I don’t know anyone that is in the field and is a serious person that entirely refuted Lecun’s ideas tbf, the problem always was that he is his own worst spokesperson and reeks of fr*nch academic haughtiness

ah yes GPT-5 is gonna simplify the model dropdown they said, it's gonna be one magical option they said it has taken a week to revert back to how things were previously, now with no capacity to tell what models will actually run (because model names are Scary!)

I’m sorry I’ve been mean at times lately, I will be going back to #LiveLoveLaugh #JustAKidFromAkron mode there is a lot of wonderful people here and a never ending supply of interesting things to talk and think about, and focusing too much on the negative things of the culture or industry etc can make you forget that

thus is good advice especially in work environments in tech (and even more for remote) I think of it as adversarial honestly in that I used to do this only because I believed people would be dishonest and try to trick me or imply that I wasn’t trying

some napkin math on why Claude Code with the 200USD plan is basically unbeatable if you have >1B tokens a month and want the the "best possible model without going broke" every model cheaper than it is worse, every model more expensive than it is rarely objectively better tl;dr the only way for GPT-5 to replace Claude Code in any heavy workflow is either: you are spending fake VC / company money or its costs are subsidized through something like Cursor or an equivalent OpenAI subscription bundling codex cli usage

OpenAI docs always returns 403 when accessed from Claude Code #Nooticing

some random TT-Evolve screenshots from August 1 (last time I worked on it 🙏) Someone please bully me into releasing this for Tenstorrent folks... it ain't much, it's basically a harness that allows you to drop a TT-metal kernel, add some "EVOLVE" blocks, captures tons of profiling informations and run an LLM-based search in parallel with the profiling metrics as guidance to find optimizations (binning kernel candidates with MAP-Elites), basically an re-implementation of AlphaEvolve. It's been done for months, I just got distracted + I can rarely turn on Big Dave due to the sound and it being in my office (and my office being occupied by my Stardew Valley addicted girlfriend in the evenings). I don't think it is particularly well made, but it has a nice TUI, it can work for tt-nn modules, has some nice devx here and there, and certainly it optimizes Metal kernels (including multi-chip ones). It certainly got every single programming_examples/ kernels I've tried to be faster, optimized NoC bandwith etc...

one interesting note is that the whole “it isn’t just {{{thing}}} — it’s {{{bigger thing}}}" written quirk that‘s associated with ChatGPT is insanely common in Sonnet/Opus 4. Sonnet spams it all the time and feels very prone to sloppifying things, and Opus will slip these in at the cringiest moment, and both are prone to subtle sycophancy in chat mode (there are ways to stomp it out through grounding w/ web search etc... but even then) i guess my point is that it’s not just a ChatGPT issue — it’s really a fundamental consequence of RLHF benchmarkmaxxing western slop!

there’s something a bit sad that predicting fMRI brain responses to naturalistic video stimuli can done incredibly well with just some Data and, most unfortunately, a couple of Python lines you know what i mean? like, the code is readable and simplistic in nature (so it is the case for the avg transformers model using torch, etc) like torch and python are too good of an api over cuda kernels, it almost removes some of the charm simple ass code that calls “fairly simple very fast math” + lot of data is enough the only complex part is the ML compiler and its target’s architecture all of this is completely self-evident and well-known, it’s not a realization or anything, i just think it’s funny that predicting brain responses is 500 loc of python and some Data in 2025 (thanks to Lovecraftian complexity of modern GPUs and ML compilers being so neatly hidden away)

synthetic data of godly nushell use would 1000000x the model's IQ vs the crusty unixslop