Typo. Thanks for pointing that out.
1,628 karma · joined October 25, 2017
https://trekhleb.dev
https://github.com/trekhleb
https://www.linkedin.com/in/trekhleb/
The comparison and the reference to the karpathy/llm-council is available on GitHub https://github.com/trekhleb/yesbrainer
In short, Karpathy's llm-council is one fixed answer→rank→synthesize pass behind a local server you have to run, while Yes-Brainer is a zero-setup browser app with three deliberation structures — including a real multi-round debate (consensus mode)
I made a thin Docker container wrapper "claude-pod" recently for my personal usage here: https://github.com/trekhleb/claude-pod
However, I wasn't using it that often, just because of that additional friction of running Claude via `PORTS="3000 5173" claude-pod` instead of just `claude`, etc.
But now I have more motivation for the containerisation :D. Not a 100% defence from the potential glitches, though, but still something...
I played with similar approach in JavaScript and built a NanoNeuron https://github.com/trekhleb/nano-neuron (it is more verbose than Python though)
Does it even make sense to ask this? Is it reasonable or feasible?
I understand there are many nuances, such as the size and source of the training data, the size of the model (which would be too large for any browser to handle), network overhead, and the challenge of merging all the pieces together, among others. However, speculative calculations suggest that GPT-3 required around 3x10^22 FLOPs, which might (very speculatively) be equivalent to about 3,000 regular GPUs, each with an average performance of 6 TFLOPs, training it for ~30 days (which also sounds silly, I understand).
Of course, these are naive and highly speculative calculations that don’t account for whether it’s even possible to split the dataset, model, and training process into manageable pieces across such a setup.
But if this direction is not totally nonsensical, does it mean that even with a tremendous network overhead there is a huge potential for scaling (there are potentially a lot of laptops connected to the internet that potentially and voluntary could be used for training)?
It inspired me to experiment with a genetic algorithm in "Self-parking car evolution":
The reason why I chose GA is because I wanted to play around with this algorithm at the first place. And only after that I’ve tried to come up with some artificial problem I could try to solve with it :)
However, there is an issue right now (https://github.com/trekhleb/self-parking-car-evolution/issue...), that the cars are not “punished” for hitting another cars (they are allowed to create the road accidents). That’s why if both cars have hit another cars they may continue driving and approaching the parking lot (only approaching matters so far). That’s not good, agree. But the app is in proof-of-concept stage, so it has the issues like this.
The mental model is like this: if sensor says 4 - it means the obstacle is 4 meters away. If obstacle is far away, then sensor may say… hm… 5 meters? 10 meters? Infinity meters? So I went with something a bit higher than max sensor distance limit of 4 meters. And, for linear equation this didn’t work for me. Cars were straggling to learn.
So I’ve switched to another mental model: if sensors says 0 - it means we just turn the sensor of, the sensor is not important. Let’s say you want to learn how to drive forward if the obstacle is behind you. Then you don’t care about the side sensors, you may just cancel them with zero variables. And with this setup, the cars started to learn much faster.
I think the correct approach depends on the brain “model”. For linear equation, canceling the sensor with the zero value of the sensor.
But if you would manage to train the cars well with the different approach - it would be really interesting to try
You may also press the "Restore Evolution" button and then press "Use demo checkpoint" to use some pre-trained data.
First issue I've encountered was the text recognition performance. Depending on the camera input (if the image contained something that looked like the text or not) I've got 2-20+ seconds per 640x640px image for text recognition on iPhone X. Not so fast as you may see. But the recognition was pretty accurate though.
The performance, as expected, improves when the image size is getting smaller and the amount of text on the image is also smaller.
Since I did't want to recognise the whole text, but only the links, I've used the TensorFlow Object Detection model to quickly find the areas with the text http://**. Then, instead of recognising the whole image I needed to do it only for smaller parts of the image. This gave some improvements to the performance: from the variable 2-20 seconds per frame I've got more stable 0.5-1 seconds. Also not good, but several times faster.
I've described the challenges in more details here https://trekhleb.dev/blog/2020/printed-links-detection/. But to sum up, I had a good recognition quality with an arguable performance with Tesseract.js
https://github.com/trekhleb/machine-learning-experiments
In the repository there are several experiments, each consists of Jupyter/Colab notebook (to see how a model was trained) and demo page (to see a model in action right in the browser).
For now I've created only 10 experiments (i.e. Digits Recognition, Object Detection, Image Classification, "Write like a Shakespeare", etc.). But the plan is to do some more experimentations with GANs and RNNs.