klabb3

22 ms·

241.
by klabb3 1y ago
Disagree. Untrusted code was thought to be a meaningful term 20-30 years ago when you ran desktop OSs with big name software like Microsoft Word and Adobe, and games. What happened in reality is that this fence had false positives (ie Meta
242.
by klabb3 1y ago
> That's asshole behavior, and worth naming, shaming, and ostracizing over. These exploits are being developed, distributed and orchestrated by Meta. The ”millions of websites” are just hummus recipe content farms using their ad SDK
243.
by klabb3 1y ago
I have no idea why you're downvoted. Why on earth would they delete their most valuable competitive advantage? Isn't it even in the fine print that you feed them training data by using their product, which at the very minimum is l
244.
by klabb3 1y ago
Yeah. I never liked the blackbox property (it works but we don’t know why). But the second order effect that we’re witnessing now is basically ”we are going to assume it’s good at everything unless proven otherwise”. This to me is religious
245.
by klabb3 1y ago
I love Go, but this is almost farcically hilarious: > The goal of the proposal process is to reach general consensus about the outcome in a timely manner. If proposal review cannot identify a general consensus in the discussion of the is
246.
by klabb3 1y ago
Yeah. Reviewing code thoroughly is extremely time consuming. When you review code from a human, you can look for choices they made fairly quickly - say they used framework X and language feature Y. Most importantly, you assume that they’v
247.
by klabb3 1y ago
> What are some notable uses of P2P these days? Im using it for Payload[1] in for LAN and WAN transfers (if possible). Reduce operational costs (especially if you run on public clouds and have to pay extortion rates for egress) and also
248.
by klabb3 1y ago
Intuitively I agree. In the long run, we’ll know better. But for now, nobody truly knows what the new equilibrium is. That said: it’s one type of work that is getting dramatically cheaper. The debate is about the scope and quality of that l
249.
by klabb3 1y ago
Well yes but there is no better way to measure without resorting to pure hearsay. How would you make an accurate assessment of something so inherently vague?
250.
by klabb3 1y ago
Wait until they hear how good dictionaries are at spelling.
251.
by klabb3 1y ago
Pareto is about diminishing returns. > but each succeeding iteration seems to be more disappointing This is because the scaling hypothesis (more data and more compute = gains) is plateauing, because all text data is used and compute is
252.
by klabb3 1y ago
I know it’s technically reproducible under the right conditions, and sure it might help in some cases. But it matters little in practice – the issue is that it’s unstable relative to unrelated parameters you often have good reason to chan
253.
by klabb3 1y ago
> If AI creates a mess, why won’t future AI be able to continue building a mess on top? Because due to how complexity works, you reach a state where the expected number of breakages from modifying the code base exceeds the value from the
254.
by klabb3 1y ago
Don’t forget that they give away the data the AI needs for training.
255.
by klabb3 1y ago
Also worth mentioning that tools have stable output. An LLM is not a tool in that sense – it’s not reproducible. Changing the model, retraining, input phrasing etc can change dramatically the output. The best tools are transparent. They are
256.
by klabb3 1y ago
> but the interconnectedness and therefore state space will explode polynomially without applying encapsulation and local reasoning I wish the complexity demon was this gentle. It's exponential generally – a single boolean flag can
257.
by klabb3 1y ago
I suspected as much. It would have been too obvious of an attack vector for something so sensitive. Then obviously my argument falls apart, since it’s no longer saves any config. That said, you can still benefit from pub keys by having good
258.
by klabb3 1y ago
Both are http requests from client to server. Servers are already authenticated through TLS. The difference is who takes the role of the client. With API requests the customer takes the client role. The endpoint is the same, eg api.stripe.c
259.
by klabb3 1y ago
> Maybe LLMs do teach us something about language... that language is complicated. It certainly teaches us many things. But an LLM trained on as many words (or generally speaking an AI trained on sounds) in similar quantities of a toddle
260.
by klabb3 1y ago
Plain data in all glory but I’d rather (deterministically) generate and modify it than having big blobs checked in. If I have too much copy pasted data I often forget to modify until the test case runs and I realize it. This creates a false
261.
by klabb3 1y ago
> If the LLM wrote a harness and proof of concept tests for its leads, then it might increase S/N dramatically. Designing and building meaningfully testable non-trivial software is orders of magnitude more complex than writing the b
262.
by klabb3 1y ago
> "Data for me but not for thee" Exactly. Consumer tech has been locking down APIs and actively working against interop since the advertising business model became dominant. Web 2.0 is the most obvious example, but there are pl
263.
by klabb3 1y ago
Im not really much of a flatpak user but to me it seems like permission system on top of Linux is an incredible undertaking. Solving both packaging and retrofitting permissions at the same time seems too big of a cookie to swallow. I don’t
264.
by klabb3 1y ago
> Saying that current LLMs are only useful for the mainstream is saying that they're incredible useful. Yes and no. If you have a hammer that works 98% of times, you might wanna use it a lot or not at all, depending on the circumsta
265.
by klabb3 1y ago
> The next major device will definitely be AI-first. Everything so far that has been named X First has been marketing woo woo, and in practice only meant "we're thinking about this use case a little more than before". Such
266.
by klabb3 1y ago
Maybe. A bigger impact factor imo is the distance between the tech you're using and the center of mass of training data. Showcase site in NextJS with a CSS gradient hero banner and a news letter? AI will be amazing. But then again, so
267.
by klabb3 1y ago
Aren’t American startups enjoying access to capital markets to a much larger extent, which in turn is a result of financialization of the US economy and demand for USD as a reserve currency (ie unrelated to tech itself)? My impression is US
268.
by klabb3 1y ago
> but what it means is left for readers questioning, your "vision" here is your text prompt? Right. Imo you have to be imagination handicapped to think that creative vision can be distilled to a prompt, let alone be the medium
269.
by klabb3 1y ago
As someone with a decent level of familiarity, this post was good (even taught me a couple new things). Unfortunately this topic is so difficult that it can only be briefly introduced in a medium sized blog post. Mastering fences, compiler
270.
by klabb3 1y ago
Ah, I think I misunderstood the sentiment. Thanks for clarifying

More ›