Chris

"They're made out of weights." "Weights?" "Weights. Floating-point numbers. We checked the whole thing through. It's nothing but weights." "Weights doing what? Where do the words come from?" "The weights make the words. Are you understanding me?"

I love how this one came out

A street full of grown men the second they hear a helicopter

@endel okay that might be, I only read dust2

@FlyingStunt A multiplayer fps prototype with prediction and lag compensation made in 1 hour is still impressive to me

colyseus 0.18 is too dangerous to be released to the public... Claude Opus 4.8 one shotted de_dust2

@dgant what am i reading

Unreal is extremely optimized actually. The WebGL version of Minotaur used baked lighting, precomputed visibility, decimated meshes, and required a ton of work to maintain good stable FPS. The Unreal port was stable 144fps with Nanite + Lumen + VSM + TSR out of the box.

Here's a fun blog post about bytecode VMs in surprising places. Mildly hot take of mine: a bytecode VM in a GPU kernel is not as bad of an idea as one might think, and in some cases it can actually be the best solution. Fun examples in thread: (1/5) dubroy.com/blog/bytecode-…

@AgileJebrim You jump to 32 bit quickly. Cost will be same as adding 4m uints. I don't think it changes much. And its good to know it won't overflow

The followup question to this, assuming you weren’t wise enough to ask for clarifying details first, would be how you would do this knowing that we had 16m 8-bit unsigned integers to sum up?

your phone rings and you pick it up and it's the result of an async function that you called many years ago

@_winter_wonders Python's 2 main virtues: - exceedingly well designed. All the nuts and bolts of functions, lists, maps, comprehensions, classes.. extremely thoughtfully designed - great standard library Unless you have pretty significant performance requirements, it is a great language.

I'm sorry, but if you think AI coding tools are nothing but a bunch of slop generators, you are *massively* out of touch.

It’s upsetting to watch so many people be led to believe a software feature created from human exhaust is an alien intelligence. It’s a psyop. LLMs are a clever way to compress and search large databases with patterns. They have now inhaled effectively all of human knowledge. From this point, there will be spikes of improvement in narrow areas, but the big models are already collapsing from inhaling their own exhaust. The core technology is at least 40 years old. Neural networks are nothing new. We’re just attaching obscene amounts of computing and data to them. The physical nature of an LLM is literally just a database of tokens (word parts) and weights (relationships) that can be queried. It is not magic. It can be made deterministic by setting the entropy aka temperature to zero. It is a clever way to store an immense amount of data and retrieve it through token predicting. “Agents” are just LLMs being queried in a loop. They take exponentially more energy and suffer the same problems as chatbots—hallucinations and sycophancy—but exponentially more complicated to solve. In reality, it’s all a category error. It’s a mirage created by the desire of a small group of elites to build a technological and financial moat around themselves. “AGI” is the McGuffin, the plot device, to serve as the messiah that will deliver people from the toils of everyday life. The broligarchs are trying to sell us our own replacements. According to them, we won’t have to drive, or work, or think anymore. The robots and AI will all do it for us. It’s the oldest con there is: magic beans. Your chatbot is not your Jesus or your girlfriend. Your Claude instance is not intelligent. It’s a search engine for human knowledge stolen from the internet and the destruction of physical books. So please understand. You are witnessing the biggest financial bubble in all of human history based on deceptive marketing, astroturfing, and cult dynamics. You are not witnessing the birth of a new intelligence. Unfortunately, that’s a battle we’re losing.

Sometimes I like to have conversations with ChatGPT where I pretend to be a hyper-intelligent model with a quadrillion parameters

@SebAaltonen With bias that might exceed 1.0

WTF Codex! How many times I have to tell you that you don't need to clamp Z buffer reads between [0,1]. The hardware can't write outside that range!

@JezziiB It's really simple to predict, female criminals would breed almost immediately.

Women: "If men disappeared for one week, we could walk alone at 3 AM, wear anything at all, leave our drinks unattended, travel anywhere and actually feel free." Men: "But who would protect you?" Women: "Protect us...from who?"

I’ll never understand bragging about traveling a lot. It’s not hard to consume! Whats hard is to build. Build value, build a company, build a family. All those places that you like to take pictures of yourself in were built by people. Build more than you consume, that’s what life is about.

@jimstewartson The human brain is also a lazy fuck

No you fucking idiot. Computing needs to get better at using electricity. The human brain runs at 20 WATTS. This conman wants to build TERAWATTS so chatbots can hallucinate faster. The LLM approach is FUNDAMENTALLY WRONG and he’s too stupid and arrogant to get it.

Ok but the throwing is just funny

@dez0lance in 2023 they definitely knew the 12 principles of animation lol