naasking

Your incredulity that such things are possible for reasons you do not yet understand is not a persuasive argument. There are many possible explanations that completely eliminate conspiracy if you'd even bothered to look.

Just think of it intuitively: we don't hold babies or the insane responsible for their choices partly because they act randomly with no rhyme or reason, but you become responsible for your choices as you grow and your choices become more understandable and predictable in a way; in some real sense, you're more responsible the more deterministic you are!

It is crazy how readily people believe wild accusations with little credible evidence though. Not addressed at you specifically, just this whole subthread is wild.

Free will is almost certainly compatible with determinism (cf. Compatibilism).

> or the retrofit is going to be painful and probably will never happen.

They're using LLMs. The economics of this sort of retrofit has changed considerably with LLMs.

> I think you at least need a formal semantics for both C

https://en.wikipedia.org/wiki/CompCert

Superdeterminism is local actually, although accepting non-locality is another way to satisfy Bell's theorem.

Remote alien knowledge doesn't give us new useful tools unless we make contact with them. Fortunately we now have contact with alien intelligence right here that is providing us with useful tools.

It's a social project where humans collaborate because they want useful tools, and sharing the useful tools they built with others gets their own tools improved. The reciprocation of sharing knowledge and knowledge artifacts compounds. The advent of AI does not change this formula one bit, if anything AI increases the compounding factor. You are of course free to do as you wish with this new reality.

Humans aren't the only systems that can wield knowledge now.

Either you care about advancing the frontier of knowledge or you don't. If you care, then you should use every tool available to you. If you don't, then society has no obligation to fund your personal hobby interests. Calling this "selling out" is ridiculous.

Why would someone contribute to open source without getting paid?

Is 50% good? It's what you'd expect by posing "will AI do X by 2026?" and then flipping a coin to get an answer.

The price went up because demand went up. If you want incentives to build more infrastructure to produce more supply so the price can come down, it has to first go up.

Do you really think reading a snippet like that completely out of context should qualify as probable cause?

What is the actual problem being solved though?

Driving the compressor with software is not intrinsically a problem, but critical functions have to be isolated to a minimal safe kernel that will always run, like the kernel space/user space distinction. Smart features can read and send messages to the minimal kernel, but should not themselves be in the critical loop. Boggles the mind that this is not standard practice.

My Samsung 3D TV is still going strong after 13 years.

There were too many unsolved problems with diffusion to commercialize it, where autoregressive loops had a more straightforward roadmap. Doesn't mean the diffusion problems are insurmountable though.

> $20M for 1.7MW of capacity. Over $10 per watt. Recent utility scale installations in California are more like $1/W.

It's a) just a trial, and b) not just generating power so that cost isn't factoring in the other benefits.

I don't think so, I typically use Opus 5 on High, and 5.5 scores lower on token use:

https://artificialanalysis.ai/models/claude-opus-5-5?models=...

llama.cpp has significant AI contributions. The Linux kernel too is receiving AI patches. You are seriously out of touch.

> LMs generate text output that appears to be useful, but regularly is not.

No, they are empirically useful, and only getting more useful. This is not even a debate anymore.

Is there a known source of internal flaws/porosity in an otherwise solid part? Presumably laser melting produces a puddle which shouldn't allow for internal pores, as long as it isn't printed too fast (or solidifies too fast, which is why I think most chambers are internally heated to near melting temp).

Re: surface roughness, I can understand that the powder grain size creates a sort of minimal structure size, and can in principle be the start of a crack if a surface grain gets knocked loose. Is that the sort of thing you mean? I can see that for any internal or external surfaces, and a rocket engine combustion is certainly applying a lot of heat and pressure on surface grains. Can this be alleviated by smaller grain sizes, or is there some limit there?

Re: repeated heating/cooling and internal stresses, this strikes me as just requiring standard post-printing stages like tempering to alleviate internal stresses.

Sure, but I mean what's the technical reason a material isn't it 3D printing friendly? Are we talking grain structure here? Is it something that can be at least partly mitigated by some post-printing heat treatments, like tempering?

Individual incidents aren't representative of accuracy or well-tuned truth seeking processes, that can only be assessed over time.

> LLMs are vectorial databases with losses that index statistically filled data

Yes, and that statistically filled data is insanely useful. It remains true that it's a relatively poorly understood how this can be applied in various scenarios and what processes are needed to ensure robust results (or quantify the uncertainty).

What's the current theory for why this is?

Model, Q4_K_M: https://huggingface.co/agentionai/Signal-3.8-27B-GGUF

DFlash2, Q8_0, --spec-draft-n-max=7: https://huggingface.co/z-lab/Qwen3.8-27B-DFlash2-GGUF

I run llama.cpp with -ctv=8, -ctk=q4. Vulkan has better throughput if you're doing single-stream decode, but ROCm has better throughput if you have "--parallel 2" or higher. If supporting parallelism, unified kv cache should be off, especially with Vulkan.

Of course, some of these may be specific to my card so try variations for your hardware. Hermes can concoct a test suite and run some tests for different llama.cpp parameter permutations to find something optimal.

I've found the opposite on my R9700 (n-max=7, no other speculative decoding like ngram-mod, which I found slows it down). I think it depends whether your workload and system are bandwidth limited or compute limited. I see draft acceptance around 0.55, so 0.55 * 7 = 3.8 tokens per pass, which on my bandwidth-limited card takes me from 30tps to a peak of 80tps on llama.cpp (MTP peaked at ~65tps). I'm also running a Qwen fine tune whose speculative execution is better than the base model.

Strix Halo has lower compute than the R9700 but the RAM is also slower, so not sure what would be the ultimate limiting factor.