Sounds like any post-secondary, graduate student, or management consultant out there being there are, very often, page/word count or hours requirements. Considering the model corpora, wordiness wins out.
The chart is actually words "thought or written" so I guess they are running up the numbers even more by counting Claudes entire inner monologue, on top of what it ultimately outputs.
It's not merely cost per word, but it is even more bizarre: "cost per word thought", whatever that is. Most of these "word thoughts" from LLMs of today are just auto-completed large dumps of text.
SRAM scaling also hit a wall a while ago, so you can't really count on new processes allowing for significantly higher density in the future. That's more of a longer-term issue with the SRAM gambit that'll come into play after the DRAM shortage is over though - logic and DRAM will keep improving while SRAM probably stays more or less where it is now.
You can still scale SRAM by stacking it in 3D layers, similar to the common approach now used with NAND flash. I think HBM DRAM is also directly stacked on-die to begin with, apparently that's the best approach to scaling memory bandwidth too.
It'll be interesting to see if we get any kind of non-NAND persistent memory in the near future, that might beat some performance metrics of both DRAM and NAND flash.
NAND is built with dozens of layers on one die. HBM DRAM is a dozen-ish dies stacked and interconnected with TSVs, but only one layer of memory cells per die. AMD's X3D CPUs have a single SRAM die stacked on top of the regular CPU+SRAM, with TSVs in the L3 cache to connect to the extra SRAM. I'm not aware of anyone shipping a product that stacks multiple SRAM dies; the tech definitely exists but it may not be economically feasible for any mass-produced product.
The issue is size, SRAM is 6 transistors per bit while DRAM is 1 transistor and a capacitor. Anyone who wants density starts with DRAM. There’s never been motivation to stack.
> AMD's X3D CPUs have a single SRAM die stacked on top of the regular CPU+SRAM, with TSVs in the L3 cache to connect to the extra SRAM.
Just FYI, the latest X3D flipped the stack; the cache die is now on the bottom. This helps transfer heat from the compute die to the heatsink more effectively. In armchair silicon designer mode, one could imagine this setup also adds potential for multiple cache dies stacked, since they do interpose all the signals, why not add a second one ... but I'm sure it's not that simple, for one: AMD wants the package z-heights to be consistent between the x3d and normal chip.
I agree with your description and conclusion. Additionally the companies that can make chip stacks like HBM in volume are the HBM manufacturers. As they are bottlenecked by the packaging/stacking right now (while also furiously building new plant capacity) I can't see them diverting manufacturing to stacking a new SRAM tech.
> So, my question to anyone in the Microsoft C-suite: have you ever tried to, like, actually use, like anything that you're selling?
Satya Nadella insists that Bing365Pilot has supercharged his productivity, but determining if he's high on his own supply or lying through his teeth is an exercise for the reader.
> Copilot consumes Nadella’s life outside the office as well. He likes podcasts, but instead of listening to them, he loads transcripts into the Copilot app on his iPhone so he can chat with the voice assistant about the content of an episode in the car on his commute to Redmond. At the office, he relies on Copilot to deliver summaries of messages he receives in Outlook and Teams and toggles among at least 10 custom agents from Copilot Studio. He views them as his AI chiefs of staff, delegating meeting prep, research and other tasks to the bots. “I’m an email typist,” Nadella jokes of his job, noting that Copilot is thankfully very good at triaging his messages.
This is just him aping every other AI CEO. Every single one has to act like the agents are super-geniuses mere moments away from achieving the singularity like people can't try them out themselves and be disappointed. Some of it is "we think this will work soon so it's ok if we pretend like it's working now", but I think a lot is just needing to constantly shove hot air into the balloon before it pops.
> he loads transcripts into the Copilot app on his iPhone so he can chat with the voice assistant about the content of an episode in the car on his commute to Redmond.
What a dorky thing to do. Does the CEO have some concept he's living a life that precisely _zero_ of his customers do? Who would even think to do this?
> “I’m an email typist,” Nadella jokes of his job
Yea, I have actual work to do, perhaps you should familiarize yourself with this?
> He likes podcasts, but instead of listening to them, he loads transcripts into the Copilot app on his iPhone so he can chat with the voice assistant about the content of an episode in the car on his commute to Redmond.
I remember reading that when it first came out and all I can think is: No, he doesn't like podcasts, if you like podcasts you listen to them.
That's like saying "He loves food, but instead of eating it he feeds it to an analyzer that tells him what elements were detected in it".
I have to assume it's all BS/lies because if that's a truthful statement (about podcasts and the other things) then I really question wtf they are doing over there. None of that sounds like "the future", it sounds like hell. I cannot imagine how shitty it would be to have all my emails/messages to the CEO being filtered through an AI and getting AI slop back in return.
The podcast thing reads like something he made up that sounds cool and futuristic on the surface, but doesn't actually make any sense. Maybe what's actually happening is he's just having the LLM give him the cliff's notes of the transcript, but that isn't interesting enough so he's making up some BS about having a conversation with the AI.
Nadella has to have his own custom agents. It isn't even possible for an enterprise like MSFT to not have custom agents that are still remotely useful.
So, his experience with Copilot agents != Average Customer's experience
Look, if the CEO of a five trillion dollar company investing hundreds of billions into AI can come up with some custom agents to handle every aspect of the CEO's work, surely your average Microsoft 365-subscribing corporation can do the same.
It's just another example of the rich being wildly out of touch. Yes, Beyonce has the same 24 hours in a day that the rest of us have, but she also has enough money to pay people to do every aspect of her life that isn't bringing her joy or wealth. Yes, AI can be used to streamline workflows or help you find signal in the noise so you can focus on the important things better, but if every company has to build that themselves then no company is going to see the value of spending a bunch of extra money on something that they can only get benefit from if they spend even more money.
If the 'AI agents' that Nadella is talking about were part of Copilot then sure, okay, I could see a benefit, but when people in this thread are saying that Outlook can't even tell you who is in a meeting then it certainly explains why Nadella doesn't understand the lack of value.
If you want a cheap rooted eReader I think you're better off getting a Kobo instead, they don't officially support rooting but AFAICT they make basically no effort to prevent it.
The latest Kobos use MediaTek SoCs with locked bootloaders. The Kobo Clara BW's MT8113, for example. As far as I know, one of the early bootloaders it, BL1, refuses to execute the next bootloader (BL2) unless its signature is valid. We can get the device into a mode where BL1 waits for upload of a BL2 via USB using an exploit called Kamakiri, but in public there is neither an exploit to get BL1 to boot an arbitrary BL2, nor an authorized BL2 image to upload. See here: https://github.com/bkerler/mtkclient/issues/1332
Kobo devices have root exposed but don't let users boot their own kernels (and the kernel they ship was not compiled with kexec either).
I really don't know the reason so many devices these days don't have an unlock method. It seems predatory. Who knows where in the chain this happens... maybe it's Kobo, or maybe MediaTek won't sell you their SoCs for mass-market devices unless you lock them.
According to the github issue it seems to be a simple checksum step, not a true signature verification? If so there is no locked bootloader in any real sense.
If the real impediment is lack of demand or low-level development effort for any given device, that's in principle a solvable issue once projects like pmOS and Mobian choose to focus on some reasonably-available hackable hardware and bring it up to true daily driver state.
mtkclient does not seem to correctly interpret the usb output of the device past some part of the early boot process. Really, any of those messages formatted by mtkclient are unfaithful to the intended meaning. So yes maybe it is "just a checksum step" or maybe something else entirely. Last year I collected some UART logs on the device during bootup in a zip here:
As far as I know, yes, it's possible. No SELinux. Kernel is a branch from 4.9.something pretty far off mainline with a few proprietary binary blob modules. As far as I know the real impediment here is lack of demand.
+1 to a Kobo, they cheaper and better than Kindles, with full Calibre support (https://github.com/kovidgoyal/calibre - OSS which has been in development for ~20 years!).
The way you install additional software is literally just moving files into folders whilst its plugged into your computer. I'm sure it could handle Tailscale.
I agree with your sentiment that the Kobo is better than the Kindle from an... ethical standpoint, if you have the money for one. However, it is worth noting that Kindles will always be cheaper than Kobo devices [0] due to economies of scale and lockscreen advertisements (removable with jailbreaking). From a pure cost perspective, and assuming the user is technically-minded enough to accomplish the jailbreak, the Kindle is likely always [1] a better deal.
[0] as of today, 12/8/25, the "base model" Kindle 11th Generation is priced at $109.99 USD, and the respective Kobo Clara BW is $139.99 USD.
[1] I say "likely always" to cover my bases. To my knowledge Calibre supports Kindle, just not as well as Kobo. That said I have found that the KOreader app is more than powerful enough for my use case (reading my own epubs, using dictionaries, etc.)
The current colour kindles and kobos don't use real eink colour. It's just a bw screen with lcd colour overlay (eink kaleido)
The real colour screens are used on the remarkable (eink gallery) and they are indeed slow for full page updates though remarkable seems to have done a lot of smarts for local updates while drawing.
Where do I get DRM-free ebooks to put on a Kobo? I don't support breaking DRM. So I'm using a Kindle because it has the best access to and integration with almost any book I want.
I kinda love that buried in the koreader menu somewhere is an option that drops me at a linux shell. I have no use really for this feature, but i like it. Good for those times you absolutely have to crank out some awk on the plane or whatever. :)
Same with the Barnes and Noble Nooks. I've never rooted one, but via ADB one can install a launcher and most Android applications run. I've used four generations of Nooks to run AnkiDroid.
Just beware to check what version of Android the Nook is using before you buy, and what your app needs.
Android on an e-reader unlocks so much potential. I've owned four or five Kindles over the years but recently switched to an Onyx Boox page 7" as my main e-reader. Expensive (relative to Kindles) but runs full Android 11 and has physical page turn buttons. I use an app called BookFusion to sync my library including reading position across all platforms. Battery life isn't Kindle grade but I can get by charging once a week which is a good enough tradeoff for the convenience of being able to run Android apps.
I used to like my Kobo a lot but recently it's got some pretty severe unreliability issues, usually around reading non-Kobo epubs and PDFs. Like, if I open of those files, the device usually crashes and when it recovers after a reboot, the file disappears.
The only (tiny) issue I've had with Tailscale on Kobo has been that the tailscale daemon prevents me from using the Kobo in Mass Storage Mode while it's active, so I have to disable/quit KOReader to be able to plug it in again, which is admittedly not frequently warranted anyways.
I think the only way to get the no-Copilot version now is to already have the Copilot version and try to cancel your subscription, and only then they'll offer the "Classic" version sans Copilot as a last ditch retention effort. If users actually wanted this stuff they wouldn't need to bury the option to not pay for it.
> “Everybody should be selling or licensing their voice and their skills to these companies,” Stewart said. “Otherwise they’re just going to take it anyway.”
If they can just take it anyway, why would they ever pay for it? Make it make sense.
Doesn't the UEFI firmware map a GPU framebuffer into the main address space "for free" so you can easily poke raw pixels over the bus? Then again the UEFI FB is only single-buffered, so if you rely on that in lieu of full-fat GPU drivers then you'd probably want to layer some CPU framebuffers on top anyway.
What type of compression would change the relative scale of elements within an image? None that I'm aware of, and these platforms can't really make up new video codecs on the spot since hardware accelerated decoding is so essential for performance.
Excessive smoothing can be explained by compression, sure, but that's not the issue being raised there.
Neural compression wouldn't be like HVEC, operating on frames and pixels. Rather, these techniques can encode entire features and optical flow, which can explain the larger discrepancies. Larger fingers, slightly misplaced items, etc.
Neural compression techniques reshape the image itself.
If you've ever input an image into `gpt-image-1` and asked it to output it again, you'll notice that it's 95% similar, but entire features might move around or average out with the concept of what those items are.
Maybe such a thing could exist in the future, but I don't think the idea that YouTube is already serving a secret neural video codec to clients is very plausible. There would be much clearer signs - dramatically higher CPU usage, and tools like yt-dlp running into bizarre undocumented streams that nothing is able to play.
A new client-facing encoding scheme would break utilization of hardware encoders, which in turn slows down everyone's experience, chews through battery life, etc. They won't serve it that way - there's no support in the field for it.
It looks like they're compressing the data before it gets further processed with the traditional suite of video codecs. They're relying on the traditional codecs to serve, but running some internal first pass to further compress the data they have to store.
If they were using this compression for storage on the cache layer, it could allow more videos closer to where they serve them, but they decide the. Back to webm or whatever before sending them to the client.
I don't think that's actually what's up, but I don't think it's completely ruled out either.
That doesn't sound worth it, storage is cheap, encoding videos is expensive, caching videos in a more compact form but having to rapidly re-encode them into a different codec every single time they're requested would be ungodly expensive.
The law of entropy appears true of TikToks and Shorts. It would make sense to take advantage of this. That is to say, the content becomes so generic that it merges into one.
The resources required for putting AI <something> inline in the input (upload) or output (download) chain would likely dwarf the resources needed for the non-AI approaches.
reply