Jeremie Pelletier

Need more tools to bind all these feature together: twitter.com/Lambda_Coder/s…

@Kekius_Sage No. Everything has to include the understanding of everything inside its definition, its a paradox :D Paradoxes generate novelty on to infinity and beyond. Understanding more shows more to understand later.

Will we ever truly understand everything?

Watching Asmongold react to Afroman was a moment worth witnessing, truly what the Internet is made for😂

you just heard the song too, didn't you?

@shauseth best take on engineering, ngl

engineering is easy. just get rid of everything that stands between you and the desired outcome

It's funny because we get both the references here 😂

@Grummz This is how you redeem a game from yellow paint !'s

Asmongold yeets a lamb in Crimson Desert 😂

@hasen_95dx Anyone fixing this will eat a large % of nVidia's market.

@hasen_95dx "Classic examples include early GPT-4 training runs hovering around 32–36% MFU on 25,000 A100s. Recent industry commentary points to clusters wasting 20–70% of paid capacity" There's your GPU/memory shortage explained quickly. Then consider the math's not even close to optimal.

CEO whose business model relies on ever more people increasingly using ever more GPUs tells other CEO they should use more GPUs

@parmita dis-ease, it's in the name I have this weird belief diseases get easier to treat only after the patients stop ruminating/stressing/worrying. Like, take that nocebo/placebo effect right into the AI. Let that do the first pass and heavy lifting to simplify :) Am very mind-first.

For AI to solve disease, we have to make disease solvable.

@notch Feels like playing a version of the game no gamer sees. Been wishing for this ever since making use-map-settings in StarCraft back in the day: "why do I need to reboot the game just to test changes to my local map?" Just want to boot an empty shell and develop the whole live :D

@notch Dangit now I want to toy with these ideas more :D This is the sort of stuff that developing while the game runs at 60FPS makes so much fun to explore; iterate on everything from CPU side render passes to GPU kernels, while the game plays the same ~30 seconds of inputs on repeat.

Voxels come with some interesting 4D-like problems. You can't just blend voxels together for the equivalent of mip mapping for 2d textures without the insides of the object becoming visible. (proof: a two voxel thick slab with a red front and a blue back. Can't downsample that!) (well ok there might be anisotropic solutions)

@headinthebox Next step would be to replace lwjgl by automating libclang->ffm bindings; give it a .h and dylib to go. Been using this for FFI over toy languages, works surprisingly well with minimal work. Lower the bar to get Minecraft-like sims outside gaming, one step closer to Bret Victor

@headinthebox Yes! Much cleaner API than JNI+NIO if you ask me. No views needed means less casting buffer types, brings it closer to C#'s Memory<T> and Span<T>. Needs a bit of glue from clojure to resolve downcalls properly, but nothing we can't codegen at runtime :D github.com/IGJoshua/coffi

Project Panama openjdk.org/projects/panam… is God's gift to the JVM ecosystem in the new era of AI.

New game's out and main quest became: Mount Benus

@JustDeezGuy @stylewarning At least for GPU targets this seems to be the default, otherwise it lowers to tensorflow styles of overhead.

@JustDeezGuy @stylewarning Am I the odd one in thinking this automates best when the data layouts and code runs go for batching first? In that going ECS/Burst style over a subset of the language means annotations are no longer needed and the compiler sees better vectorization than most devs could?

OpenMP-style loop parallelization annotations in Coalton?