This small learning engine is part of a larger project aimed at creating the ultimate Clash Royale bot. The reason I started the project is actually quite funny: my brother Duddy is a pro at the game, and since I couldn't beat him myself, I launched this initiative. Building an engine that could train directly on the actual game was impossible; the game lacks an API and offers no way to speed up gameplay. Consequently, I had to build a complete game simulator that runs faster, making it feasible to train a bot on it.
The link provided points to a "mini" version utilizing components I developed for the project. This version features just one attacking card and one defensive card. You begin by running a series of defensive sequences to see how well you can hold off the attack; then, you let the bot train from scratch and attempt to defend against the attack as effectively as possible. This is a very primitive version of the neural network; while the actual version contains two million parameters, this one has only a few thousand.
As for the project's current status, the simulator is very solid, but I haven't managed to produce a decent bot capable of playing at a human level, let alone at my brother's level. I am not a machine learning expert; I only possess basic knowledge from my CS degree. I received coding assistance from Claude, who acted as my partner in pair programming. Additionally, the game's simulation engine was built primarily by my friend Ambash, who is the project's second major contributor.
The project is MIT Licensed. I would really appreciate some help. Currently, the build process only works on Windows, not Linux. I would also love for someone with extensive machine learning knowledge to review the training and learning mechanism; I’m unsure of its quality right now and whether I’ve made fundamental errors that are ruining the training process every time.
They waited a whole year—until the "free year for students" promotion ended—to release their flagship model. I can't believe I've been stuck with a crappy model like the 3.1 Pro until now.
This small learning engine is part of a larger project aimed at creating the ultimate Clash Royale bot. The reason I started the project is actually quite funny: my brother Duddy is a pro at the game, and since I couldn't beat him myself, I launched this initiative. Building an engine that could train directly on the actual game was impossible; the game lacks an API and offers no way to speed up gameplay. Consequently, I had to build a complete game simulator that runs faster, making it feasible to train a bot on it.
The link provided points to a "mini" version utilizing components I developed for the project. This version features just one attacking card and one defensive card. You begin by running a series of defensive sequences to see how well you can hold off the attack; then, you let the bot train from scratch and attempt to defend against the attack as effectively as possible. This is a very primitive version of the neural network; while the actual version contains two million parameters, this one has only a few thousand.
As for the project's current status, the simulator is very solid, but I haven't managed to produce a decent bot capable of playing at a human level, let alone at my brother's level. I am not a machine learning expert; I only possess basic knowledge from my CS degree. I received coding assistance from Claude, who acted as my partner in pair programming. Additionally, the game's simulation engine was built primarily by my friend Ambash, who is the project's second major contributor.
The project is MIT Licensed. I would really appreciate some help. Currently, the build process only works on Windows, not Linux. I would also love for someone with extensive machine learning knowledge to review the training and learning mechanism; I’m unsure of its quality right now and whether I’ve made fundamental errors that are ruining the training process every time.
It’s simply ridiculous. None of their 20 new features seem genuinely usable for day-to-day life; I’m struggling to find any use cases for them. I have no idea what they were thinking with "Dots". The "Super Fast" addition is nice and important, but it’s just one small bright spot in a hellscape of useless features.
This is a wild take to me, and I don't even use Codex! I've been on Cursor for a long time, but this is finally making me consider switching (especially since OpenAI models are leaving Cursor).
- Cloud Agents are awesome, super useful, I can't go back to working locally only
- Dots have a cheesy name, but they remind me of a more polished version of Cursor Projects, which I just started using this week. You talk to a single agent who serves as orchestrator and organizer, and spins up cloud agents as needed to do research and implementation tasks. So you can just dump stuff into the thread without worrying about interrupting ongoing work. Works great.
- Sign in with ChatGPT: how is this not major news? They're launching with partnerships with a bunch of harnesses so you can use your subscription outside of Codex.
- Slack integration: now that they have cloud agents, this is a no-brainer, cursor's slack integration is endlessly useful to me. I'm constantly tagging notifications from various services that get dumped into slack and having a cloud agent dig into more details or fix an issue
I could go on, 6.1 sol, ultrafast, etc. Wild that you think all this is useless. I guess we just work very differently.
> Sign in with ChatGPT: how is this not major news? They're launching with partnerships with a bunch of harnesses so you can use your subscription outside of Codex.
Maybe the sign on process is new but OpenAI has explicitly allowed subscriptions to be used in 3rd party apps at least as long as I've been a subscriber (a few months)
> - Dots have a cheesy name, but they remind me of a more polished version of Cursor Projects, which I just started using this week. You talk to a single agent who serves as orchestrator and organizer, and spins up cloud agents as needed to do research and implementation tasks. So you can just dump stuff into the thread without worrying about interrupting ongoing work. Works great.
I don't get it, I could already do this right? I have lots of agents working and I just dump stuff in and they continue working with subagents, not sure why they need a separate thing for the same idea.
"Cloud Agents" is just a blatant copy of Anthropic's offering. "Dots" is a copy of the Grok bot. Regarding the release of 6.1 Sol - I’m not dismissing it, and the model will likely be useful for many things, but come on: there is no way your flagship model doesn't compete with Sonnet 5.5. I can't say for sure, but there must be a reason they didn't show any benchmarks. The live demos were absolutely ridiculous. Sure, there are some useful features, but overall, it feels like they're missing the mark. Instead of simplifying AI usage, they're complicating it to the point where it's no longer clear where you're supposed to do anything.
I was talking about Cursor's Cloud Agents, which predate Anthropic's offering. Obviously they're all going in this direction, it's a no-brainer. I don't think calling it "useless" makes any sense just because Cursor and Anthropic were ahead of them there (if they were, I know nothing about Anthropic's cloud agent offering).
Well, it’s a bit of an exaggeration to say it was bad. It’s simply a matter of falling in line with the rest of the industry. I was expecting something of a higher standard or something new, so I’m disappointed.
> I just don’t seem to have any problems that AI could solve in my personal life.
My microwave handle broke. I looked up the part needed to re-attach it; it's a a tiny piece of plastic. $30 each, and I need two of them, so $60.
So I grabbed photos of the part from online, which showed it against a 1inch square grid. I gave the photos to ChatGPT, told it the grid size, and asked it to make a 3D model.
It did it. I made a test print. It works perfectly.
¯\_(ツ)_/¯
I guess for me, it's been partly a matter of trying to unblind myself to problems I've grown accustomed to solving in a certain way.
I’m a bit disappointed with Sol 6.1. I suspect they didn't show the benchmarks and test results because it would have been embarrassing to reveal that their flagship model can't compete with the capabilities of Sonnet 5.5. That said, I still think this model is useful for a great many things, but it looks like Anthropic has the upper hand this time.
If its price is almost the same as Opus 5.5 but it’s ultimately not as good, I don’t understand the use case for this model. Perhaps Haiku 5.5 could fill a real need.
Cool engineering, but 17s p90 latency kind of defeats the point of a Jev-class model, which is supposed to be fast and cheap. Losing 10 points on MMLU along the way doesn't help.
A short work with Sonnet 5.5 showed me beyond any doubt that it is also an amazing model. The benchmarks that Entropic published also look crazy. Entropic has undoubtedly cracked something. It also seems to be driving openAi crazy, who have made some outrageous steps in the last few hours. openai shelves 6.1 astra. In addition to the S-1 files for IPO. I'm waiting for tonight to see what they will publish, but overall they seem to be at a loss.
The link provided points to a "mini" version utilizing components I developed for the project. This version features just one attacking card and one defensive card. You begin by running a series of defensive sequences to see how well you can hold off the attack; then, you let the bot train from scratch and attempt to defend against the attack as effectively as possible. This is a very primitive version of the neural network; while the actual version contains two million parameters, this one has only a few thousand.
As for the project's current status, the simulator is very solid, but I haven't managed to produce a decent bot capable of playing at a human level, let alone at my brother's level. I am not a machine learning expert; I only possess basic knowledge from my CS degree. I received coding assistance from Claude, who acted as my partner in pair programming. Additionally, the game's simulation engine was built primarily by my friend Ambash, who is the project's second major contributor.
The project is MIT Licensed. I would really appreciate some help. Currently, the build process only works on Windows, not Linux. I would also love for someone with extensive machine learning knowledge to review the training and learning mechanism; I’m unsure of its quality right now and whether I’ve made fundamental errors that are ruining the training process every time.
https://github.com/itzik123/ClashRoyaleAi