 1/5 the price. The price jump from 2.5->3->3.5 has been so high.

I feel like the solution is a better common crawl. As nice as it would be to block the frontier AI labs from getting access to information, we should reset the baseline of information accessibility so there's less marginal advantage on these labs.

I worry a lot of the anti scraping rhetoric will just injure the open web and put somebody like cloudflare in charge.

I’m uncomfortable with how much focus ai labs have focused on replacing workers. It’s easy to tune a model on an HR job description but a lot harder to enable people to be more capable.

Maybe micron/samsung/sk won't expand enough, but trust me the Chinese will. It might mean we don't get access to HBM for a while, but you can do a lot with infinite cheap DDR5

The marginal cost of making RAM is zero and it's clear to everyone that there's near infinite demand, so the fab capacity is coming. Maybe the high end won't expand fast enough, but you can do a lot with Chinese ddr5.

Yeah but Samsung/Micron/SK hynix are building new fabs, and Chinese Ram will be able to fully supply the lower end ddr5 ram market within 2 years. The crunch is temporary and on the other side of it there will be incredible hardware at the $10-20k price point.

I know there's a lot of reasons to think that everyone will just use AI inference in the cloud, but I think if everyone had access to a dgx gb400 class machine with 512gb of hbm4 vram and 1tb of lpddr8x a lot of people are going to be running finetunes of models locally. Like the dgx gb300 is $94k now, but i bet this class of machine will come down to $20kish in the next 2-3 years.