bjackman

My router is slightly loadbearing for my homelab so probably much more often than average!

I recently upgraded from a WiFi 5 router to a WiFi 7 one, and my bandwidth increase was a grand 0% (that's ok though I was actually just sick of my old router's shitty UI!)

There is about 10m and a brick wall between my AP and my desk so yeah, a fancy new MCS that can theoretically deliver 25Gbps was just never gonna be useful to me.

Also I've worked on (other non-WiFi) 802.11 products where our competitive edge was all about interference motivation. Our competitors at trade shows would see the demos and say "that's fake, you guys are cheating". So I have the feeling that in general there might be alpha in more resilient wireless links.

> LLMs are also terrible at these new standards.

This seems like a common enough issue that it would be worthwhile publishing a "Markdown library" of "stuff HTML can do" that you can "call" from your AGENTS.md.

Edit: lol, a sibling commenter shows I am not the first to have this idea! https://developer.chrome.com/docs/modern-web-guidance/get-st...

(They made it a skill, which brings me to another thing I've had on my mind lately: that's annoying if you know when you start the harness that you will want it to read the skill. It's a roundtrip through the model to activate it. But maybe there's a way to fix that?)

(Edit2: Claude says if you activate the skill via /skill-name in the initial prompt, you get what I want)

That's exactly what I am saying. I don't care if it's AI, I care if it's good. But unfortunately because AI makes creating software very cheap, the old signals to tell you if software is good, no longer work.

If you click the link at the top you'll find it goes to an article. The article will often expand upon the text in the link!!!

In this case:

"While we would expect children's lungs to grow year on year, our results indicate that the London children's lung growth had accelerated during the four years to 'catch up' with the control group in Luton to reach similar levels of lung capacity," researchers say.

"I was absolutely stunned when I first saw the results," Prof Chris Griffiths, a senior author on the study, at Queen Mary University of London, told the BBC.

"The speed of catch up in lung capacity in the London group was surprising and impressive."

Nothing wrong with the AI website but it suggests the code is AI too.

Nothing wrong with AI code either but it suggests the project may have been built in 2 weeks, has no community and low likelihood of ongoing maintenance.

So now overall instead of a flashy website I think the best signal of quality is a boring GitHub README and several months of commit activity, lots of merged PRs.

Well this is why, as an enthusiastic AI user, even I usually want to disable intrusive AI!

If it's popping up in some random product I don't pay for I can only assume it's such a shitty model that it's probably not worth talking to.

(Exception: my bank has replaced its shitty phone tree with an AI chatbot. This is faster and the model has a "transfer to a human" tool. It's perfectly capable of following its playbook, determining that I need a human, and putting me in touch with one. So actually I started with "now I've gotta debase myself by talking to a fucking 32B Chinese LLM" but ended with "oh that was great actually").

I deployed Forgejo on my homelab yesterday. Now GitHub is just a free backup service.

I guess they're quite happy with that TBF, it doesn't seem like they're that interested in being a git forge anyway.

Don't know actually, it might be I they are just putting cloud-init in, installing their VM agents, and maybe like tweaking the bootloader setup and kernel cmdline or something. But equally, they might be compiling it themselves so they don't have to trust distros' build farms so much.

But anyway I don't think it matters. The hard part here is having a distro image supply chain, not compiling stuff. If they do need to start compiling loads of stuff they didn't before... They just need to add CPUs.

Yeah my x86 comment wasn't meant to suggest RISCV can easily beat x86 it was pointing out that "being a good arch" is very low on the list of priorities for making the arch successful.

Exactly as you say, x86 is successful coz there are really good x86 CPUs and a SW ecosystem. Not because it's a good arch. Similarly I think, if it wins, RISCV will win because there's enough demand for a nonproprietary ISA, and this trumps basically everything else.

Another way to say this: I've seen internal slide decks about ISA strategy that target C-suites at megacorps. Not once do those slide decks ever go into technical dis/advantages, coz the people steering the 9-figure investments don't give a shit. If you piped up in those meetings and said "RISCV sucks though!!!" you would just be ignored and not invited back.

It's not really any different from startups. Tech stacks matter but there's a huge number of successful companies out there running total garbage and who knows how many perfectly engineered products go out of business every day coz the thing they perfectly engineered was not useful.

> RiscV is going to win despite its limitations, simply because they can be worked around freely, and the cost per core is zero.

This feels right to me. I've worked on embedded systems and OS my whole career and it just doesn't really feel like riscv wins or loses on technical merit to me.

Boots-on-the-ground engineers might grumble but the decisions that matter here just aren't made by them. Otherwise why would we still be using x86 so much?

I've worked on platforms that had barely-functional CPUs, we had to regularly be in touch with the engs that implemented them and they'd regularly go "oh yeah you've found a bug in our memory subsystem". This was kinda annoying for us but those CPUs were cheap as fuck and that opened up business models that weren't otherwise available. This is always gonna be more important than "ah but the CPUs are suboptimal!!"

> too much fragmentation to make binary distribution feasible

Sorry I haven't actually read the original article so this is a bit of a driveby.

But, I'm skeptical of this? Surely whichever cloud platform first offers RISC-V compute, the next cloud platforms will ship a CPU that's compatible. Probably they'll be RFTing the same CPU vendors and they won't need to coordinate explicitly for this to happen. Then everyone else building RISC-V servers will essentially be forced to align on the "AWS variant" or whatever.

So yes you'll need a -cloud build of the distro you use but most people are already doing that and the cloud platforms are already providing the infra for distros to ship it (I assume they are also contributing to the -cloud distro builds directly).

RISC-V laptops and phones I could see this being an issue but for server compute it feels like there's gonna be a Schelling point.

(Actually, for phones can't Google just fix this by fiat?)

"Lagging" is putting it rather lightly. GDM is no longer a frontier lab.

FWIW neither is xAI, there is no "big 3". xAI has had momentary peaks (I think they are having one right now) but they have never been able to claim to consistently push the frontier in any particular direction. You can also infer they aren't a frontier lab from the fact that they sell their compute.

It is not a competitor to those it competes with Sonnet.

Google's Opus competitor is 3.1 Pro Preview which is essentially obsolete (competed with Opus 4.6). They do not have a Fable/Sol competitor.

> With that difference in training speed, it should be impossible for Chinese models to close the gap that easily.

They have a big advantage in that they can directly distill from frontier models.

I think ROCm is just a total second class citizen in the space TBH.

It's a shame coz it's not even really what we want, we would obviously all be better served if we could use Vulkan or something. But I guess it's inevitable that a generic framework lags behind here.

If I was AMD I'd hire a whole ecosystem team to sit next to the ROCm people and just support big users like llama.cpp to work better on their HW, e.g. giving OSS maintainers access to their board farms. Maybe they have already done that, in which case I guess I should say I'd double the size of that team.

It's a unit of mass used in the firearms world

I believe the fully baked-in nature is pretty important for the perf they get. With a cartridge port you now have a bus between the weights and the compute and the weights and that bus can become a bottleneck.

So I think we're looking at a spectrum here:

- Fully fixed function - i.e. Taalas

- Fixed function transformer unit (or whatever other AI architecture) with a "cartridge" for weights etc - i.e. Etched

- Flexible TPU/GPU type stuff

I think the middle of the spectrum is a bit of a dead space ATM because new models have recently been coming with significant updates to the architecture (like MoE, MTP) so by the time you have new weights you wanna load, you also want to replace the compute too. So really you probably either say "I can tolerate an old model, but I want it fast as FUCK" and go for Taalas-style, or you say "I want a near-frontier model" and you have to use flexible compute anyway.

But, caveat: this comment seems to be making me sound more knowledgeable than I actually am. Take this with a grain of salt.

For anyone who wants to learn more about this claim, I thought this video was cool: https://youtu.be/M9Hoy9Kn1u0?is=1O5AJPEziihH1MSJ

Dwarkesh recently pointed out [0] that these guys are almost forced to spend most of their compute on training instead of inference. This is because they need to maintain the appearance (which may also be the truth) that future models will make current models obsolete and be much more valuable.

Completely fixed-function HW can't be used for training, it's inherently a statement that "this model is Good Enough and we are now gonna start just extracting its value instead of extending it". So yeah it's an inference moat but it's not a growth moat.

Makes perfect sense for a company trying to get into the compute business, not companies who wanna be in the creating-ASI business.

Still, I guess/hope they have teams doing it in-house anyway. Just not something they'd wanna make a huge amount of noise about, it doesn't look good for To The Moon valuations.

[0] https://www.dwarkesh.com/p/why-compute-might-get-10x-more-ex...

No I think you are thinking of Etched/Sohu.

Taalas' approach (at least for their demo'd product) is to bake the whole thing in completely. IIUC the optimiser can even see the weights while generating RTL. It's like there's an "uint8_t weights[] = " in the source code.

I'm not satisfied with any of the text editors I've used (I've used all the big ones for a few years each).

Zed was a promising new direction I.e. VSCode minus the bullshit. But it's not getting the polish / productionisation it would need to actually be a better editor than the others. And now it doesn't look like it will.

I don't want someone "to make it simpler for developers to follow changes applied by an LLM in an ergonomic fashion". I want a text editor.

I want "VScode without the bullshit". That's what I thought they were building.

this prompted me to search and find: https://addons.mozilla.org/en-US/firefox/addon/xcancel/

ABP works for the most common of these too (e.g. popups/banners)! You have to enable it in the settings I think.

JS annoyances it won't fix are like idiotic scrolling behaviour and stuff. And for that, yeah reader mode is ideal.

Firefox has this too, it's useful. However I only find that I use it for websites that have shit CSS that makes it hard to read on mobile. For everything else AdBlock Plus works fine.

You know if you tap on the text at the top, it takes you to an article that expands upon the question. In this case you'll find that article discusses the strengths and weaknesses of the materials bottleneck!

In the Linux kernel I've had a lot of luck with:

1. "find the code that does X"

2. go read that code

3. When you hit a bit you don't care about, go back to the model and ask it for the pertinent details

4. When you hit a really confusing bit, ask the model for hypotheses about what's going on. (I always phrase it as "give me some hypotheses" not "what is going on here". I dunno if this changes the output but I think it helps me stay in a mindset of uncertainty, it's important to avoid locking in any misunderstandings. Anyway I find the models do well at this task, and when they bullshit here it has a strong smell).

Before AI, parts 1 and 3 could be insanely time consuming, sometimes it felt like a infinite breadth-first-search. And part 4 was basically: either you find a human who knows the code, or you just make a mental note and hope that later on you find something that makes you go "oh, THAT'S why they <do weird thing that should 100% have a comment>!".

So yeah even though you're still reading code with your wetware the AI makes you dramatically more powerful.

This is also extremely helpful for unpicking undocumented API contracts. E.g. you can say "the x86 implementation of this API is safe to call under a spinlock, go read the other arch versions and tell me if that's true there too".

I have been pondering this and I think it's likely a gap that will get filled sooner or later.

Right now there's just so much value in building LLM tools for experts that everyone is focusing on that. But surely at some point we'll have bespoke harnesses that exist exactly to solve this kind of thing.

I think this can start with constrained problem spaces like "you are a WordPress developer, you solve problems for people with enough expertise to know they are looking for a WordPress developer" and incrementally expand from there. Maybe I'm naive but I think you can probably get pretty far with this today just by writing loads of skills and picking the right technical preferences to encode in them.

> The model outputs are much more concise than when I try and talk to GPT-5.6 Sol about mathematics. By signalling expertise, Tao shunts the model into “talking-to-mathematicians”

Maybe, but FWIW my first thought when I skimmed Tao's session was that he probably has a personal system prompt requesting this style.

E.g. even if you get it into "talking to an expert" mode I've found AI waffling through filler like "given your background in Linux kernel engineering, I'll skip the surface level and go straight to the technical meat". You do have to explicitly tell them if you don't want this.