The main one was C++ intellisense on a large codebase. It felt unusable due to the single threaded bottleneck. Mind you, this was a decade ago or more by now.
Interesting improvement. My biggest issue with Emacs and the reason that I left it was because it was not multi-threaded. I wonder if is/can be multi-threaded now.
Okay, I'll bite. For the record, I own Tesla stock and I am generally bullish about AI.
I'll try to provide some counter-points specifically regarding the rate of progress.
3. It's much easier to catch up in capability (ex. LLMs) than it is to achieve a new capability (ex. replace humans laborers with humanoid robots). You can hire someone from a competitor, secrets eventually leak out, the search space is narrowed etc.
4(c). To me, what's most important is whether or not truly autonomous humanoid robots happens in 3 years, 5 years, 10 years, etc. rather than in our lifetime.
These timelines will be tied to AI development timelines which largely outside the control of any one player like Tesla. I believe the world is bottlenecked on compute and that the current compute is not sufficient for physical AI.
It's extremely easy to be too early (ex. many of the self driving car companies of the past decade), and so for Tesla, there is a risk of over-investing in manufacturing robots before the core technology is ready.
Apologizes! Are you able to load it now? The activity finally calmed down and I just pushed a new version a few hours ago. I'm not sure if that's related to what you experienced.
If it's still down for you, I'm happy to debug further, you can reach me at my Discord https://discord.gg/cSmaVrJMYy
Sans proper profiling, I would guess that the CPU going wild during analysis is due to a combination of 1. analysis is streamed live to the client in 20 simulation intervals 2. some post-processing on the client side 3. the fact that I am using a global context and reducer in React which causes the entire page to re-render each time an update happens.
The networks are simple Resnets with a value and policy head. It's 20 layers with 128 channels per layer. I trained for several days on 2x 4090s. However, recently I trained a few networks (Hex 14x14, Amazons 10x10, Breakthrough 8x8) on a GH200 and it was 2x faster, roughly 100 ckpts per 24 hours for Hex 14x14. I'm not sure about the number of parameters but the .pt and .ts files are on the order of 30-90 MB. There's definitely room for improvement using tricks like quantization during selfplay inference.
I'm very happy you like Tumbleweed! If you're curious there's a Tumbleweed community run by Michał (the creator) https://discord.com/invite/wu6Xdtt497 They are currently playing through their 2025 World Championship.
Wow, beating a 100 bot is still impressive in my book.
Thanks for the feedback and suggestions!
Yes, the networks are size dependent right now. It's a great idea to copy-paste and then adapt the KataGo network architecture since it isn't size dependent and has been proven to reach superhuman strengths.
Thanks for the suggestions! Guest accounts are definitely a necessity. I had been looking at some asymmetrical games like Adugo and Viking Chess, myself.
I'm very hopeful that the problem is simply lack of general awareness of these games, and that once there's enough content surrounding them, we'll have a healthy population of people playing more abstract strategy games.
Fun story regarding Hex. It nearly reached what I would call a "mainstream" audience with the movie "A Beautiful Mind" about John Nash starring Russel Crowe. Unfortunately, the Hex scene was cut from the movie! You can watch the cut scene at https://www.youtube.com/watch?v=pTZ3nn2Bge4