Awhile back, I was fortunate enough to be invited to a semi-weekly “build night.” It’s a casual gathering where a bunch of random folks hack on all sorts of things: software, hardware, AI tooling, art. It is awesome! There are some seriously inspiring people there.

One longtime attendee showed off a project involving 3D-printed tiki sculptures lit from within by LEDs. Each sculpture had an ESP32 inside, controlling the LEDs and communicating via Bluetooth LE. A scroll wheel controller mounted on top could change the brightness and pattern, and the effects would sync across all the sculptures. It was ridiculously cool!

I happened to randomly have an ESP32 dev kit sitting inside a box of random electronics in my garage. I ordered a few more (these things are ridiculously cheap) and decided to experiment. Despite not having touched ESP32s before, it felt like a fun challenge.

Ultimately, I ended up cutting off a few LEDs from a WS2812 LED strip, soldering some wires up to the necessary connectors and then plugged them into a breadboard.

After snipping a few LEDs off a WS2812 strip and wiring them up to a breadboard, I used VS Code and the PlatformIO extension to vibe-code a basic Bluetooth host and client (shout-out to Claude for the help!).

The result?

1. The host ESP32 registers itself as a Bluetooth service using a UUID, loops through a series of colors, and broadcasts the current color value.
2. The client ESP32 scans for that UUID. If it finds it, it listens for the color data and updates its own LED. If not, it pulses red with a slow “heartbeat” to signal it’s waiting.

You can see it in action, below. Pretty wild! It starts off with the host disconnected. The only communication between the two is via Bluetooth.

Have you ever wondered how fast your favorite LLM really compares to other SoTA models? I recently saw a Reddit post where someone was able to get a distilled version of Deepseek R1 running on a Raspberry Pi! It could generate output at a whopping 1.97 tokens per second. That sounds slow. Is that even usable? I don’t know!

Meanwhile, Mistral announced that their Le Chat platform can output tokens at 1,100 per second! That sounds pretty fast? How fast? I don’t know!

So, that’s why I put together TokenFlow. It’s a (very!) simple webpage that lets you see the speed of different LLMs in action. You can select from a few preset models / services or enter a custom speed, and boom! You watch it spit out tokens in real time, showing you exactly how fast a given inference speed is for user experience.

Check it out: https://dave.ly/tokenflow/

The code is also available on Github.

I’ve been playing around a lot with Ollama, an open source project that allows one to run LLMs locally on their machine. It’s been fun to mess around with. Some benefits: no rate-limits, private (e.g., trying to create a pseudo therapy bot, trying to simulate a foul mouthed smarmy sailor, or trying to generate ridiculous fake news articles about a Florida Man losing a fight to a wheel of cheese), and access to all sorts of models that get released.

I decided to try my hand at creating a simplified interface for interacting with it. The result: Super Simple ChatUI.

As if I need more side projects. So it goes!

There was a time (really, the past 15 years or so) where responding to things with an animated GIF was so perfect and encapsulated so much (e.g., if a picture is worth 1,000 words, what is a series of pixelated images moving a 8 frames per second worth?).

For example. see the rise of services like Giphy. I even have a random 10 year old project myself that involves animated GIFs!

Now though, it’s becoming generative AI all the way down.

For example, I just received a meeting invite that increases the frequency of meetings I’m having related to a certain project to… every single day.

Me: Hey, robot! Please create a meme image of a programmer jumping up on a desk and excitedly cheering “MOAR MEETINGS!”

Robot:

Now to figure out a way to send it in my place…

My Mastodon landscape painting bot, Mr. RossBot keeps kicking along, generating some fun landscape art. It’s been powered by the AI Horde (the open source project behind ArtBot) and has tried to utilize whatever image models provided by the API to the best of its abilities.

For the most part, the code behind it is a bunch of spaghetti that looks like this:

An update to the AI Horde late last year added support for SDXL. However, the SDXL model on the Horde did not use a refiner. Because of this, images tended to come out a bit soft and lacked texture.

You can see examples of this in my announcement post about Mr. RossBot being back, here. See also:

More recently, the Horde added support for a new image model: AlbedoBaseXL. It’s an SDXL model that has a refiner baked in. Now images will come out a lot sharper looking.

Coincidentally, I was also playing around with various prompts and discovered I could get much better image results that look more painterly (rather than simple digital renderings) by utilizing the following prompt:

A beautiful oil painting of [LITERALLY_ANYTHING], with thick messy brush strokes.

And that is it! No more messy appending various junk to the end of the prompt to attempt to get what I want. The results speak for themselves and are pretty awesome, I think!

Various Stable Diffusion models massively benefit from verbose prompt descriptions that contain a variety of additional descriptors. Much recent research has gone into training text generation models for expanding existing Stable Diffusion prompts with relevant and context appropriate descriptors.

Since it isn’t feasible to run LLMs and text generation models inside most users’ web browsers at this time, I present my “Poor Man’s Prompt Expansion Model“. It uses a number of examples I’ve acquired from Fooocus and Hugging Face to generate completely random (and absolutely not context appropriate) prompt expansions.

(For those interested in following along at home, you can checkout the gist for this script on GitHub).

How does it work?

We iterate through a list of an absolute crap ton of prompt descriptors that I’ve sourced from other (smarter) systems that tokenize user prompts and attempt to come up with context appropriate responses. We’re not going to do that, because we’re going to go into full chaos mode:

1. Iterate through a list of source material and split up everything separated by a comma.
2. Add the resulting list to a new 1-dimensional array.
3. Now, build a new descriptive prompt by looping through the list until we get a random string of descriptors that are between 175 and 220 characters long.
4. Once that’s done, return the result to the user.
5. Create a new prompt.

For our experiment, we’re going to lock all image generation parameters and seed, so we theoretically get the same image given the exact same parameters.

Ready?

Here is our base prompt and the result:

Happy penguins having a beer

Not bad! Now, let’s go full chaos mode with a new prompt using the above rules and check out the result:

Happy penguins having a beer, silent, 4K UHD image, 8k, professional photography, clouds, gold, dramatic light, cinematic lighting, creative, pretty, artstation, award winning, pure, trending on artstation, airbrush, cgsociety, glowing

That’s fun! (I’m not sure what the “silent” descriptor means, but hey!) Let’s try another:

Happy penguins having a beer, 8k, redshift, illuminated, clear, elegant, creative, black and white, masterpiece, great power, pinterest, photorealistic, award winning, vray, enchanted, complex, excellent composition, beautiful composition

I think we just created an advertisement for a new type of beverage! It nailed the “black and white”, though I’m not sure how that penguin turned into a bottle. What else can we make?

Happy penguins having a beer, volumetric lighting, Digital, intricate, awesome, futuristic, cartoon artstyle, vector, solid, detailed, dramatic light, realistic photograph, wonderful colors, dramatic atmosphere

The dude in the middle is planning on having a good night. Definitely some “wonderful colors”. Not so much realistic photo or vector, but fun! One last try:

Happy penguins having a beer, 35mm, surreal, amazing, Trending on Artstation HQ, matte painting hyperrealistic, full focus, very inspirational, pixta.jp, aesthetic, 8k, black and white, reflected on the matrix studio background, awesome

As you can see, you can get a wide variety of image styles by simply mixing a bunch of descriptive elements to an image prompt.

I’ve wanted to implement a feature like this on ArtBot for a long time. (Essentially, if the user allows it, automatically append these descriptions behind the scenes when an image is requested). Perhaps this will come soon.

Having a project hit the front page of Hacker News is a noble goal of any software engineer. Someone posted about ArtBot and it got up to position #5!

Alrighty, I updated the logic this weekend and have Mr. RossBot operating on the hairy elephant website (Mastodon). (It’s also posting on Threads, if you’re into that sort of thing.)

I also updated the image model to use Stability.ai’s swanky new SDXL model. I’m pretty impressed with the results.

Alright, this is pretty awesome! 🎉

Last night at 11:27PM, my little generative AI side project, ArtBot, crossed 10,000,000 images created. People have created some pretty cool stuff.

Some random stats from Google Analytics since it launched on October 9th, 2022 to now:

• 71,000 unique users
• 8.3 million page views.

ArtBot got another callout in PC World in the article: “The best AI art generators: Bring your wildest dreams to life.”

Though a bit of (fair) criticism at the end of the blurb though:

Why use Artbot? The vast number of AI models, and the variance in style those images produce. Otherwise, generating images via Artbot can be a bit of a crapshoot, and you may expend a great number of kudos simply exploring all the options. Since there’s no real setup besides figuring out the API key, Stable Horde (Artbot) can be worth a try.

Hey, I’ll take it!