stymaar

9 ms·

151.
by stymaar 2mo ago
Both passive corruption (accepting a bribe) and active corruption (giving a bribe) are a problem. “But the officials accepted it” is not a valid way to discount Amazon from their nefarious practices.
152.
by stymaar 2mo ago
> Basically you built a miniature, integrated, embedded, open-source web browser which can parse CSS and render to SDL... very impressive!: Modern LLMs are indeed impressive.
153.
by stymaar 2mo ago
Why the Apple copyright and ToS at the bottom of the page? Is that an Apple product?
154.
by stymaar 2mo ago
> then you should be able to add features by just reading the acceptance test and scanning unit tests. That “just” is bearing a lot of weight though as tests are often even longer than the code itself, in addition to being excruciating t
155.
by stymaar 2mo ago
Someone had to come up with needle in the first place though. And it's the kind of thing that's going to come from a big lab with an AGI monopoly ambition.
156.
by stymaar 2mo ago
True. I'm a big fan of Cactus's work on the needle family of Simple Attention Network: https://github.com/cactus-compute/needle
157.
by stymaar 2mo ago
> though vibed software (thoroughly used) can be all good. Yes, but the problem with these vibe-coded crap is that they are pretty much always less than a week old, which means it wasn't even used before the “author” submitted it he
158.
by stymaar 2mo ago
> They are highly adaptable even without fine tuning, in-context learning is still superior to fine tuning in most cases also. Good luck relying on in-context learning for a 600M LLM. > The actual adapter training is automated and put
159.
by stymaar 2mo ago
> . But specifically when it comes to LLM engineering, no there's really not much you can do, they are called "large" The “large” qualifier dates back to pre-transformer language models, where even training a multi-million
160.
by stymaar 2mo ago
The problem is that even Fable still make trivial yet high impact mistake when let on their own, and then you'd need to read the whole code to catch them… Meanwhile they are very good at implementating an explicit algorithm that you fe
161.
by stymaar 2mo ago
Yeah, especially since most of these are already available for purchase from data brokers.
162.
by stymaar 2mo ago
> It took him two hours of passing errors to Claude for the endpoint to start working What? It's literally three actions and you're good: download llama.cpp, download the model on Huggingface, and run it with. I have no idea ho
163.
by stymaar 2mo ago
> If your software has the affordance of a waiting dialogue or loading wheel for many of its UI controls, you are building with this default blocked assumption. Even if you are building something web based, ask yourself if that's ac
164.
by stymaar 2mo ago
Illegitimate abuse of border control power to restrict freedom of speech.
165.
by stymaar 2mo ago
> A small team today, running 20-100 agents in parallel, might generate 500 commits/200 pushes/100 PRs Who the fuck is writing the requirements in that story?! Setting aside the problem of pushing AI-generated code that nobody
166.
by stymaar 2mo ago
If the underlying probability distributions are the same, then DFlash can lead to an invalid Python Syntax iif the autoregressive process could have generated one if the random sampling picked a different token. If a model can output a “wro
167.
by stymaar 2mo ago
> I mean, I guess I disagree that you need anything close to a perfect model. It's an adversarial setting, economic actors are incentivized to find any loophole and exploit them, so even if it doesn't need to be perfect it ne
168.
by stymaar 2mo ago
Can you explain a little bit more please?
169.
by stymaar 2mo ago
Related: Stop Anthropomorphizing Intermediate Tokens as Reasoning/Thinking Traces! [1] > Our findings consistently challenge the prevailing narrative that intermediate tokens constitute a semantically meaningful reasoning process.
170.
by stymaar 2mo ago
Gosh, what kind of conspiracy theory is that?! The “copper plate” rule appeared in 1998, long before the phaseout of nuclear and the move to wind power. Your argument completely confuses cause and effect: the reason why wind was pursued was
171.
by stymaar 2mo ago
Even if they don't match current-day Opus in everything, they do beat 6 month old Opus, which we have no reason to believe it was smaller than the latest version.
172.
by stymaar 2mo ago
The fact that Musk claims Opus is 5T to justify why Grok is far behind should be taken with a massive grain of salt given he's a recidivist mythomaniac. Honestly if Opus is 5T parameters while being matched by the biggest open models t
173.
by stymaar 2mo ago
This paper, as well as the Chinchilla one, aged like milk though.
174.
by stymaar 2mo ago
Storing general knowledge in VRAM has always been a dumb idea in the first place.
175.
by stymaar 2mo ago
> If you moved the ownership and control of the transmission network and all generators into a single operator that was responsible for optimising its operation in near-real-time and that bore all the costs within a single accounting per
176.
by stymaar 2mo ago
June 2025 Someone else shared the up to date info about this feature: https://doc.e.foundation/os/apps/voice-to-text > The earlier Voice to Text was a Premium-only, online feature: it streamed your speech to a
177.
by stymaar 2mo ago
Unfortunately it looks like it's not just a issue of missing hardware security features: > Their partnership with Murena along with promoting it themselves with misleading marketing means no possibility of working with us. I want to
178.
by stymaar 2mo ago
It's not about border, I'm talking about power flow through Germany. And of course it's a market design issue, but as I said above, the problem is that you'll always face market design issues because the market designers
179.
by stymaar 2mo ago
> insanely high tokens-per-second especially when served from hosted providers, though, given how tiny it is (37B!) It's a dense model so it will use all of its parameters per token. 37B active parameters isn't tiny at all, i
180.
by stymaar 2mo ago
There's no Qwen3.8-35B-A3B though.

More ›