I do think this is a pretty generous estimate he ends up at. I don't think emissivity is perfectly well dealt with. I don't know if his earth reflection emissivity includes the high energy thermal transfer from particles that do exist on the near vacuum (exosphere be hit, ya'll) or if that matters. But his figures seem ballpark correct & my concerns are at best marginal.

My guess is that there's a bit of a circular issue, where the team doesn't particularly see lots of good/useful feedback when they do try hard to explain the work, and drive-by folks like me and you don't have much to see or grasp at when we do take a pass through the repos to try to see what's going on. They do have regularly scheduled meetings though, which are open attendance, if you want to try to ask in person for more help: I don't know how that would go, but I doubt you've tried.

It has been nice to have LLMs for the past year, which have helped summarize what's been going on in various wasm repositories.

That said, I agree with you on two fronts. Yes: dynamic hotload and linking is very high on my want list. Trying to get to a place where we have lightweight isolates like things that can use existing loaded libraries has been the northstar, and one I'd love to see start to become real.

Second, it is really hard to figure out what is and isn't possible, to assess and try the tools to see how far you can get. There's also a lot of old documentation floating around, so if you for example go use rust there's a 33% chance you'll happen upon material that points you in the wrong direction.

The big obvious central smoking gun that you'll get to in computer science 200 level classes is Amdahl's Law, which states:

> the overall performance improvement gained by optimizing a single part of a system is limited by the fraction of time that the improved part is actually used

You queue up some work for an agent. The LLM is going to do a bunch of work over time, and spend 20 minutes crunching on a task. Let's generously say it takes your PC 2 minute of it's CPU time for it to do the tool calls, to run the build, to run tests. If we expand this to 10 minutes to run it on a phone, that's indeed starting to be a big enough difference to notice. But in 99.9999% of cases, I don't think the harness consumes that much CPU and I don't think the growth factor is 5x to move to phone, and even if it did, it's still only an increase from 22 to 30 minutes: it's an async job either way, and the time budget is not dominated by the phone or PC running the harness.

Ideally yes, there's some intelligence to see: oh, we are about do to a build. Send the build to the build server, that's a 384 core 1U with terabytes of memory bandwidth and let it do that. But most work is not like running builds and tests. The harness doesn't need that. We need some small local computers cheap that we can have lots of running.

Model performance might radically improve in time, and that might change the Amdahl's Law calculations here. If you're paying for Turbo or Plaid or whatever, yeah, you maybe have the money to spend on a better harness too. I'd say that ideally these workloads become live migrate-able, that we can CRIU checkpoint/restore them across systems, ideally, anytime, so that we can give performance people performance when it actually counts, like the build concern above, when the agent is fast. LLM's built for speed like LFM2.5-8B-A1B (DiffuseGemini feels unlikely as it's fast, but low concurrency, but perhaps?), double the speed of many models, so that 20 minutes could become significantly less. But right now it feels like we need a lot of cheap not-performance critical harnesses that can sit around running, and that performance for them is not critical. https://www.liquid.ai/blog/lfm2-5-8b-a1b

If folks aren't aware, I also suggest taking a peak at Google Ax and Google Scion, two agent runtimes designed for scaling out, that are both kind of neat. https://github.com/google/ax https://github.com/googlecloudplatform/scion https://news.ycombinator.com/item?id=47675213

NVMe reads were faster! (Some interesting potential wins there, assuming you can get data from NVMe onto the core without going through main memory, a feature available since Sandy Bridge-EP (2011), in the form of Data Direct IO aka DDIO). I crack jokes about "PCIe speed ahead", but that's seemingly real here (at huge cost to latency, which CXL promises to remedy).

There is a non-zero chance the main cores cannot saturate what the memory controller can do, that the AI cores have some reserved bandwidth to themselves. I doubt it's going to double the memory bna

One absolute ecosystem gem from this article that I didn't know before: the fact that Orange PI 6 uses CrosEC, the embedded controller for Chromebooks (RIP i guess?). I wonder if this is the newer Zephyr Iot (awesome, also underlies Framework's new embedded controllers) or the older legacy version of CrosEC. Not spoken of flatteringly in this implementation, but it's super notable to me the borrowing of firmware from this place I didn't expect it! But there's good upstream kernel support so makes sense! https://chromium.googlesource.com/chromiumos/platform/ec/+/H...

One architectural nit I need to dig into that's interesting: the shared AI processors on the AI cores appear to have shared AI units. This reminds me a lot of AMD Bulldozer (2011), which had semi-independent CPUs but shared FPU. It was an interesting chip (still haven't disposed of my old FX-8320 server), but not well loved.

Really appreciate the dive into the matrix cores. That's going to take more time for me to look at, but: thanks. I notice the architecture diagram says all cores have AI instructions, not just the A100's. Presumably it's the same instruction set/features?

The memory bandwidth situation here feels so off. We've lived in a world where it's a battle for cores, where how many cores one could ship made chip empires rise and fall. Today, the memory bandwidth wars are on, and supplies are scarce. This looks like a fascinating board with amazing capabilities, but wow, that lack of memory bandwidth here is most surprising.

Apple just shipped iOS 27, which has support for 2019's iPhone 11. So we are around 7 years there. It's probably fine for many people's use!

For a task like openclaw or hermes, or even something more aggressively graphical & GUI, it's not hard to imagine an 8 year old phone doing fine.

I got downvoted to -4 for my rather strong worded praise of this virtue, one that the media & politicians seem to directly hate & be opposed to. There's a great Battlemesh presentation on FCC rule changes in 2014 that made it illegal to ship devices that users could modify, and since then OpenWRT has been in deep woes, has been having incredibly hard time supporting routers, for one example, with the Table of Hardware for support barely growing. https://www.battlemesh.org/BattleMeshV8/Agenda?action=Attach...

I'm pretty sad to see so many people so quietly hating, blasting a piece that is as close to my heart on the right to hack as it gets. No one even replied, to something I obviously care so much about, just downvoted me out. It is a very over the top piece, but does it really deserve to be un-seen like this? I hope not. I hope it has some food for thought, even if you don't agree. https://news.ycombinator.com/item?id=48486657

Watts just kept going with his universe. It was and is so good. Such an incredible reflection of the world at the time of writing, and I've found it's lost so little of it's capturance. That it gets so many of the plights of the over-civilized world, and the perils lurking in the economic and governmental an attention systems of the planet. From the old site (https://www.rifters.com/attic.htm) to the new site (https://www.rifters.com/), Watts just really, across mediums, wanted to get his world out, to show it's timelines. Incredible.

Starfish is where it all started, and I remember it as both a slow burn, but also so hard core, so real. In a world both so our own but so far away, so separated (insert follow up deep joke here), but still within the world, still immersed (pun!) in the Earth of the story. Maelstrom, the second book, is also incredible, in very different ways. Watts reflected on Maelstrom 18 months ago, and it captures some of the amazing titular sceneage, of an overrun net, a howling wasteland from accelerated technological adversarialism. Incredible book. He goes to talk more to his own background, biology, but upon re-reading it, I think of LLMs, of the GPU milleniums burned recently, doing not that far askance competitive training, forcing our own gradient descents in ever increasing numbers upon the world. Thanks Peter; your visions are cherished. https://www.rifters.com/crawl/?p=11220

But man, to see this story show up around the world has been really such a terrormongering frenzy of Fear Uncertainty and Doubt. Oh no, the users: they are uploading their own firware! What if they do bad things?! Everyone is at risk!

This is such toxic terrible dreck. Even ArsTechnica, who I've loved for decades, was in full on "be terrfieid, be afraid, hide your children, no one is safe" mode over this: https://arstechnica.com/security/2026/06/highly-reviewed-spe...

And they're not even wrong here. Yes. It's some kind of a risk. The device looks like a keyboard, and it could be programmed to type. It could perhaps possibly script a way to open a terminal and exfiltrate some data, before your very eyes (but if you tab away you're ok! it's just a keyboard!)

There are so many forces (and definitely among them much of the press) who seems desperate to build such a sterile, closed world, that drive such a conservative clutching for certainty against any chance or possibility for good fear before them. We've seen similar pants wetting over wireless. The FCC in 2014 changing Part 15C rules for U-NII devices was exactly this sort of "someone might do something bad possibly" rule setting, that demanded that device makers lock down their devices:

> All U-NII devices must contain security features to protect against modication of software by unauthorized parties.

> Manufacturers must implement security features in any digitally modulated devices capable of operating in any of the U-NII bands, so that third parties are not able to reprogram the device to operate outside the parameters for which the device was certied. The software must prevent the user from operating the transmitter with operating frequencies, output power, modulation types or other radio frequency parameters outside those that were approved for the device. Manufacturers may use means including, but not limited to the use of a private network that allows only authenticated users to download software, electronic signatures in software or coding in hardware that is decoded by software to verify that new software can be legally loaded into a device to meet these requirements and must describe the methods in their application for equipment authorization.

Via the excellent Battlemesh conference, https://www.battlemesh.org/BattleMeshV8/Agenda?action=Attach...

This all is such an anti-human infernal hell. A world where we can not see our devices, much less be allowed to touch or manipulate the world around us. We are being robbed of our god given right to explore, understand, experiment. This is our heavenly purpose, our god given nature, our mission to understand and shape the world around us. Yet there are these confounding legalistic and media denials of the Enlightenment project, that are against understanding the world, against humankind gaining our footing. The government is outlawing Homo Habilus, is using the law to send us back not to a pre-Great Deal (US) time a hundred years ago, nor pre-Reconstruction (US again) two hundred years ago, but to pre Homo Habilis, pre man the skillful, 2 million years ago. Anti-circumvention anti-access laws are an affront to god, an affront to our deepest spiritual nature, an affront to our species.

Rasmus is doing amazing work. The circumstance of them being able to see what was happening, being able to observe the world around them, capturing & seeing what he had there, and then doing some work to modify & change things: that is glory. That is divine. This is virtue. That is why our species was created, and why it exists, our making and our ongoing purpose, and is what has made our species better at every turn. That Conde Naste or the FCC doesn't like it, and wages war against our species is infernal. Shame on them. Shame on the terrormongers, the fear-makers of this world, who spin human access to the world about us as bad, as scary, as something to be stopped and shut down: that fear is what we have to fear, that fear is a little death for our species. That fear diminishes not protects, that fear is against god.

The god or gods made us all hackers, and that has been the best thing going for our species for millions of years: to try to unwind this is spiritual/religious treason. The terrormongering needs to be shown the door.

I confess I was quite frustrated at first when browsers all said no to wasi / wasm components. But honestly, it was the right call. It's taken so long to make wasm components happen, to get them far enough along to start really consider implementing. I can accept that as just the reality of what it takes for a small team to do such amazing work. I am so thankful for the folks who have kept this going, kept advancing.

But it's time now. 0.3 delivers an incredibly comprehensive & gorgeous suite of capabilities that offer a winning combination of characteristics (fast, lightweight, sandboxable, runtime composeable components) that is ideal for the web. I hope browsers can help get us set up for 1.0, help steer us forwards towards that spec, and I hope they're moving quickly towards being ready to implement!

(Anthropic has said they want to get Fable back on to subscription plans at some point.)

I would not expect them to dump cheap ram. That is a false hope. The world needs volume, massively more volume, and it feels like everyone else is going to take a sizable fraction of a decade to even start responding. Maybe perhaps possibly CXMT can scale fast, but they have many multiples to grow before they are more than a drop in the bucket.

It's also unclear when if they too will want to start stacking 12 then 16 then 24 rams atop each other, to sell chips that cost what multiples of what GPUs used to.

If you break it, then what do you pay? But what if you fix it after? Oh sorry, that device has been deactivated forever, of course of course of course.

Nightmare hell mode from such an evil evil evil murderous inflaming (literally, last night) chainsaw wielding government (DOGE) & person murdering (USAID) man. You will own nothing intensifies.

They also give their thanks to the Internet Archive for doing so much to preserve the spark, for keeping the lightcone alive for so so long! Also they say to Get The Fuck Off Twitter what the hell are you fools still doing there on that lightcone-less dark site?

The video demo of the svg sword is an interesting example of what is so interesting about diffusion models: it's not just putting one token after another to make edits to a file. It's skipping around, it's re-editing previous lines. I feel like forcing it to write too calls is maybe not its best nature.

I feel like perhaps instead of a monolithic edit file tool call, perhaps the diffusion model would be better suited to posting a change stream, a series of edit ops, across multiple files.

Wow, get fucked!