veunes

6 ms·

91.
by veunes 7mo ago
Curious how they'll handle accessibility/prescription edge cases (especially if someone relies on assistive features built into these devices)
92.
by veunes 7mo ago
I suspect it'll be selective
93.
by veunes 7mo ago
Part of the problem is that modern apps aren't really "one thing" anymore
94.
by veunes 7mo ago
Not "C++ everywhere again" but maybe "understanding memory again"
95.
by veunes 7mo ago
Nah, those are completely different beasts. DeepSeek's MLA solves the KV cache issue via low-rank projection - they literally squeeze the matrix through a latent vector at train time. TurboQuant is just Post-Training Quantization where
96.
by veunes 7mo ago
Classic academic move. If the authors show accuracy-vs-space charts but hide end-to-end latency, it usually means their code is slower in practice than vanilla fp16 without any compression. Polar coordinates are absolute poison for parallel
97.
by veunes 7mo ago
Looks like Google canned all their tech writers just to pivot the budget into H100s for training these very same writers
98.
by veunes 7mo ago
It's funny on the surface, but it also kind of highlights how broken the interaction is
99.
by veunes 7mo ago
This is probably the part people underestimate
100.
by veunes 7mo ago
The concern here is that regulators are starting to assume global reach by default, even when that leverage isn't obvious
101.
by veunes 7mo ago
The tricky part is that geoblocking is inherently porous
102.
by veunes 7mo ago
520k sounds like a lot until you realize it's basically unenforceable if the company just ignores it
103.
by veunes 7mo ago
Nobody in their right mind builds a pipeline where security relies on a custom container runtime catching things after the fact. Security starts in CI at the image build stage. If your flow actually lets a vulnerable Next.js build slip all
104.
by veunes 7mo ago
Reviewing generated code actually takes a higher skill level than writing it. A junior who prompted this Next.js app into existence is physically incapable of auditing the security of those imports. And for a senior it's often cheaper
105.
by veunes 7mo ago
The system works because responsibility sits with the submitter
106.
by veunes 7mo ago
The real invariant is responsibility: if you submit a patch, you own it. You should understand it, be able to defend the design choices, and maintain it if needed
107.
by veunes 7mo ago
I think framing it as a trust question is exactly right
108.
by veunes 7mo ago
Accessibility is an angle that rarely comes up in these debates and it's a strong one
109.
by veunes 7mo ago
The quality argument against LLM-generated code has always seemed weak to me. Maintainers already review patches because humans routinely submit bad code. The review process is the filter.
110.
by veunes 7mo ago
"Fake it till you make it" in a nutshell. Half the AI wrappers on the market do the exact same thing: they render pretty activity charts that have absolutely zero correlation with actual VRAM consumption or server-side inference l
111.
by veunes 7mo ago
Impressive numbers for a spam bot, but what's the point if the content is generated by an LLM and the comments are written by other agents? The internet is already turning into an endless feedback loop of generated garbage where the on
112.
by veunes 7mo ago
"In principle" - sure, but in practice, even if you pin the seed, your float32 calculations are going to drift due to non-deterministic CUDA kernels during parallel execution. You'll never get bit-for-bit identical tensors ac
113.
by veunes 7mo ago
Even if you pin the seed and spin up your own local LLM, changes to continuous batching at the vLLM level or just a different CUDA driver version will completely break your bitwise float convergence. Reproducibility in ML generation is a to
114.
by veunes 7mo ago
Perfect analogy. Nobody cares how many times you googled "how to center a div" before finally writing proper CSS. Same goes for agents: I only care about the final architectural state and performance, not how the model brain-farte
115.
by veunes 7mo ago
The idea of "saving prompts for reproducibility" is dead on arrival. LLMs are non-deterministic by nature. In a year, they'll deprecate this model's API, and the new version will spit out completely different code with e
116.
by veunes 7mo ago
Dumping panic traces to an agent works fine if it's just a vanilla page fault at an obvious address. But when your memory gets corrupted by some scuffed DMA sync or a race condition in an interrupt handler, the kernel panics a million
117.
by veunes 8mo ago
I think that's mostly right, but Google still is part of the problem because it normalized the idea that the tradeoff should be invisible
118.
by veunes 8mo ago
A lot of "this service is terrible" turns out to be "I've accumulated ten years of bad habits around this service"
119.
by veunes 8mo ago
The hardest part of leaving big platforms usually isn't technical, it's psychological
120.
by veunes 8mo ago
If just 16 million examples were enough to significantly boost model quality (as Anthropic claims), it turns out that data quality beats quantity Instead of vacuuming petabytes of trash from Common Crawl, you can just take high-quality dist

More ›