Yep, and if the system isn't built to shield those doing the "verifying" then there should be a clear audit log proving exactly that.

Want to bet that log is incomplete and/or missing? I can't read the paywalled article to find out.

The AI does that without any prompting. That is the problem.

My (jokey) point was that the actual data between the samples doesn't exist in the file; you're always hearing an approximation. An infinitesimally close approximation at high enough sample rates, but still an approximation.

Your name checks out, tho. :D

Uh, those screenshots were clearly taken several seconds apart. The background characters had just... moved in the scene. That wasn't DLSS deleting entire characters.

Hey, wait until you hear about linear interpolation in animation. Did you know some of the animation frames in game models don't exist? The data between keyframes is being made up! The models are moving by themselves. Spooky.

And audio! Did you know we only sample audio at around 48Khz? It's not even a continuous analogue wave! The audio in the gaps between the samples every 0.000020833333333333 seconds is being made up! That's not even real data!

Oh, and don't get me started on pixel latency on LCD displays. Did you know that the image you see is actually different from what your graphics card puts out? Those blurred pixels in the few milliseconds between each frame don't exist! They're being made up by the LCD panel!

It's all a conspiracy, I tells ya. Big Tweening is out to get us.

Not if you look at how Musi actually worked.

I'd agree if Musi were offering a different product, and not just repackaging an existing product owned by someone else while collecting fees.

Indeed. From what I can gather from 5 mins of searching, this Musi app was essentially a YouTube wrapper. It streamed music directly from YouTube without actually showing any YouTube assets or - more importantly - ads.

Calling themselves a streaming service seems like a stretch. What they did was stream YouTube's content and place their own ads over it. Apparently they made over $100M in one year doing this.

You can argue about the actual legality of what they were doing, but it definitely breached YouTube's TOS. Obviously, so do most ad-blockers, but this wasn't on an individual level, this was enabling millions of people to avoid YouTube ad impressions and eat YouTube's bandwidth while siphoning the ad revenue from those views/listens.

Perhaps YouTube discovered there was no simple way to prevent what Musi was doing through technical means, so instead YouTube complained to Apple. Apple said "yeah, fair enough" and took the app down. Musi sued Apple and lost because Apple's TOS says "we can remove your app whenever we like for any or no reason". Musi's gravy train is permanently derailed.

I'm finding it hard to get worked up about this, tbh. Boils down to "whose flavour of greed do you prefer?"

You're completely missing the point. What's the alternative here? Get someone to produce actual SuperFX chips in 2026? Good luck with that.

This is not remotely equal to what you're talking about. This is using an off-the-shelf part to replicate the functionality of a component that is no longer manufactured. It is functionally identical to putting an actual SuperFX chip in the cartridge, only - you know - actually feasible.

It seems unthinkable that this won't be hacked to run Asahi.

Haha, unfortunate word placement there.

I am not associated with Anthropic or Claude beyond being a very recent paying user. It might sound like I'm shilling, I'm really, really not. I'm just surprised to be this excited by an LLM.

Claude is the most capable coding LLM I've ever used. I paid for it after trialling it for just a couple of hours. I am an AI - and particularly LLM - sceptic. I had used ChatGPT and Copilot and dabbled very briefly with Gemini and thought they were fine as Google replacements/augments, but they suck at coding, especially from the ground up. They seem to expect you to get started and they'll just tag along.

Claude changed my mind about coding LLMs. In my opinion, Claude with Sonnet 4.5 is pretty exceptional. Certainly not perfect, it makes mistakes, sometimes gets trapped and needs you to dig it out, but its reasoning, capabilities and the sheer speed at which it seems to ingest information is mind boggling.

I tasked Claude with writing a BSP map viewer for Quake (yes, the 1996 FPS) that runs in a browser. I am an enthusiastic novice programmer at best, and I have zero experience writing 3D code. None. I have a basic familiarity with the Quake engine, I've made some maps for it, I know the basic file formats etc. but if I had to learn how to write a BSP parser, how to handle UV mapping, textures, lightmaps, (and, frankly, relearning high school maths that I haven't used in 25 years) etc. it would take me - I'm guessing - somewhere between 8 and 12 months.

I knew that Claude used the three.js library for 3D rendering and it's already very familiar with it. I gave Claude a very basic prompt outlining what I wanted; essentially just render whatever BSP you threw at it, no game logic or anything... and Claude just went off and did it. Seeing its first iteration was genuinely amazing - it wasn't at all feature-complete but it was functional; you could fly around the textured map with keyboard and mouse controls and it had thrown in a basic UI. My jaw kinda dropped a bit.

Obviously there's lots of technical information about Quake online, not to mention the entire source code of the original game on GitHub. Over the course of 5 days - without me writing a single line of code, just several hours of prompting - it has managed to make a very capable Quake map viewer with some very nice bells and whistles. I'm still working on it as it keeps hitting the token limits with my basic 18GBP/month subscription, but what it has already generated is extremely impressive.

Currently the project features full texturing with toggle between various texture filtering levels, lightmaps, entity visualisations and filtering with tagging and bounding boxes, image controls with gamma, brightness and contrast, quake-style sky and liquid rendering including mostly Quake-accurate underwater shaders, embedded texture browser, lines and arrows showing connected entities, wireframe mode and wire-overlay mode, rendering the quake character set from an embedded base64 gif, basic collision detection, alias models, animations, dynamic lighting and a lot more I don't recall right now. Again, I have not touched the code even once. I just give it prompts and if something doesn't work I describe how it doesn't work and expect it to figure it out. All of this in a (currently) ~240KB html file (and obviously it pulls in three.js from CDN).

I should note that someone has already ported Quake in its entirety to three.js using Claude, so I'm somewhat reinventing the wheel, but the only time I ever referenced that project was pointing Claude to one single source file that handled lightmaps, as Claude was having a hard time getting the implementation right from just random Internet-sourced documentation and "intuition". The only other thing it struggled with repeatedly was conversion of Quake angles to three.js angles. Even though the formula was extremely simple it took a long time to get Claude to implement it consistently, and it had to be prompted to standardize it in a utility function instead of having multiple near-identical formulas in several places. It did get there eventually, but required a lot of handholding. When I say a lot I mean multiple rounds of prompts. It's more about the tokens being wasted than the time, honestly.

Virtually every other feature I asked for it has managed to do completely by itself, coming up with fixes when I report issues, throwing in smart logging and diagnostics when it needs more information and generally just being surprisingly - even astonishingly - competent.

Maybe I need to give Gemini another shot, but frankly I'm blown away by what Claude has managed to do. It's already earned its £18 for this month, as just watching it work has been a fascinating journey, and I've learned a lot even if it is just for a niche and fairly useless hobby project.

Believe me, I think AI - and especially LLMs - are an extremely unbalanced double-edged sword. I worry for white-collar employment in the future. I hate the RAM and HDD shortages that the bubble has produced. I hate the environmental impact of datacentres and the massive energy costs. Despite all of this, I am coming around to the fact that LLMs really do have a place, and really do provide value in some scenarios.

I look forward to trying Sonnet 4.6 and seeing how it compares to what I've already experienced.

AC is the flight instructor in the Simpsons who says "You know who else I like is that, uh, Leo Sayer."

I am not associated with Anthropic or Claude beyond being a very recent paying user.

Claude is - by far - the most capable coding LLM I've ever used. I paid for it after trialling it for just a couple of hours. I am an AI - and particularly LLM - sceptic. I had used Gemini, ChatGPT and Copilot and thought they were fine as Google replacements/augments, but they suck at coding, especially from the ground up. They seem to expect you to get started and they'll just tag along.

Claude changed my mind about coding LLMs. In my opinion, Claude is pretty exceptional. Certainly not perfect, it makes mistakes, but its reasoning, capabilities and the sheer speed at which it seems to ingest information is mind boggling.

I tasked Claude with writing a BSP map viewer for Quake (yes, the 1996 FPS) that runs in a browser. I knew that Claude used the three.js library for 3D rendering and it's already very familiar with it. I gave Claude a very basic prompt outlining what I wanted; essentially just render whatever BSP you threw at it, no game logic or anything... and Claude just went off and did it. Seeing its first iteration was genuinely amazing - it wasn't at all feature-complete but it was functional; you could fly around the textured map with keyboard and mouse controls and it had thrown in a basic UI. My jaw kinda dropped a bit.

Obviously there's lots of technical information about Quake online, not to mention the entire source code of the original game on GitHub. Over the course of 3 days - without me writing a single line of code, just lots of prompting - it has managed to make a very capable BSP viewer with some very nice bells and whistles. I'm still working on it as it keeps hitting the token limits with my basic 18GBP/month subscription, but what it has already generated is extremely impressive. Currently features full texturing with toggle between various texture filtering levels, lightmaps, entity visualisations and filtering, image controls with gamma, brightness and contrast, quake-style sky and liquid rendering, embedded texture browser, lines and arrows showing connected entities, wireframe mode and wire-overlay mode, rendering the quake character set from an embedded base64 gif, basic collision detection, and more I don't recall right now. Again, I have not touched the code once. I just give it prompts and if something doesn't work I describe how it doesn't work and expect it to figure it out. All of this in a (currently) 160KB html file (and obviously it pulls in three.js from CDN).

I should note that someone has already ported Quake in its entirety to three.js using Claude, so I'm somewhat reinventing the wheel, but the only time I ever referenced that project was pointing Claude to one single source file that handled lightmaps, as Claude was having a hard time getting the implementation right from just documentation and "intuition". Almost everything else it managed to do itself, coming up with fixes when I report issues and generally just being surprisingly - even astonishingly - competent.

The other AI companies should be worried, because right now, in my humble opinion, Claude outperforms all of them at coding tasks. Anthropic knows what it has.

Genuine question: does anyone have anything good to say about HP (not HPE) anymore?