andrew0

~~I think it might be Magic Research 2?~~ Nevermind, I couldn't find that review on the Steam page, so it must be another game.

Organic Maps. Make sure you download the areas beforehand. The resolution for walking paths is much nicer, and you can clearly see the routes of temples and other points of interest.

It also shows data about drinking faucets, toilets, and other things provided by users.

You're right. I read past the "I want to learn ML" and went straight to "do something useful with the data".

If the goal is to understand how modern LLMs work, it's also good to read up on RNNs and LSTMs. For this, 3Blue1Brown does an amazing job, and even posted an in-depth video about transformers. I'd watch that next, followed by implementing a simple transformer in PyTorch (perhaps using the existing blocks).

You could argue that it's important to design everything from scratch first, but it's easier to first go high level, see how the network behaves, and then attempt to implement it yourself based on the paper. It is up to OP how comfortable he is with the topic though 😁

Depending on how much compute you have available, you can look into finetuning models from HuggingFace (e.g. Llama 3, or a smaller Phi model). Look into LoRA, and try to learn how the model you choose calculates the loss.

There are various ways to train, and usually involves masking the input by replacing random input tokens with the mask token. I won't go into too much detail with this, because it's a lot to explain, and I suggest you read an article on this (link1 or link2)

Hmm, might be fun to try and install custom firmware on these. I saw an article that showed a way to root them. Could be a nice mini display to link to home assistant!

The problem currently seems to be that no one sells it for less than 100$ around me right now 😅

That person clearly hasn't witnessed Dutch students carrying a whole bedroom on the back of their bike.

Good luck! You can try the huggingface-chat repo, or ollama with this web-ui. Both should be decent, as they have instructions to set up a docker container.

I believe the Llama 3 models are out there in a torrent somewhere, but I didn't dig to find it. For the 70B model, you'll probably need around 64GB of RAM available, but the 7B one should run fine with just 8GB. It will be somewhat slow though, compared to the ChatGPT experience. The self-attention mechanism can be parallelized, which is why you will see much better results on a GPU. According to some others that tested it, if you offload some stuff to RAM, you could see ~10-12 tokens per second on an RTX 3090 for certain 70B models. But more capable ones will be at less than 1 token per second, all depending on the context window you use.

If you don't have a GPU available, just give the Phi-3 model a try :D If you quantize it to 4 bits, it can apparently get 12 tokens per second on an iPhone haha. It should play nice with pooling information from a search engine, or a vector database like milvus, qdrant or chroma.

What db2 already said. Microsoft just released Phi-3 mini, which could, allegedly, run locally on newer smartphones.

If I understood correctly, the Rabbit thingy just captures your information locally and then forwards it to their server. So, if you want more power, you could probably do the same by submitting the same info to a bigger open source model than Phi-3, like Llama 3, hosted on your homelab. I believe you can set it up with huggingface/gradio, which sort of provides an API that you could use.

That way, you don't need a shitty orange box, and can always get the latest open source models with a few lines of code. There are plenty of open source frameworks in the works at the moment, and I believe that we're not far off from having multi-modal LLMs running on homelab-level hardware (if you don't mind a bit of lag).

That is good to know. Tried the free version of Roll20 before, and it definitely felt lacking in certain areas. Oh, and thanks for letting me know about the sale! I'll definitely keep an eye out for that one :)

How will you move to WhatsApp if everyone else uses iMessage? Europe has the same issue, but reversed. Everyone uses WhatsApp and can't jump to Signal/Telegram because they're not as popular.

With the way current LLMs operate? The short answer is no. Most machine learning models can learn the probability distribution by performing backward propagation, which involves "trickling down" errors from the output node all the way back to the input. More specifically, the computer calculates the derivatives of each layer and uses that to slowly nudge the model towards the correct answer by updating the values in each neural layer. Of course, things like the attention mechanism resemble the way humans pay attention, but the underlying processes are vastly different.

In the brain, things don't really work like that. Neurons don't perform backpropagation, and, if I remember correctly, instead build proteins to improve the conductivity along the axons. This allows us to improve connectivity in a neuron the more current passes through it. Similarly, when multiple neurons in a close region fire together, they sort of wire together. New connections between neurons can appear from this process, which neuroscientists refer to as neuroplasticity.

When it comes to the Doom example you've given, that approach relies on the fact that you can encode the visual information to signals. It is a reinforcement learning problem where the action space is small, and the reward function is pretty straight forward. When it comes to LLMs, the usual vocabulary size of the more popular models is between 30-60k tokens (these are small parts of a word, for example "#ing" in "writing"). That means, you would need a way to encode the input of each to feed to the biological neural net, and unless you encode it as a phonetic representation of the word, you're going to need a lot of neurons to mimic the behaviour of the computer-version of LLMs, which is not really feasible. Oh, and let's not forget that you would need to formalize the output of the network and find a way to measure that! How would we know which neuron produces the output for a specific part of a sentence?

We humans are capable of learning language, mainly due to this skill being encoded in our DNA. It is a very complex problem that requires the interaction between multiple specialized areas: e.g. Broca's (for speech), Wernicke's (understanding and producing language), certain bits in the lower temporal cortex that handle categorization of words and other tasks, plus a way to encode memories using the hippocampus. The body generates these areas using the genetic code, which has been iteratively improved over many millennia. If you dive really deep into this subject, you'll start seeing some scientists that argue that consciousness is not really a thing and that we are a product of our genes and the surrounding environment, that we act in predefined ways.

Therefore, you wouldn't be able to call a small neuron array conscious. It only elicits a simple chemical process, which appears when you supply enough current for a few neurons to reach the threshold potential of -55 mV. To have things like emotion, body autonomy and many other things that one would think of when talking about consciousness, you would need a lot more components.

I got NFS Most Wanted (2005) working in Wine, and was somewhat impressed how easy it was at the time. Game worked quite well, and would only crash once in a while with some cryptic errors that I don't remember. Made me hopeful for the future of linux gaming :)