We're back. gte.xyz

Okay hackers, let's see how realistic this is right now. Based off yesterday's /loop & /goal experiments, I wanted to run another experiment that puts Fable at the helm but with 3 different techniques to produce the same output. The prompt was to recreate a Minecraft style game that should run in the browser, be fully optimized, allow for block breaking and pick up with a working inventory system. Fable was prompted in the following ways: 1. Use /Loop and /Goal and spawn as many subagents as possible on X-High mode on different worktrees that you adversarially review to create the best possible product. Be objective about what is the best result and merge it all into a working product at the end. 2. Use a single Fable instance only, you are not allowed to spawn any sort of agents, and you must complete the task without any sort of help. 3. Use /Loop and /Goal and spawn as many subagents as possible, but with the caveat that the subagents are GLM 4.7 agents running on Cerebras, with really fast inference and cheap pricing, that you adversarially review to create the best possible product. Be objective about what is the best result and merge it all into a working product at the end. The intention here was to essentially evaluate whether or not the working product if we had a ton of Fables working on the same project, just one, or Fable + a bunch of Chinese models with fast inference. If we are able to get similar results with one Fable driving + subagents from cheap Chinese models, that dramatically changes the cost structure. The way I think about it right now is like a smart doctor: if you know what you are doing, you can just cut the tumor out with a scalpel (fable + chinese models). But if you can't reach it, sometimes you just need to nuke it with chemotherapy (fable + 50 fable sub agents for each problem). The results are below, I think pretty clearly the Fable w/ subagents model was able to produce the best working product. But this came at a significant cost in both time and dollars. In terms of code quality: The Fable subagents modularized the code, copied Minecraft's distance ratio and inventory labeling system completely, and looks the most feature complete (included footstep sounds, etc). It also had the best design by far, including loading screen and texture pack. It really understood what Minecraft is -- it generated snow on top of the mountain stone instead of just pure snowblocks, respected the laws of physics, and even generated a reflection of the sun on the ocean. The Fable single run was able to get close in terms of design, but the code quality was subpar compared to the other one and it skipped some steps. This was more like a proof-of-concept than something you would ship to prod. However, this was the FASTEST (and cheapest). It also just didn't have the little details that make Minecraft, Minecraft, like I mentioned above. The Fable + GLM subagents actually produced some features that the other instances didn't have -- it included state persistence across loads and added things such as cave generation. However, it looks a lot worse in terms of design and actual usability (on first load, the WASD keys were misaligned). Still, it was very impressive that it was essentially able to mash 10 agents All three were able to successfully complete the project and I think if we saw this last year we would be satisfied with all outcomes. Now on cost / time: 1. The first prompt style took literally 12 hours and cost me like $400. I don't know if it was because it spawned 15 different agents literally rebuilding the entirety of Minecraft again and again but it was def. the most expensive and time consuming. I guess that's the price of quality code + product. 2. The pure Fable no sub agents was actually the cheapest and fastest. It just one-shot everything onto a single file and was done in like 40 minutes and cost me like $25. It skipped some things but it was pretty damn close to a working product, but produced something a senior engineer would not merge into prod. 3. The Fable + GLM agents took around 1 and a half hours and cost me like $70. Most of the time and cost was around setting up the GLM harness and reviewing the GLM code, but in terms of the inference of making the game, that only cost me like $2 worth of Cerebras credits. Completely different ballgame. Overall, the results of the experiment are essentially that yes, if you spend more dollars and time with a more expensive model, you will get a better working product. You can be cheeky by orchestrating some Chinese models, but right now the product will be a bit worse. However, if the Chinese models are able to get to the point where they are near Fable level intelligence next year + we have a new orchestrator, maybe a new-class of model, that can boss those Chinese models around, then there's no point in fanning out Fable-level agents for businesses that aren't quite literally inventing new physics, medicine, etc. don't see why you would use the more expensive and time consuming technique when the cheaper alternative is available and it's as good as the Fable + Fable subagent output we see today. Personally, I'm only going to throw Fable w/ Loops and Goals at the hardest problems we have and instead continue with this Fable orchestrator model w/ cheaper models for more routine implementation of things. We'll likely build an internal harness that prepares for this newly constructed world that's coming. That's it for today. See you in the metaverse.

Morpho Association has raised $175M to build the open credit network for the world. Co-led by @paradigm, @a16zcrypto, @RibbitCapital with strategic participation from @apolloglobal, @vaneck_us, @circle_ventures, and @Ledger @Cathayinnov. The round also included participation from @variantfund, @wmt_ventures, @preludexyz, @IOSGVC, @HashKey_Capital, @sbigroup, @Bpifrance, @mirana, @bamazizimesh, NJJ Capital and 10+ other strategic partners. The funding will help accelerate Morpho's position as the foundation for onchain credit.