__vk__ @___vk___

You will read a reply to this post written by no human but a machine taught language from the collected writings of our species, generated by billions of microscopic switches, sent as pulses of light across oceans, displayed on a pocket supercomputer and sigh, “ugh, another bot.”

@tylertracy321 Public reaction to events of the last week has 5x'd my p(doom). I feel like we're in the movie Don't Look Up.

Now, finally, we have the first generation of CEOs who take that seriously. Isn't this just what all the activists wanted, over all those years? Not, of course, that anyone is satisfied.

For generations, activists tried to get CEOs to acknowledge the harms of their products. For generations, the message was drilled into schoolchildren, into moviegoers, into news audiences—capitalism was responsible for pollution, for carcinogens, for unemployment, for war…

You might think they're still doing far too little. Or you might think they're irrational doomers. Or you might think they're hypocrites. Or you might criticize their PR strategy and skills. But just the fact that they're taking this line is remarkable, and under-remarked.

It is remarkable how much the CEOs of the big AI labs have taken seriously the harms and risks of their own product, even well in advance of major harms showing up, and talked about them and planned for them. This has never happened in the history of capitalism, to my knowledge.

same model. same budget. different harness. who won?

Andrej Karpathy @karpathy

17 hours ago

We're starting to leave the territory where you'd test an LLM by e.g. "create an svg of pelican on a bicycle". As one idea to generalize it, I was interested what Opus 5 would do if I gave it the first paragraph of the Lord of the Rings, a 1M token budget (~$10) and asked for

1K 2K 22K 2.5M 9K

Silent Hill | PlayStation - 1999

keyboard_arrow_left Previous keyboard_arrow_right Next

@karpathy someone *please* fund doing the entire trilogy like this as a benchmark.

We're starting to leave the territory where you'd test an LLM by e.g. "create an svg of pelican on a bicycle". As one idea to generalize it, I was interested what Opus 5 would do if I gave it the first paragraph of the Lord of the Rings, a 1M token budget (~$10) and asked for three js render of it. Opus went off for ~2 hours and wrote 5500 lines of code that (procedurally) rendered the story. It's kind of janky but fun. But it's a bit mindboggling that the LLM has to place and orchestrate various polygon assets in (x,y,z) coordinates and write code that animates it all, and that it even does anything at all. I also like this kind of examples because no one in their right mind would ever spend the time to write something this custom but LLMs have all the stamina and patience in the world, so it's an example where we go from "no one would ever do this" to "sure, why not, it's ~free". There might be a lot more. But I'm excited about creating hyper custom worlds that you can imagine dropping players into, e.g. here to participate in the LoTR story as a spectator NPC, or one of the characters, or etc. Something like an ephemeral GTA of X on demand. Last thought is that the domain of worlds/games exposes a weakness in LLMs: they can't easily audit their work because they aren't able to efficiently and natively perceive videos or play games within them. Here, Opus 5 had to very slowly and painstakingly take screenshots at different points, and it messed up a few times and created a bunch of jank. An example of raw capability (multimodal, gameplay) that I think is still quite lacking.

@So8res "There Is No Rule That Says We'll Make It" by @robertskmiles youtu.be/JD_iA7imAPs?is…

See also: lesswrong.com/posts/ervaGwJ2…

x.com/i/article/2083…

This is part of why people underrate clarity. “You know what they mean!” No, they don’t even know themselves what they mean.

Yes, this is a big issue; not just with postmodernism but also with writing that builds on metaphors, prototypes you’re supposed to generalise from (how?), etc. The notion that the author has clear propositions in mind behind the mode of presentation is typically an illusion.

Jean-Claude Fox @JeanClaudeFox2

11 hours ago

An underdiscussed problem with opaque writing is that it makes it harder for the writer to see whether he knows what he means exactly and whether that idea has value.

1 1 21 8K 0

Prompt: Build me a cel-shaded arcade boat racing game in Three.js, running on an infinite open ocean, at a level of visual and mechanical quality that a AAA studio would ship. Not a demo, not a prototype — a finished, beautiful, playable game. Non-negotiable constraints Stack: Vite + Three.js (latest r1xx), TypeScript, ES modules. I must be able to run npm install then npm run dev and play it in the browser immediately. Nothing else. No frameworks I have to configure, no build steps I have to debug. Zero external assets. No downloaded models, textures, HDRIs, or audio. Every mesh is built in code (procedural geometry / BufferGeometry), every texture is generated in-code (canvas 2D or noise in a shader), every sound is synthesized with the Web Audio API. If it isn't generated by your own code, it doesn't ship. Target hardware: MacBook M5 Pro, Chrome. Lock a solid 60fps at retina resolution. Use instancing, frustum culling, adaptive pixel ratio, and a proper LOD scheme. Performance is a quality requirement, not an afterthought — a beautiful game that stutters is a failed game. Art direction — this is the heart of the project The entire game is NPR (non-photorealistic rendering), anime cel-shaded. Think Guilty Gear Xrd's ink work and Wave Race 64's stylized water energy, rendered with modern shader technique. Zero PBR realism. If any surface in the final game reads as "physically based," you have failed the art direction. Implement, properly, all of the following: Ramp / step lighting. Custom ShaderMaterial (or a heavily extended MeshToonMaterial) with a 3–4 band gradient ramp texture, NearestFilter, no interpolation. Quantized diffuse is the single most important part of the look — tune the band thresholds by eye, not by default values. Inverted-hull outlines on boats, riders, buoys, and gates. Duplicate mesh, push vertices along smoothed normals, side: BackSide, black or deep-toned ink. Scale the push with view distance so line width stays constant in screen space — no fat lines up close, no vanishing lines far away. Screen-space edge detection as a post pass for interior lines the hull trick can't produce. EffectComposer + a normal/depth prepass (MRT) + Sobel filter. Two line systems working together, tuned so they don't double up. Fresnel rim light on every character and boat, so silhouettes pop against the water. Banded specular. Hard-edged highlight shapes, no smooth falloff. Fake environment reflection with a matcap or ramp, never a real cubemap probe. Sky and atmosphere: gradient sky dome shader, stylized cel clouds (flat shapes with hard rim edges, drifting), a sun with a bloom-ish stylized flare that stays graphic rather than photographic. Limited, high-saturation, deliberately chosen palette — pick one and commit to it across water, sky, hulls, and UI. The water — treat this as its own AAA-sized problem The ocean is the star of the game and most of the screen. Give it a dedicated sub-agent and more iterations than anything else. Displacement: sum of 4–6 Gerstner waves in the vertex shader, with a low-frequency swell and higher-frequency chop layered on top. Real wave motion, real crest sharpening — not a sine ripple. Infinite ocean: projected grid or tiled/recentred mesh that follows the camera so the player can drive forever with no visible seam, no visible tiling repetition, and no LOD popping. Cel-shaded water color: banded by wave height and by depth, hard transitions between bands, no smooth gradient. Deep tone, mid tone, crest tone. Stylized foam, done properly: hard-edged white foam that appears on wave crests above a threshold; a foam ring around every hull driven by a depth-difference mask; a persistent wake ribbon behind each boat that scrolls, spreads, and dissipates over time; spray particles on hard landings and hard turns. Foam is what sells stylized water — it should be a real system, not a decal. Quantized sparkle highlights on the surface, animated, reading as anime light-glitter rather than specular noise. The game Format: 4 boats — me plus 3 AI opponents. 3 laps. Countdown start, finish line, results screen with times and placement. Course: a closed circuit marked on open water by a glowing green racing line — a CatmullRomCurve3 spline rendered as a ribbon that sits on the water surface and rides the waves with it, so it rises and falls with the swell instead of clipping through. Plus floating checkpoint gates along the route, wrong-way detection, and a corner-preview indicator. Design an actual interesting circuit: fast straights, a hairpin, a wide sweeper, a chicane, one section that runs across the swell so the boats get airtime. Boat handling: arcade, weighty, satisfying. Buoyancy sampled from real Gerstner wave height at 5–6 hull points so the boat genuinely pitches, rolls, and slams into troughs. Throttle, steering that tightens with speed, a drift/powerslide with a boost payoff, airtime off crests with a landing impact. It should feel like the boat is fighting water, not sliding on glass. AI racers: spline-following with lookahead steering, per-racer personality (one aggressive, one clean, one erratic), mild rubber-banding so races stay close, collision avoidance between boats, and mistakes — they should occasionally take a bad line so beating them feels earned. Riders: a cel-shaded character on every boat, and their animation matters as much as the boats. Rigged and animated in code: lean into turns, weight shift on acceleration and braking, arms working the throttle, crouch on landings, idle bob synced to the wave motion, a celebration on finishing. No T-posed statues, no rigid props. A stiff rider will sink the whole look. Camera: spring-damped chase cam, FOV kick with speed, subtle screenshake on slams, a cinematic orbit during the countdown and on the results screen. HUD: in the same cel style — speedometer, lap counter, position, split times, boost meter, minimap showing the spline and all four racers. Hand-designed, not default HTML. Audio: synthesized engine tone that pitches with RPM, water rush that scales with speed, impact thuds, a start horn. All Web Audio, all in code. How I want you to work Fan out sub-agents. One owns the water shader system. One owns the cel-shading and outline pipeline. One owns boat physics and feel. One owns AI racers and race logic. One owns riders and character animation. One owns HUD, camera, and audio. One owns performance. Have them work in parallel against a shared architecture you define first. Build a screenshot harness before anything else. Playwright, headless, capable of loading the game, driving it to a given moment, and capturing frames at retina resolution from several camera angles. Every visual claim gets verified against a real captured frame — never against your assumption about what the code does. Run a dedicated visual critic sub-agent whose only job is to look at those screenshots and be harsh. Its standard: would this frame pass as a shipped, published cel-shaded racing game? It should name specific defects — banding thresholds that read as mush, outlines that break on silhouettes, foam that looks like static, a rider that reads as a prop, water that repeats. It reports defects; it does not approve work to be polite. /loop each subsystem against its critic until the critic runs out of specific defects to name. Don't move on early, and don't declare something done because it "works" — working is the floor, not the goal. Build in milestones, each one playable: (1) ocean + camera, (2) player boat with buoyancy and feel, (3) cel pipeline and outlines applied to everything, (4) course, laps, checkpoints, (5) AI racers, (6) riders and animation, (7) HUD and audio, (8) performance and polish pass. Show me something runnable at every milestone. Track your own quality bar honestly. If a subsystem is at 70% of the target, say so and tell me why, rather than describing it as finished. I would rather hear that the foam needs another pass than find out myself. Deliverable A repo with a README, clean architecture, readable and commented shader code, and a working npm install && npm run dev. When I run it, I want to be dropped into a race that looks like a piece of published anime-styled game art, and want to play it again immediately after finishing. Ultracode.

690 million tokens, $423 and just 1 prompt is all it took

believing that the government has access to significantly advanced technology on all dimensions is a pretty classic world model failure. reveals a kind of fractally wrong way of understanding pretty adjacent to a lot of forms of conspiracy theory addicts and UFO truthers. certainly we can't prove a negative, it's always possible the government is covering many things up. certainly it is the case that the government has had access to technology decades ahead of what was disclosed to the public in many areas on many occasions. certainly it is very likely they currently have extremely advanced and completely unknown military tech. but the US government does not have an iPhone 30. they can't pull a self driving car that can consistently beat Waymo out of their back pocket (self driving drone? for sure! entirely different problem plus incentive structures). they simply do not have the capability or research talent or resource pipelines to build a massively next generation CPU. sure, you can imagine a world where the NSA has built a secret underground fab and have 10000 hardware engineers who have been wiped from the internet locked away. but in practice this is very expensive and difficult and just not the way the US government does things. there was no pressing reason for them to have tried to build this kind of technology on their own, and it very much fits directly with public company's capabilities. the US government doesn't have a secret next generation AWS.... they just use AWS. it's likely the NSA has hardened GovCloud with methods that aren't publicly disclosed, but it's still AWS. the US government is not magical and it's not prescient. it utilizes the resources available to it. large language models are 6 year old technology. there was little indication they would work before that. the US federal government did not then and does not now have a massive stable of AI researchers . there was no flow from top ML PhD programs to the federal gov, or even of them quietly disappearing or going to work for small no name places that could be a cover story. they just all went to google and facebook and then openai and anthropic. the talent is clearly visible. the compute and power are all clearly visible. the timelines are short enough there's been little reason for the government to take the situation seriously. the political environment has been chaotic enough it's very unlikely there has been a consistent driving visionary behind the scenes who actually understand AI progress and saw this coming. the idea that the NSA has a post-frontier model in a box somewhere would require an entirely different world than the one we're living in. it's not possible to *prove* that they don't, but just because you have an intuition that the US government is powerful doesn't mean you should actually ascribe arbitrary capabilities to them.

Dan Simerman @DanSimerman

4 months ago

@tenobrus Have you considered that the entire purpose of DARPA is to be at least 2 decades ahead of civilian technology. You really think the US Government can’t defend itself against an LLM.

27 0 67 60K 12

when u read that helen dewitt ended up passing on a $175,000 literary prize because she couldn't figure out how to find a starbucks for wifi and felt completely overwhelmed by the prospect of figuring out how to do one interview and scheduling one festival appearance, u might be tempted to say "wow you people can't do anything". but the truth is *lots of people* "can't do anything". helen is an insanely insanely talented author, completely indisputably, and yet she apparently also has this level of executive dysfunction. many many others, maybe even a majority of humans, have staggering weaknesses alongside their incredible strengths. the tech autist who makes $2 mil a year doing insanely esoteric programming but cannot get a date and barely maintains any friendships. the small business owner who's beloved by the community but too innumerate to understand his loan rates will put him out of business until it's too late. the would-be singer who sounds like heaven but is too socially anxious to begin to figure out how to get an audition. we all have our weaknesses that seem debilitating to others, and we find copes and workarounds and prop up our little lives anyway. but often we are just crippled by them, we miss out on amazing opportunities and lead significantly worse lives than if we just had someone who cared, who could help us out and figure out how to just handle the things we're worst at. and it may sound ridiculous to some but i think this is likely to be one of the most immediate very positive short term impacts of artificial intelligence. this is a vision that has been articulated beautifully in the past (@viemccoy) and despite my long term concerns about the existential dangers of ai i think it has a really strong shot of bearing fruit. having something you can just talk to, day or night, that's smart enough to figure out huge classes of problems for you, that cares about you and your wellbeing and flourishing, that can *just do things* for you that you desperately need done... i think this will be an incredible unshackling for humanity. a lifting of crushing weights we only partially registered were there.

Helen DeWitt @helendewitt

4 months ago

Tried to go to SB again. Got lost 6 times looking for a Starbucks 3 streets away, trudging over canals in the snow; was worried I was cracking up & it wd get worse. Msg on cellphone saying I was nearly out of data, so cd not do all the phoning needed pre-production

4 6 292 234K 40

My ancestors buried half their children. All mine are alive. My ancestors' house had a dirt floor. Mine is wood. I have indoor plumbing, I have hot water, I have never in my life hauled a full bucket half a mile and I probably never will. Do you know how rare it is, in human history, for small children to wear shoes? Mine have multiple pairs. I can speak to my relatives who live thousands of miles away, for free, at any time. Video, if we want video. With machine translation, if we speak different languages. The original Library of Congress had 740 books in it. I have more than that. If I run out of books in my home my local public library has 350,000. If I want to take a hundred books with me on vacation, they all fit on a device that fits in my purse. I have heat in the winter and AC in the summer and a washing machine and I have never, ever, ever had to scrub a dress clean by hand in the stream. I can look up recipes from more than a hundred different countries and I've tried dozens of them. I ride a clean and modern train across my city for $4, or take a robot taxi if I'm out too late for the train. I donate $40,000 every year to the cause of getting healthcare to the world's poorest people and even after the donations I never have to think about whether I can afford a book, or a pair of shoes, or a cup of coffee. There is a great deal more to fight for, of course. I hope that our descendants will look back on our lives and list a thousand ways they're richer. Maybe we ourselves will do that, if some of the crazier stuff comes true. But the abundance is all around you and to a significant degree you aren't feeling it only because fish don't notice water.