last steps in the work on the #moxie bootstrap, #claude is at 13h43m continuous work. it's fixed many dozens of bugs, fully integrated new compilation pragmas, corrected flaws in the modified language protocols that diverge from #Go, and probably uncovered underlying errors inherited from #TinyGo that have been little landmines sitting inside the code and inherited into the moxie translation also probably messily. no idea how much longer it needs to work but i'm pretty sure i'm into the last day and a bit by now.

the moxie bootstrap is proceeding nicely. the bootstrap was MVP as of yesterday, converging at stage 3-4, stage 1 is the go compiler compiling the moxie version of moxie to stage 2, stage 3 is the moxie compiler compiling to the stage 4, and when 3 and 4 match up, the compiler is correct on its own code. so i've taken that, and now it's being exercised on the stdlib, and then will be used on iskra, iskradb, transdb (the translator engine that uses iskra and iskradb, smesh relay and web app, and the gio GUI library. it's now at the stage of the stdlib, and already nearly a dozen further code generation bugs have been found and fixed, at 7 hours autonomous run on claude code. i don't think there is another LLM that is capable of running coherently at even this amount of time like claude does, which is very convenient since i find it to be the best and i would hate to have to try and do this kind of work with other models. the big benefit for this compiler bootstrap/debugging work is that i can just fire it up and it keeps running tests, feedback from errors and other checks of the output immediately can be then used to identify bugs, figure out the fixes, and then recompile the compiler, and test its convergence. it now already converges at stage 3 where yesterday it was converging at stage 5 (ie, the hash of the binary is the same now between 2, the output of the Go version, and 3, where yesterday that took until 3->4). so as well, that step is tightening, once it applies a bug fix, it only takes two compilations to get the identical binary. meanwhile, i can just go back to bed or do something else knowing it's gonna keep doing this until it's tested everything thoroughly and maybe sometime tomorrow it will have full stage 3 convergence and compile every bit of moxie code that currently exists. next thing i plan to do after this is use the iskra language lattice to replace all this work. the converged, correct moxie compiler will still exist, but it will eventually be bootstrapped to iskra, which is the point at which iskra will be compiling itself. the bootstrap spawn of iskra, the tiny little black kitten in my dream (that's where my profile banner image comes from, i had a dream about this tiny little kitten that kept on running under my feet and i was worried i would squash her). but with a fully tested and correct, deterministic compiler available, i can then apply it to fixing all the other codebases fully. the smesh web app particularly is more challenging, as is the moxie-gio GUI library. it uses a lot more memory than it should be. with the compiler confirmed to be correct, i can now move to focusing on fixing smesh, which will soon become a lot more reliable. there is a number of bugs in smesh that i know need fixing, stuff to do with memory with the notification systems, the overall amount of memory it requires (around 500-700mb) and it's already bleeding fast as it is, most likely fixing the memory allocation errors will make it even faster and more reliable. this will also help with completing the marmot messaging protocol implementation. with the forum, filtering, identity stuff all flattened, and marmot fully working and interoperable, i can put smesh to the side and start on the moxie OS kernel. eventually, i'm going to have a full desktop, and eventually mobile (ie running on an unlocked pixel). the OS also will eventually have a full cluster distributed compute capacity that will mean you can just attach new devices running the moxie kernel to a network and instantly distribute heavy compute workloads to it without any more configuration than setting its cryptographic identity. that's a big part of the whole point of this. imagine you can come back home with a bunch of video, and then run some processing on it immediately that uses all of the computers in your house in parallel to process the data. the overhead probably somewhere around 10% for the distributed compute protocol, so every device you have it running on can contribute to doing the work. even just simple stuff, like let's say you have an LLM running and several moderately capable devices, like let's say a bunch of 16gb GPUs attached to the systems, the LLM app can run on these and coordinate and 8 of these devices will be able to do 70b models. attach another 8 and it's all transparently coordinating to do processing that will give you a token output rate starting to get into the range of full self-hosted. and that is before i get to adapting the iskra lattice to doing inference itself. very likely iskra lattice will let you do AI workloads equal to a 70b model on 128gb of memory... on your phone. offline.

after learning about the special enzyme in fresh ginger, when i experimented with dicing it finely and mixing into milk, i tried adding it to my egg-tomato-onion thing that i like making, quick and easy thing that can be cooked like an omlette or just a big muddled gob, and it is really nice - the way that the enzyme attacks the protein in the egg is different from simple acid coagulation, which tends to make eggs into a kinda grainy mess. similar to the curd it makes out of milk, which has a texture a bit like coconut flesh, but without the fibre, the egg protein is actually broken down partly, so it also sits really nice in your stomach immediately and digests better. just to explain how it works, the enzyme in fresh ginger cleaves proteins cysteine bonds, so it breaks the protein, not just deforming it as acid does. this is normally done by bile enzymes, whether you acid treat it or not. hydrolysis, using an alkali, also breaks the protein inter-amino-acid bonds but is indiscriminate and you wind up with a fine powdery substance. what the ginger enzyme does is only break some of the bonds and so it's a very nice addition to any liquid or ground form of protein cooking process. you can actually press the juice containing it out of the fresh ginger too. the proper way to do the milk is exactly this, because in the milk, the fibre of the ginger kinda ruins it. probably by cutting it moderately finely, you can separate the fibers better using an onion press, lever style, or the screw type, both work. ginger juice also has other things in it that are really great for digestion. dried ginger only gives you *some* of the anti-inflammatory oils that help heal the gut, but the fresh ginger at moderate temperature (about 70'C) actually partially digests the protein, so this trick gives you both in one. predigests part of the protein, *and* soothes irritated gut membranes. this is why confucius was famous for advising people consume ginger daily. it makes your digestion better by partially breaking down the proteins, and helps your gut heal from anything that irritates, which improves their efficiency as well.

wild stuff. i figured out how i can get claude to use a web browser with a headless python playwright thing. it's trawling websites to find fitting job openings for me right now. if you'd told me 6 months ago i'd have an AI working a web browser to save my ass i totally would have laughed at you.

the hilarious claude narrator voice is hilarious, but the point of it is quite good. nato round is for long barrel long distance engagement. the warsaw pact 7.62 is for up to 300m and pushes through light obstacles better but drops badly at long distance. .556 carbines are the wrong platform for the round, but do actually fit well with the defense use case. the lighter bullet doesn't overpenetrate as much. long barrel better if you want accuracy and lethality at range. if you get hit by one or the other, it's preferable to be the 7.62 because if it doesn't instantly kill you, you probably can survive. interesting difference between the requirements between east and west. you probably could say that the american style would work better in american landscape, the russian style, fits better in most of europe, outside of the interior of buildings. hah. anyway. these ai narrators are kinda annoying. claude especially has a really obama-style emphatic, rapid fire cadence to it. i'm used to it for discussing technical subjects but hearing it in a voice is weird.