tim-fan

Is anyone making LLM-in-a-box for emergency supply kits yet?

I feel that would be handy in all sorts of situations when networks are down.

Hey me too, I also have a DIY inventory system, maybe yours is more developed. But I also had the concept of items for n arbitrarily nested containers.

I also track historical movements to see which items are never used.

Recently I've moved towards everything being stored in numbered bags, which are hung in order on a line for O(N) retrieval. For storage it tells you which bag to put it in, for retrieval it tells you where it is.

I'm thinking more and more the optimal system will have a physical as well as digital component.

Also, I feel this system would be great for shared workshops at work places and maker spaces etc. I was just rummaging through our lab at work today, there's so many parts in the lab no one would know about, if it was inventorised with a good integrated (AI?) search function the equipment could be much more useful/available.

https://github.com/tim-fan/hordor

OK right, so they're at 170 million now, maybe ~1 life saved. I see it might be a bit early to be making claims about that.

Also in comparison the cases of disruption/blocking intersections/emergency-services do seem significant - seems plausible in the right circumstances you could lose lives.

Seems they need to 10x their miles before they can start making confident claims about lives saved.

Can they just report directly how many human lives they have saved since beginning operation? (of course, within some error bars)

Maybe that's too much of a statistical stretch.

But would be a good to-the-point number to have on hand for some waymo debates.

"yes they caused some disruption in an intersection in so-and-so scenario, but on the other hand they saved X number of human lives last year"

Here's the repo, I've just added a demo video to the readme:

https://github.com/tim-fan/hordor

I was learning Django when I wrote it, today you'd probably get further quicker vibe coding from scratch.

I have about 100 items in storage today, I intend to add more, would like to optimize the workflow as I scale up.

Going forward I'd like to add:

  * more optimized storage/retrieval flow. The overall goal for the project is to  minimize this friction, as far as possible
  * AI enrichment - generate descriptions, aid with search etc. I'd love to be able to query my storage "how can I connect this thing to this old speaker?" and the storage responds eg "you have this cable, this adaptor, plug that into this cable, etc"

I do enjoy thinking about the different options and tradeoffs for cost and storage/retrieval time. Also tradeoffs between time and (physical) space.

edit: formatting

I'm storing small items in numbered bags. I take a photo and have a web UI for easy searching. It also tracks which items are used frequently vs not used at all in years.

Bags are stored in numerical order for quick storage and retrieval.

With this you do your decluttering from the web interface: search for items that haven't moved in years, flag for removal.

For frequently used items the system doesn't make sense - the storage and retrieval overheads are too high. But it pays off for any item you might forget the location of, or forget if you have it at all.

I feel we're overdue to have these types of digital front ends over our household item storage.

On the other hand, insurance costs for robotaxis should be lower if they are able to drive significantly safer.

Then the one I'm more interested / excited for: optimizing the fleet for the cargo. If most trips involve single passengers, then most cars can be small electric single seaters. This can further reduce insurance costs as well as fuel, maintenance and depreciation.

I'd hope that's enough to offset the price of the sensors, compute hardware, and engineers to maintain the system.

But yes paying back investors; not sure how long that would lead to elevated costs for riders.

Not to be confused with Bonxai (Fast, hierarchical, sparse Voxel Grid)

https://github.com/facontidavide/Bonxai

Is there some connection between Voxel grids and Bonsai trees that I'm missing??

Hey, cool to see!

I'm running a similar but smaller project (5k MAU), my oldest map is central London in 1561

https://onamap.me/maps/London1561/

I got into it because I was interested in the technical challenge of registering GPS to maps which are very warped compared to reality, like very old maps or illustrated tourist maps.

My home page is here for more: https://onamap.me/

I also came across this similar project a while ago:

https://www.verbeeld.be/2024/11/17/using-gps-in-the-year-156...

Good luck continuing to build out the project!

I'm hoping for a self driving taxi + trains combo to maybe solve the problem.

For one a self driving taxi fleet could take up vastly less space - you'd no longer need one car per person, you'd need far fewer parking spots, most cars could be single or double seaters again taking less space and running more efficiently.

The space savings could be used to boost rail-based public transit options, which would see more adoption as self driving taxis make last-mile transport cheaper and easier. A bunch of positive feedback loops driving public transit adoption and improvements.

Result is cleaner and more efficient transport for all, and vast amounts of space returned from serving cars to serving people.

At least that's the dream!

Doesn't the mechanical part need to position the optics with a fair amount of precision to get a good view? Id imagine very slight movements or misadjustments could cause issues.

I don't know much about telescopes :p

Hi I've been thinking about the same thing, in the context of beancount / plain text accounting.

https://www.reddit.com/r/plaintextaccounting/s/BKsaLrfy3A

I already have thousands of labeled examples and a list of valid categories. I'm also hoping an llm will do a reasonable job.

At the moment I'm wondering what to do with all the example transaction data, as it's likely larger than the context window. I guess I could take a random downsample, but perhaps there's a more effective way to summarize it.

Cool!

Have you looked at producing 3D reconstructions over the thrown trajectory? And/or something like a gaussian splat-based representation for viewing the whole trajectory at once?

Sounds like you're describing my carpet-color classification project! [1]

Built as part of a larger carpet based localisation project [2]

1: https://nbviewer.org/github/tim-fan/carpet_color_classificat...

2: https://github.com/tim-fan/carpet_localisation/wiki/Carpet-L...

Any ideas on if/when (image+text)->image models will be available?

Would be good to be able to iterate on images (keep this, change that etc).

The use of the seed looks useful but I'm guessing it has its own limitations.

For me the intuition came from Sebastian Thrun in his Ai for robotics online course. Running through everything in 1D definitely helped get the concepts before extending to multidimensional problems.

This looks to be the course playlist: https://youtube.com/playlist?list=PLAwxTw4SYaPkCSYXw6-a_aAoX...

The kalman filter stuff starts at video "Tracking Intro - Artificial Intelligence for Robotics"

Also the free course appears to be available here, although a login is required to access:

https://www.udacity.com/course/intro-to-artificial-intellige...

I appreciate the rigour, but it sounds like the argument is just "I did stop but you didn't see it"

As two flat hands collide, the air between them is forced out increasingly quickly, ultimately exceeding the speed of sound. This creates an abrupt pressure change, resulting in shock waves that make up a large part of the noise we hear

So the clap is a mini sonic boom between your hands?

Yes! New Zealand published current covid locations of interest (where infected people had been) to a github repo - I created an animation of locations over time by crawling the repo history. It gave an idea for how it was spreading through the country, and Auckland in particular.

Animation: https://raw.githubusercontent.com/tim-fan/media/main/2021092...

Code: https://gist.github.com/tim-fan/5f601c274a30505b1ae6b989a015...

Or nook (rhymes with book)? At work I say nuck, but most others say nook, so I'd like to get to the bottom of this!

I believe in some conditions fat and/or protein can be used to restore glycogen via Gluconeogenesis https://en.wikipedia.org/wiki/Gluconeogenesis

I remember reading that for keto diets it's important to not eat too much protein, otherwise the excess protein may be converted to glucose, replenishing glycogen and knocking you out of ketosis.

When I was relocating internationally, it was keeping track of the task dependency structure that stressed me out. After doing X and Y do Z.

I ended up jotting down the tasks/dependencies in dot syntax to quickly get items recorded and out of my mind during the day, then at the end of the day updated rendered the whole depency graph to give the overview.

Here's the scripts I used

https://github.com/tim-fan/task_dependency_tracking_tool

I just give you two books that you've read and you tell me which is better.

I played around with a tool for sorting through my personal photos based on that idea.

It gets interesting when you start thinking of how to optimally choose pairs to compare, to maximise signal and minimize redundant comparisons. This becomes more important as the size of the set of objects you are comparing becomes large.

https://github.com/tim-fan/image_sorting

I made a robot location tracking system specifically for use on the carpet at my office.

The carpet was an arrangement of 4 particular colors tiled in squares, so I manually made a carpet map (a few hours in excel!), wrote a carpet color classifier to run from under-robot camera data, then integrated with a particle filter for location tracking. Write up is here:

https://github.com/tim-fan/carpet_localisation/wiki/Carpet-L...

I wrote it only ever expecting usage in this particular office, but if anyone has a similar carpet and a robot that needs localizing, please reach out!

Seems like the key to consistency will be generating multiple images in a sequence simultaneously, rather than generating frames independently. Are models being developed along these lines?

Trying to understand exactly the function that is being computed - is it a weighted sum of the inputs?

Relevant xkcd: https://m.xkcd.com/2682/

And from the alt-text:

"Friction-driven static electrification is familiar and fundamental in daily life, industry, and technology, but its basics have long been unknown and have continually perplexed scientists from ancient Greece to the high-tech era. [...] To date, no single theory can satisfactorily explain this mysterious but fundamental phenomenon." --Eui-Cheol Shin et. al. (2022)

Further to indexing conversations, it would be interesting/helpful to be able to pull out:

  * conversation length
  * participants
  * location

I'm not sure how easy it is to identify/track different participants in a conversation.

Edit:formatting

The quest won't work in a car as it uses gyros.

I did try the quest 2 on a plane and it worked well, apart from when the plane was turning or in turbulence. I suppose the visual tracking makes an assumption that the world is static, which is usually true but breaks down when the plane is bouncing up and down.

Would anyone know where to look to aquire that sort of monochrome projector?

And in general how to implement a camera->projector loop with that low latency / high frame rate?

I've been looking into a similar idea and have bodged together something like: USB camera -> laptop (opencv) -> consumer LED projector (via Hdmi). Probably unsurprisingly, latency is a problem.