msdz

10 to 15 years from now the scale of LLM efficiency improvements is, quite literally, unpredictable.

It could be that the company valuations crash tomorrow, and (almost) only performance gains achievable on hobbyist-level hardware come to fruition from there on out.

Or it could be that in the future, we have a custom "model FPGA" à la Taalas [0] in every home, and that it turns out we can still massively boost inference efficiency due to novel discoveries like TurboQuant [1] or a somehow-improved quantization method [2] again and again ten times over.

Point is, Moore's law in this context shouldn't be applied to just hardware spec sheets alone, but more the total number of "parameters potentially improving", IMO.

[0] https://chatjimmy.ai

[1] https://research.google/blog/turboquant-redefining-ai-effici...

[2] https://prismml.com/news/bonsai-27b

Exactly. AFAIK none of the popular "open Chinese" models have published the full pre- and post-training pipeline, so the models are only partially open, if at all (plus the openly shared final weights, of course).

Yeah, I think that’s the one case where it’s actually appropriate to use that card.

Like it or not, banks are legally mandated to do KYC, and this is essentially the same, except for information or communication being “moved” instead of money.

I’m amazed that it is even legal for businesses to ask for that at all, outside of a few exceptional cases of which “I want to rent a server” most certainly isn’t one.

How about the case of “I want to rent a server to distribute illicit and or dangerous materials and then it’s your problem”?

Is engineering capability the issue, or time pressure/constraint?

Context: https://en.wikipedia.org/wiki/Uyghurs

Since 2014, the Chinese government has been accused of subjecting Uyghurs in Xinjiang to widespread persecution, including arbitrary arrest and detention, forced sterilization, and forced labor. This is denied by China.

TIL, that makes a lot of sense, and thanks for the correction.

We also further increased the sparsity of the Mixture of Experts (MoE): with the Stable LatentMoE framework, the model efficiently activates 16 out of 896 experts. Together with improvements in training methodology and data recipes, these structural advances give K3 roughly 2.5x the overall scaling efficiency of K2, converting compute into capability more effectively.

Assuming experts are uniformly distributed (I’m really not that familiar with the deep details there), that’s 2800/896*16 = 50 billion active parameters just for the active/expert part. Wild stuff, and I’m glad there’s at least some companies still publishing (and pushing, for open-weight models) total parameter count.

And: It sounds very believable that this would result in efficiency gains wrt. to compute necessary for “good”-quality inference. Does anyone know whether there currently even are any SOTA or near-SOTA models that are dense still?

LLM’s amplify what you already have: opinions, structure, frameworks.

So far, so agreeable, but…

If you have thoughts, they come out sharper and faster.

I can’t help but wonder whether constant use of “agent” harnesses will lead to an atrophy of the software engineering (or really any field) muscles.

Actual muscles need exercise to stay in shape (let alone grow), so does the brain. Can we really be sure that thoughts, opinions, taste will still come out sharper and faster after five, ten, 20 years of using these tools almost every day?

Conversely, I also am a user of LLMs (true shocker these days, I know), and am noticing a speedup in areas I was already familiar with, and a quicker introduction to new ones. The obvious benefit cannot be denied, and doing so regardless makes you look uninformed. [0]

So what’s the ideal “middle ground” in this situation? Stoically continuing to sharpen your skills on your own, but risking being left in the dust productivity-wise? Or taking an “agent first” approach and trying to learn and improve more only on the side, as more of an afterthought?

[0] Excluding people who don’t want anything to do with LLMs out of moral principle, which curiously just like the overarching topic I also both respect and understand, but on the other hand don’t do myself.

Glad to hear!

By the way, I never got a chance to thank you: The fireside you held at CDTM InnoLabs (I believe early 2024 was when I attended?) completely changed my thinking on direct air capture/DACCS and, to be frank, on carbon credits in general. Thanks!

Hi Felix,

I know the situation here in Germany kinda sucks for non-incorporated founders (or simply any website administrator trying to commercialize anything [0]), but gating imprint/Impressum behind a login wall makes it not legally compliant. It needs to be easily accessible from anywhere (that’s why most people place it in the footer), without auth or signup; and if you put it behind either Outlook or Gmail logins, you may as well just not include it at all (realistically, who’s gonna complain if you don’t include “made in Germany”).

All the best for your project, though!

[0]: Personally I’ve given up and just include my name and address on the public web in projects now, which I guess is what the federal government wanted to achieve.

First off: I’m not really well-versed in terms of UX design, so these next guesses are as good as anyone’s:

- Six boxes communicate “== six digits, != password” (try to imagine the least tech-savvy user)

- Some people might not be able to quickly hold six digits in their head at once (especially seems to be the case with older people, from my experience), so seeing “I’ve already got two/three/four” visually is potentially an efficiency boost. This also correlates (again, just from my anecdotal experience) with the population that doesn’t have a password manager handling 2FA for them automatically…

- This is probably down to preference, but IMO it also feels faster somehow than a single input field with six digits, when it works properly (with the caveat that it feels much worse/slower if it doesn’t work)

- Now that everyone does it, you’re kinda forced to adopt it as well for recognizability.

But it’s an interesting question – we can’t ever imagine entering postal ZIP codes like that and yet that’s an area where it’d actually make more sense since there’s usually some notion of “more significant” digits at the start vs. end, so you could perhaps do some cool tricks with regard to pre-filtering state, city etc. names. Whereas OTP codes are (AFAIK) essentially fully random digits with no meaningful distinction.

If you allow, I’m curious: If you felt familiar with the language just as much as with C/Zig/Odin, would you prefer Rust for a completely greenfield project that requires no C interop (or none more detailed than say providing a general ABI)?

a unique message ID in the system prompt broke the entire KV cache. Since the cache requires byte-exact matches, that changing ID forced a full re-compute on every turn, turning warm contexts into cold fills.

This is (part of) the same problem that initially lead Anthropic to ban non-Claude Code clients from using the subsidized subscription: A full to-the-second datetime stamp in the system prompts of OpenCode, and I believe Pi as well, invalidated the caches, making this a very expensive use of their compute very quickly.

They even had Anthropic employees submit PRs (or maybe just open issues, I’d have to check) to these other clients/harnesses because the cache misses were hitting them so hard.

Is the server side open-source too, as gruez brought up in the sibling comment?

Technically they can still do potentially any- and everything undetected there; and for what it’s worth, even with a closed-source client bad behavior would get detected eventually through network inspection.

Same goes for money: It enables greedy jerks to be more greedy and more of a jerky, and it enables people who e.g. voluntarily donated already to do much more in that direction, too.

It’s just inherently a difficult problem to solve, I think (just like with human individual contributors, where there famously also doesn’t exist one universal, automated process).

“Good” benchmarks to gauge development skills at the moment seem to be:

DeepSWE [0] by Datacurve

FrontierCode [1] by Cognition

And then there’s TerminalBench, which I’m certain has been saturated in post-training to no end, so I wouldn’t think of it as a gold standard anymore.

But yeah, in general, it’s not going to get easier knowing which benchmarks are actually measuring “frontier” capability, and which are just getting results inflated by way of time/token budget [2], ever again.

[0] https://deepswe.datacurve.ai

[1] https://cognition.com/blog/frontier-code

[2] https://xcancel.com/i/article/2064210146558136827

Thanks for the link (and underlying thoughts), I really hadn’t considered that.

So essentially, due to technological progress and other factors inducing price collapses (or at least cycles), you can’t start stockpiling insane amounts of finished-product semiconductor, which means you can’t scale production at current technology levels to infinity either?

Unless the raw materials have an inherent limit on mining/production due to the amount present on the planet, why should or would companies not ramp up to eventually meet demand?

Edit: Okay, this doesn’t mean that that’s actually possible in the short-term, so I think you’re right. But that means as the silver lining, in the medium term horizon there’ll be enough supply again? :’)

Genuinely thank you for linking to him and his case.

I live in Germany and am -principally- a massive advocate for and proponent of the free (or liberal) democratic basic order ("FDGO" [0]) we have had here for the last decades, and apart from Chat Control, I’m usually very highly pro-EU, too.

But reading this has genuinely left me in a bit of a shock now, and created some (for lack of a better word) FUD I haven’t felt before with regard to these two, eh, institutions governing us.

[0] https://en.wikipedia.org/wiki/Liberal_democratic_basic_order

tsc isn't fast but it's fast enough to get the work done, the annoyance is there but it's objectively minimal.

For our small codebases, maybe. The tsc-in-Go announcement had VS Code compilation go from >120 seconds to ten.

While that's true and pretty funny, it does make for a meaningful distinction IMO: Just like you wouldn't look at the "all Ds" student, but not the "straight A" student in order to get an impression of students' capabilities, you wouldn't ignore a newer, potentially SOTA-pushing model to then come out and say something along the lines of "all models suck"… right?

Has this been just pure lack of funding and infra?

For the most part, yes.

France and Germany are the two biggest EU economies. France has well, Mistral, and we here have a government-funded VC entity that is way too proud [1] to be able to offer a whopping… €125 million (<$150 M USD) for helping European researchers achieve new SOTA in sovereign models. And that sum is not even going to a single challenge winner, it'll be split up among multiple recipients. Don't get me wrong, this is a cool first step, or rather, would have been one about three to four years ago.

It's a pity, really.

[1] (in German) https://www.sprind.org/worte/magazin/verkuendung-next-fronti...

However, due to things such as batching and even different kinds of floating point imprecisions for different algorithm implementations, the probability distribution itself often differs run-by-run, so what you sample from it also differs.

Exactly. While I’m assuming this won’t be news for most here, for those that are still new and/or curious about some more explanation on e.g. the floating-point imprecisions, see this nice article: https://thinkingmachines.ai/blog/defeating-nondeterminism-in...

Maybe two things can be true at the same time?

https://swelljoe.com/post/open-model-censorship/

although quality seems very bad

The weights they “etched” into the FPGA card that’s used for the ChatJimmy demo are that of a Llama 3-something 8b model.

The actually impressive and novel thing is that Taalas’ve managed to automate that process (clearly – nobody transforms 8 billion numbers into a physical representation by hand).

So now, they can work on scaling this process up, and with low enough lead times (I’ll be convinced they have inside connections to TSMC if they can actually deliver on the promised mere 3-4 months delay), will be able to offer 30-100b+ parameter models under half a year after they’re released, at thousands of tokens per second while probably drawing less wattage (per token, not sure about overall).

Exciting times ahead, folks.

Seems anecdotal and/or a recent change. OpenRouter doesn't have my phone number either, but I do have an account. And I didn't sign up with them that long ago.

Correct, and/or in addition, most nowadays prepend something like `set -euo pipefail` to the scripts in the line immediately after the shebang which results in stopping on errors, including things such as syntax errors stemming from e.g. incomplete installer transmission over wire.

(At least for bash scripts, I’m not sure whether these are POSIX syntax to be frank.)

it hallucinates like crazy but looks like its by design to boost benchmarks.

Wasn’t there a discussion around some new-ish benchmark _punishing_ hallucination answers (over not replying at all) recently? Maybe in the not-so-distant future, this “spam replies until one’s correct” strategy won’t be able to game a benchmark much at all anymore.

You forgot about the best part, in terms of the “GDPR threat” effectiveness:

Fines can be up to €20 million or 4% of global revenues…, _whichever is greater._