???

Among others.

Crucial is dead. There's a finite amount of rare earth. Wars and floods can bankrupt industries, supply chains are tight.

If the market is big enough, competitors will appear. And if the margins are high enough, competitors can always price-compete down to capture market-share.

Also, "price-compete down to capture market-share"? Prices are going up because all future production capacity has been sold. It makes no sense to lower prices if you don't have the capacity to full fill those orders.

Micron stopped selling to consumers to focus on the high margin enteprise market. Might change in the future.

Rare earth metals are in the dirt around the world.

Supply and demand curves shifting, hence prices increasing (and decreasing) is an expected part of life due to the inability to see the future.

They are. The problem is, the machinery to extract and refine them, and especially to make them into chips, takes years to build. We're looking at a time horizon of almost a decade if you include planning, permits and R&D.

And given that almost everyone but the AI bros expects the AI bubble to burst rather sooner than later (given that the interweb of funding and deals more resembles the Habsburg family tree than anything healthy) and the semiconductor industry is infamous for pretty toxic supply/demand boom-bust cycles, they are all preferring to err on the side of caution - particularly as we're not talking about single billion dollar amounts any more. TSMC Arizona is projected to cost 165 billion dollars [1] - other than the US government and cash-flush Apple, I don't even know anyone able, much less willing to finance such a project under the current conditions.

Apple at least can make use of TSMCs fab capacity when the AI bros go bust...

[1] https://www.tsmc.com/static/abouttsmcaz/index.htm

I guess people might be mixing up all the headlines of all the articles they did not read by this point.

As long as you are going from high quality sources, you should be fine. The issue is each transcoding step is a glorified loop-(find something we think humans can't see and delete it)

In other words: the AV1 encoder in your example works by finding 47GBs of data TO DELETE. It's simply gone, vanished. That's how lossy compression works, delete the right things and save space.

In my experience, this often deletes purposeful noise out of animation (there are often static noise / VHS like effects in animation and film to represent flashbacks, these lossy decoders think it's actually noise and just deleted it all changing the feel of some scenes).

--------

More importantly: what is your plan with the 50GB BluRays? When AV2 (or any other future codec) comes out, you'll want to work off the 50GB originals and not off the 3GB AV1 compressed copies.

IMO, just work with the 50GB originals. Back them up, play them as is.

I guess AV1 compression is useful if you have a limited bandwidth (do you stream them out of your basement, across the internet and to your phone or something? I guess AV1 is good for that) But for most people just working with the 50GB originals is the best plan

With that reasoning, lossless compression of .wav to .flac destroys >50% of data.

In actuality, you can reconstruct much of the source even with lossy compression. Hell, 320kbps mp3 (and equivalent aac, opus, etc) are indistinguishable from lossless and thus aurally transparant to humans, meaning as far as concerns us, there is no data loss.

Maybe one day we'll get to the point where video compression is powerful enough that we get transparent lossy compression at the bit rates streaming services are offering us.

> In my experience, this often deletes purposeful noise out of animation

AV1 specifically analyzes the original noise, denoises the source then adds back the noise as a synthetic mask / overlay of sorts. Noise is death for compression so this allows large gains in compression ratio.

If said noise still exists after H265.

And there's no guarantee that these noise detection algorithms are compatible with H264, H265, AV1, or future codecs H266 or AV2.

1. the prediction tools of AV1 are better than those of h265. Better angular prediction, better neighboring pixels filtering, an entirely new chroma from luma prediction tool, an intra-block copying tool, more inter prediction tools, non-square coding units.

2. If the prediction is better, the residuals will be smaller.

3. Those residuals are converted to frequency domain with better tools for AV1 as well (more options than just DCT), so that you have a better grouping of coefficients close to the DC component. (Less zeros interleaving non-zero values.)

4. Those coefficients compress better, with a better entropy coding algorithm too.

You can have exactly the same video quality for h265 and AV1 yet still have a lower bitrate for the latter and with no additional decision made to “find out what humans can’t see.” The only place in the process where you decide to throw away stuff that humans can’t see is in the quantization of the frequency transformed residuals (between step 3 and 4) and the denoising before optional film grain synthesis.

To be clear: you can of course only go down or stay equal in quality when you transcode, due to rounding errors, incompatible prediction modes etc. That’s not under discussion. I’m only arguing about the claim that AV1 is better in general because you throw away more data. That’s just not true.

The web standards are growing faster than non-profit engines can implement them. Google & Apple are bloating the web specs in what looks like regulatory capture.

If Blink/Webkit dominate for long enough, they will lock everything down with DRMs & WEI. Maybe it's time to work on lighter protocols like Gopher & Gemini that don't need 20GB of RAM to open 20 tabs ?

Each generated token takes the equivalent energy of the heat from burning ~.06 µL of gasoline per token. ~2 joules per token, including datacenter and hosting overhead. If you get up to massive million token prompts, it can get up to the 8-10 joules per token of output. Training runs around 17-20J per token.

A liter of gasoline gets you 16,800,000 tokens for normal use cases. Caching and the various scaled up efficiency hacks and improvements get you into the thousands of tokens per joule for some use cases.

For contrast, your desktop PC running idle uses around 350k joules per day. Your fridge uses 3 million joules per day.

AI is such a relatively trivial use of resources that you caring about nearly any other problem, in the entire expanse of all available problems to care about, would be a better use of your time.

AI is making resources allocated to computation and data processing much more efficient, and year over year, the relative intelligence per token generated, and the absolute energy cost per token generated, is getting far more efficient and relatively valuable.

Find something meaningful to be upset at. AI is a dumb thing to be angry at.

Compared to traditional computing it seems to me like there’s no way AI is power efficient. Especially when so many of the generated tokens are just platitudes and hallucinations.

How concerned should you be about spending 0.8 Wh? 0.8 Wh is enough to:

Stream a video for 35 seconds Watch an LED TV (no sound) for 50 seconds Upload 9 photos to social media Drive a sedan at a consistent speed for 4 feet Leave your digital clock on for 50 minutes Run a space heater for 0.7 seconds Print a fifth of a page of a physical book Spend 1 minute reading this blog post. If you’re reading this on a laptop and spend 20 minutes reading the full post, you will have used as much energy as 20 ChatGPT prompts. ChatGPT could write this blog post using less energy than you use to read it!

I found this helpful.

The energy usage of the human body is measured in kilocalories, aka Calories.

Combustion of gasoline can be approximated by conversion of its chemicals into water and carbon dioxide. You can look up energy costs and energy conversions online.

Some AI usage data is public. TDP of GPUs are also usually public.

Also, for AI specifically, depending on MoE and other sparsity tactics, caching, hardware hacks, regenerative capture at the datacenter, and a bajillion other little things, the actual number is variable. Model routing like OpenAI does further obfuscates the cost per token - a high capabilities 8B model is going to run more efficiently than a 600B model across the board, but even the enormous 2T models can generate many tokens for the equivalent energy of burning µL of gasoline.

If you pick a specific model and gpu, or Google's TPUs, or whatever software/hardware combo you like, you can get to the specifics. I chose µL of gasoline to drive the point across, tokens are incredibly cheap, energy is enormously abundant, and we use many orders of magnitude more energy on things we hardly ever think about, it just shows up in the monthly power bill.

AC and heating, computers, household appliances, lights, all that stuff uses way more energy than AI. Even if you were talking with AI every waking moment, you're not going to be able to outpace other, far more casual expenditures of energy in your life.

A wonderful metric would be average intelligence level per token generated, and then adjust the tokens/Joule with an intelligence rank normalized against a human average, contrasted against the cost per token. That'd tell you the average value per token compared to the equivalent value of a human generated token. Should probably estimate a ballpark for human cognitive efficiency, estimate token/Joule of metabolism for contrast.

Doing something similar for image or music generation would give you a way of valuing the relative capabilities of different models, and a baseline for ranking human content against generations. A well constructed meme clip by a skilled creator, an AI song vs a professional musician, an essay or article vs a human journalist, and so on. You could track the value over context length, length of output, length of video/audio media, size of image, and so on.

Suno and nano banana and Veo and Sora all far exceed the average person's abilities to produce images and videos, and their value even exceeds that of skilled humans in certain cases, like the viral cat playing instrument on the porch clips, or ghiblification, or bigfoot vlogs, or the AI country song that hit the charts. The value contrasted with the cost shows why people want it, and some scale of quality gives us an overall ranking with slop at the bottom up to major Hollywood productions and art at the Louvre and Beethoven and Shakespeare up top.

Anyway, even without trying to nail down the relative value of any given token or generation, the costs are trivial. Don't get me wrong, you don't want to usurp all a small town's potable water and available power infrastructure for a massive datacenter and then tell the residents to pound sand. There are real issues with making sure massive corporations don't trample individuals and small communities. Local problems exist, but at the global scale, AI is providing a tremendous ROI.

AI doombait generally trots out the local issues and projects them up to a global scale, without checking the math or the claims in a rigorous way, and you end up with lots of outrage and no context or nuance. The reality is that while issues at scale do exist, they're not the issues that get clicks, and the issues with individual use are many orders of magnitude less important than almost anything else any individual can put their time and energy towards fixing.

(For example, simplistically there's 86400s/day, so you are saying that my desktop PC idles at 350/86.4=4W, which seems way off even for most laptops, which idle at 6-10W)

The server shares resources!

- Gemini 3.0 Pro : 84.8

- DeepSeek 3.2 : 83.6

- GPT-5.1 : 69.2

- Claude Opus 4.5 : 67.4

- Kimi-K2 (1.2T) : 42.0

- Mistral Large 3 (675B) : 41.9

- Deepseek-3.1 (670B) : 39.7

The 14B 8B & 3B models are SOTA though, and do not have chinese censorship like Qwen3.

Maybe an architectural leap?

Will synthetic data and documentation RAG really be enough? Or will we be stuck at 2022 debugging knowledge forever?

Gmail, Notion, Facebook, are painfully slow on my high-end laptop with gigabit ethernet. Something is wrong in our modern engineering culture.

What is surprising is that a few years ago, these apps weren’t so terrible on this exact hardware.

I’m convinced that there’s an enormous amount of bloat right at the application framework level.

I finally caved and bought a new M series Mac and the apps are much snappier. But this is simply because the hardware is wicked fast and not because the software got any better.

I really wish consumer apps cared less about user retention and focused more on user empowerment.

Technically, yes. But for many large tech companies it would require a large organisational mindset shift to go from more features is more promotions is more money to good, stable product with well maintained codebase is better and THAT would require a dramatic shift away from line must go up to something more sustainable and less investor/stock obsessed.

The others, probably, VCs are incentivized to fund the people who allocate the most resources towards growth and marketing, as long as the app isn't actively on fire investors will actively push you away from allocating resources to make your tech good.