This feels similar to not finding a game fun once I understand the underly system that generates it. The magic is lessened (even if applying simple rules can generate complex outcomes, it feels determined)
Once you discover any minmaxxing strategy, games change from “explore this world and use your imagination to decide what to do” to “apply this rule or make peace with knowing that you are suboptimal”
Maybe that's why I like Into The Breach so much, and keep coming back to it. It's a turn based strategy game, but one with exceptionally high information, compared to pretty much all the rest. You even fully know your opponent's entire next move!
But every turn becomes a tight little puzzle to solve, with surprisingly many possible outcomes. Often, situations that I thought were hopeless, do have a favorable outcome after all, I just had to think further than I usually did.
It's often a bit of a choice, though. You definitely can minmax Civilization, Minecraft, or Crusader Kings III. But then you lose out on the creativity and/or role-playing aspect.
In Minecraft, I personally want to progress in a "natural" (within the confines of the game) way, and build fun things I like. I don't want to speedrun to a diamond armor or whatever.
In Crusader Kings, I actually try to take decisions based on what the character's traits tell me, plus a little bit of own characterization I make up in my head.
My gripe with all procedural generated content in games, like e.g. Starbound. There's a tiny state space inflated via RNG, and it takes me just moments to map out the underlying canonical states and lack of any correlation between properties of an instance, or between them and the game world. The moment that happens, the game loses most of its fun, as I can't help but perceive the poor base wearing random cosmetics.
Right. However, we don't need infinite variety, or even much variety at all, if that variety makes sense, if it fits the game world and experience in some way. Pure randomness is boring and easy to dismiss wholesale once you realize there's no meaning behind it.
It seems like a very powerful injection point to provide connections between many different services, clients (besides claude desktop - I kinda want to try connecting vercel's ai client to it) and llm providers (as more providers add MCP or other adapters are added)
Carl and I are working on a PaaS for deploying agents, and emcee has given us a nice way to quickly prototype these kinds of workflows.
We can take a handful of OpenAPI specs for code deployed on our platform, and have Claude calling tools without any extra setup. That's been really helpful for figuring out how to get the LLM to pick the right tool for the job.
It's also been interesting to see just how far you can get with tool calling on its own. Normally, you have to write some kind of UI or client code to interact with a platform. But with emcee, we've been able to take a tool-driven development approach, using Huma [1] to generate OpenAPI endpoints for our Go web API, and talking to our service first through Claude, before making a UI or SDK.
And to your point, I think MCP could be really powerful as a connection point to various clients, servers, and hosts — a lot like LSP [2] for IDEs and programming languages. I'd love to hook emcee up to Zed, for example, to give the AI assistant even more context to do its thing.
Back in the day we had a CouchDB MapReduce view (on Cloudant) which took a full month to rebuild (while an angry customer was waiting). The I/O inefficiency was absolutely off the charts.
I miss my Hiptop(s). Nothing has come close to that experience since. I could carry on 10 AIM conversations, IRC, and be doing browsing and email while typing at ~100 WPM.
> To sample on Mac, uncomment line 21 in sample.py. To train on Mac, rename train_shakespeare_char_mac.py to train_shakespeare_char.py
The `mac` file changed several things - I decided to try running training with the original config file - changing device to mps / compile to false
iter 100: loss 2.0268, time 815.43ms, mfu 3.24%
iter 200: loss 1.8523, time 818.79ms, mfu 3.24%
iter 300: loss 1.7799, time 823.05ms, mfu 3.23%
iter 400: loss 1.6887, time 819.08ms, mfu 3.23%
This (the device='mps' version) already uses the unified memory plus GPU on M-series Macs.
It's possible MLX has some additional micro optimizations, but in general most people who have tried it out against hand-written MPS based training implementations haven't found great speed ups yet.
