logankeenan

3090 was released six years ago and is still very relevant for running models locally.

You uncover a better design and submit a string of diffs not only implementing the task but simplifying other parts of the code too. Bonus points for doing this before you implement (make the hard change easy then make the easy change).

The last part of this really stands out. A high performer understands that software is malleable. However, the way you shape it, when things change, and how much is changed at one time matters a lot

Why’s that? What would be the alternative?

Data centers are critical infrastructure and having them near military bases could be seen as a security benefit.

The AWS status page still shows UAE as disrupted https://health.aws.amazon.com/health/status

Iroh has been amazing to work with and the engineers are so nice in the discord channel. The pragmatic approach to making p2p just work has been easy to understand. Their YouTube channel has great content too. Congrats on v1!

https://youtube.com/@n0computer

Do you have any recommendations for a live digest tool?

Do you have any recommendations for books on the history of movable type?

Patrick Rucker, one of the authors, spent time at our company, Claimable, last fall so he could investigate this issue further. Claimable is a startup where we leverage AI to help people appeal their health insurance denials. Ask me anything, I work in this space every day.

https://www.getclaimable.com/

Which brand did you get? I’ve always been hesitant to get one, I’d like one that I could reduce the amount of data sent out to the cloud

Do all drones do this now? Is this required by law for manufacturers to implement?

Regardless of how you feel about content moderation, 48 hours is a ridiculously long time given what AI can do today. That “bad” image could have been propagated around the world to millions of people in that time. It can and should be removed in minutes because AI can evaluate the “bad” image quickly and a human moderator isn’t required anymore. However, the compute costs would eat into profits…

Again, I’m not judging about content moderation, but this is an extremely weak initiative.

This holiday season, I wouldn’t buy high priced high quality items from Amazon due to concerns about counterfeit. I probably still won’t even after they’ve made this change. DTC from quality producers now have decent websites, free shipping, and good customer service. If I’m going to buy a premium expensive product, why risk it.

how many of the people backing the protest are basically holocaust deniers

I don’t understand this? My assumption is that hardliners in Iran would be holocaust deniers and not want the regime to fall?

Software that connects your own personal devices to form a p2p mesh network and also connects your friends personal mesh network. On top of that technology, the ability to chat, share media, or any other basic computing done between friends. Data is synced between devices. The goal is to make the underlying tech mostly invisible to the users. They pair devices and start doing standard software things.

We need to support farmers because a market spread too thin on farming means people would starve.

People would not starve if we stopped the ethanol mandate. In fact, corn prices would fall because the government would no longer force ethanol to be mixed with oil. Less demand would decrease the price.

I guess I’ll start to look at an alternative to GitHub self hosted runners.

It’s been awhile since I looked. What’s a good alternative?

How would we effectively regulate social media? Being the regulator could be a very powerful political tool and used to capture or maintain political power.

It’s been about a year since I looked into this sort of thing, but molmo will give you x,y coordinates. I hacked together a project about it. I also think Microsoft’s omniparser is good at finding coordinates too.

https://huggingface.co/allenai/Molmo-7B-D-0924

https://github.com/logankeenan/george

https://github.com/microsoft/OmniParser

I had no idea Romanization of Perian was a thing. Thanks for sharing!

https://en.wikipedia.org/wiki/Romanization_of_Persian

The BBC article spelled the dam's name wrong in their interactive image. It's Latyan Dam if anyone else wanted to look up more on it.

https://en.wikipedia.org/wiki/Latyan_Dam https://maps.app.goo.gl/UzQrPMR4iHRdbsuP7

Edit: TIL there can be different translations/spellings of Persian to English

Using Rust instead of a Go would provide a smaller binary since it doesn't need a runtime. Compared to JavaScript apps, it's not terrible but also not great. One thing WASM has over JS is that it can decode and compile code in parallel across multiple threads as it streams in.

https://hacks.mozilla.org/2018/01/making-webassembly-even-fa...

It's really cool to see someone else gravitate toward this idea. I think there might be some real potential in the future. I wrote about a similar idea in 2022 [0] and 2023 [1]. The service worker approach was heavily inspired by Richard Anaya's work [2]. HTMX migrating to fetch [3] makes this even easier. I had to create a xhr-fetch-proxy [4] to intercept requests with htmx today. I'm not the author, but would be happy to answer any questions.

[0] https://logankeenan.com/posts/a-rust-server-app-compiled-to-...

[1] https://logankeenan.com/posts/client-side-server-with-rust-a...

[2] https://github.com/richardanaya/wasm-service

[3] https://news.ycombinator.com/item?id=45803358

[4] https://github.com/logankeenan/xhr-fetch-proxy

Oh that’ll be much nicer. Would myCustomFetch then need to return a Response?

https://developer.mozilla.org/en-US/docs/Web/API/Response

Yep, it's super nice. The Service Workers API also makes this really easy too. I experimented with compiling a Rust Axum server to WASM and then ran it in my service worker. Also, thanks for incorporating fetch into htmx!

I thought I'd include an example of replacing fetch for anyone that come across this.

    const originalFetch = window.fetch;
    window.fetch = function(url, options) {
      if (url.includes('/api/user')) {
        const mockUser = {id: 1, name: 'John Doe', email: 'john@example.com'};
        return Promise.resolve(new Response(JSON.stringify(mockUser)));
      }
      return originalFetch(url, options);
    };

This is great! I had to create an xhr-fetch-proxy to use fetch and htmx. This change should open up a lot of fun possibilities.

https://github.com/logankeenan/xhr-fetch-proxy

Usage their customers would not pay for if it did not have a positive ROI.

I don't think we can assume that's true. Their customers are paying for it, but we don't know how profitable they are being with the AI compute they pay for.

Can you elaborate on this a bit more? Does that mean OpenAI had a ~$9.4 billion loss so MSFT needs to put 49% of that loss in their books?

I'm on the binary install with version v2.0.19. It never showed in the `/model` selector UI. I did end up typing `/model haiku` and now it shows as a custom model in the `/model` selector. It shows claude-haiku-4-5-20251001 when selected.

I'm not seeing it as a model option in Claude Code for my Pro plan. Perhaps, it'll roll out eventually? Anyone else seeing it with the same plan?

I’d really like to understand the total compute cost it takes to accomplish these tasks. I assume the compute is happening in a DC somewhere and not all onboard. Is the total cost of compute plus electricity to power the machine less than the cost of human labor to do the same task? At some point it’ll be less. If so, far out in the future until the prices make economic sense