←
Arthur54,911 posts
Arthur avatar

Arthur

@knowclarified

Currently: smallbusinesschatbot.com On the side: searchthatsong.com (20k users/mo), AGI in Minecraft Do something worthy of mockery

◷ Joined July 2022▧ 6,024 Photos and videos
Arthur avatar
Arthur@knowclarified·26 Feb 2023

“I wish it need not have happened in my time," said Frodo. "So do I," said Gandalf, "and so do all who live to see such times. But that is not for them to decide. All we have to decide is what to do with the time that is given us.”

Post attachment
Arthur avatar
Arthur@knowclarified·3h

People hacking old hardware and pushing them to new limits is so interesting

I've been doing research on the Dreamcast's PVR and found a solution to introduce subsurface scattering--or at least something that looks exactly like it. Here is a real hardware capture on Sega Dreamcast. Despite being a fixed-function GPU, there is a LOT more you can do with it than you think. I've worked in visual effects for film and TV for almost 9 years now, so when I started thinking of the PowerVR chip as a supercharged compositing and layering suite (not unlike a living Photoshop), I started to get some real workhorse out of it. Every wave, colour and reflection is computed on the console: rolling swells that move the actual geometry, fine ripples catching the light, a sky gradient fading into horizon fog, the sun reflected across the water, and wave crests that shadow the water behind them. How it works: The SH4 CPU builds the water every frame. Four sine swells, each travelling in its own direction, displace the water mesh. Three shorter waves only bend the surface normals, so you get fine ripple detail without adding triangles. Slopes come straight from the wave formulas, so neighbouring heights never need recalculating. Mesh rows are spaced evenly in screen depth rather than world distance, which puts detail where you can actually see it. Far away, ripples too small for the mesh to resolve fade out instead of shimmering. A small heightfield pass then checks which crests block the low sun and shades the water behind them. All of the maths goes through SH4ZAM, a maths library built around the SH-4's vector and fast-reciprocal instructions built by @falco_girgis. I also had contributed by having Fable 5.2 Max Ultracode ("Lord SH4ZAM" as I call him) create additional SH4ZAM routines in a layered approach to find approaches to maths that Falco hasn't implemented yet: 1. shz_dce_project_vertices: a fast fused vertex routine where every row of the water's roughly 11,700 vertices goes through, with transform and screen projection in one pass. 2. shz_dce_vec3_normalize_safe: turns the view and sky directions into unit length. The PowerVR2 GPU handles the look. The whole upper sky lives in one 512×512 texture, and each vertex's bent normal decides where the water reads from it. That means the sun's reflection streak forms out of the waves themselves rather than being painted on. A Fresnel term sets how much the water reflects versus shows its own colour: the PVR multiplies the sky texture by vertex colour for the reflection, then adds its "offset" colour for the deep-water tint and the light glowing through thin crests. The PVR's built-in fog table blends far water into the exact horizon colour between 55 m and 520 m, at no cost. The sky gradient is dithered when it's generated, so 16-bit colour doesn't band. Where this could go: Ocean and lake surfaces for racing, sailing, flight or island games, and big outdoor horizons, all at almost no disc space since the scene is generated. The crest-shadow trick works for terrain too, with hills shadowing valleys from a low sun. Building reflections from bent normals is a cheap way to fake environment reflections on fixed-function hardware.... but you will need to scope down the effects. Obviously, you would not run the water as heavy as I did (turn off SH4DOW-DC or limit its range), but I am pretty pleased with this result. I pushed this to the limit for demonstration purposes to show what our little white console can do.

2:58
Arthur avatar
Arthur@knowclarified·4h

I assumed I was in the minority holding the same POV as Dean here. We must build the future where AI is an extension of human capability, not to be replaced by it. The Beginning of Infinity means problems will always be inevitable and soluble, so therefore our primary obstacle is ensuring humanity is still the center. History has not been written and machine dominance is not inevitable.

Post attachmentPost attachmentPost attachment

openai.com/index/the-silent-…

Arthur avatar
Arthur@knowclarified·Nov 18

Prediction for GTA VI: the game will be well received but will run on a very low fps for most consoles Almost certainly they’ve made the game for future hardware knowing how long it will last, and seeing so many dynamic NPC’s and graphics lifts makes me think there’s no way they can efficiently squeeze that into current console hardware

Arthur avatar
Arthur@knowclarified·Nov 17

I feel like I’ve heard nothing about GTA VI and it’s coming out in 2 days

Arthur avatar
Arthur@knowclarified·Nov 17

OpenAI has lost the sauce recently My suspicion is always that their consumer positioning has forced them to be the Walmart of foundation model companies Output will be RL’d into hieroglyphics before we know it All of Python will just be the snake emoji 🐍

Ethan avatar
⟳ Arthur retweeted
Ethan@torchcompiled·Nov 17

Possible explanation: it may be a genuinely strong model but may be extremely particular about right prompts and/or overly specialized in frontier problems. If it doesn’t generalize to common user cases and has crap UX… a tool is only as good as you can reasonably control it.

This is exactly the issues I had with Astra and Sol GPT SOL 6 was better than 6.1 and I found Astra 6 to be worse than 6.1 So I cancelled my Codex subscription and then went all in on Anthropic. It simply is just better Found it super odd Astra had saturated most benchmarks but in my own testing was demonstrably worse than Fable and even Opus 4.5 I will give both of these new 6.2 models a shot but expecting more of the same

Arthur avatar
Arthur@knowclarified·Nov 17

Huge fan overall of OpenAI and use them in my production applications. They have the best low cost reasoning models, no question. Surprising to see this all happen

Arthur avatar
Arthur@knowclarified·Nov 17

This is exactly the issues I had with Astra and Sol GPT SOL 6 was better than 6.1 and I found Astra 6 to be worse than 6.1 So I cancelled my Codex subscription and then went all in on Anthropic. It simply is just better Found it super odd Astra had saturated most benchmarks but in my own testing was demonstrably worse than Fable and even Opus 4.5 I will give both of these new 6.2 models a shot but expecting more of the same

It pains me to say something this negative, but I wouldn't be doing the honest thing if I didn't say what I actually felt. I've had mixed experiences with OpenAI models lately, but this is the first launch that is just downright negative. Both models are worse than Opus 5.5 (and massively worse than Fable 5.2) in essentially every meaningful way... they are impossible to work with, they don't follow instructions, they speak in gibberish, the list goes on. I have no idea why they were released. More time in the oven would've made a huge difference.

Arthur avatar
Arthur@knowclarified·Nov 13

Post attachment

I asked Fable 5.2 to study all of music theory and generate the best house bangers. It tries 1000s of random patterns for each instrument and chooses the optimal one and arranges it into a 72 bar track. I thought it did a really good job for not being able to hear [sound on]:

5:51
Arthur avatar
Arthur@knowclarified·Nov 13

If you really, truly love something, you’ll modulate your own desire toward it in order to get it

Arthur avatar
Arthur@knowclarified·Nov 12

Post attachment

I have a major result to share. I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China. This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model. Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves. I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems. I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used. It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20. These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs. I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo…. There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.

Post attachment