pona-a

21 ms·

361.
by pona-a 2y ago
Inaction is an action. Facilitating the development and sale of malware while benefiting from it through tax revenue and hard power with full awareness is, to me, enablement. The action taken against such vendors proved nominal, as they sti
362.
by pona-a 2y ago
And... they're gone. While I'm glad it must have gotten someone's attention, hopefully the moderators and not the spammers themselves (does HN allow one to delete 2 day old comments?), now there's no reference point. I p
363.
by pona-a 2y ago
Previous discussion here [0] and here [1]. [0] https://news.ycombinator.com/item?id=17047348 [1] https://news.ycombinator.com/item?id=18854854
364.
by pona-a 2y ago
Just browsing around right now, I already found these [0] [1] [2], the format is identical to those I've been seeing. This one [0] even seems somewhat successful at blending into a discussion, even getting a human to reply once. I foun
365.
Ask HN: What's up with the ChatGPT spam here lately?
19 points by pona-a 2y ago|18 comments
366.
by pona-a 2y ago
Unlike ChatGPT's Android app, this doesn't actually seem to use Google Play Services for anything but OAuth, which is nice. Got it running on my deGoogled phone. It feels quite snappy, hopefully isn't just a webview.
367.
by pona-a 2y ago
I hope this ruling does enough to cast recent decisions of the Supreme Court into doubt, such as the end of Chevron deference. It should have become clear to anyone sufficiently observant that it no longer serves the interests of the people
368.
by pona-a 2y ago
One compelling reason not to would be a region block... [0] https://ai.google.dev/gemini-api/docs/available-regions
369.
by pona-a 2y ago
https://old.reddit.com/r/dataisbeautiful/comments/acow6y/ask... According to this, humans are heavily biased towards 7 when picking random [0, 10] numbers. This is somewhat reflected in LLM's output
370.
by pona-a 2y ago
I tried running the numbers myself. Only did N = ~250 samples, which isn't much, but given OpenAI API pricing and rate limits, this was as much as I was willing to. ~> seq 0 250 | par-each {|x| llm "Sample a random number, from
371.
by pona-a 2y ago
That's what model interpretability research is. You can train an interpretable model from the uninterpretable teacher, you can look at layer activations and how they correspond to certain features, or apply a hundred other domain-speci
372.
by pona-a 2y ago
(defn fac [n] (->> (range 1 (inc n)) (reduce *))) or in Clojure.
373.
by pona-a 2y ago
You mean the Pinephone? I've used a Nexus 5 running Ubuntu Touch for about a year, and it was quite a pleasant experience, discounting unmaintained Canonical abandonware at every corner. I did have to write quite a few apps, but turns
374.
by pona-a 2y ago
Sadly, this didn't seem to track for LLMs. Even OpenAI gave up on trying to detect its own outputs. > As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy. We are working to incorporate feedb
375.
by pona-a 2y ago
Did exactly that, hacked together a small pipeline in Nushell with simonw/llm. Even with GPT-4 turbo and given direct guidelines on common spam heuristics, it seems to perform worse than a Bayesian BoW. Endless trails of questions with
376.
by pona-a 2y ago
And there comes the fact LLMs are actually terrible at coding beyond advanced regurgitation, precisely because their single-pass architecture does not give them Turing completeness to model a Turing machine, which may also be the reason the
377.
by pona-a 2y ago
> ...If they answer doesn't make sense LLM has no ability to retrace the steps. Beam search already exists, it precisely allows the model to backtrack on its past tokens and find the global maxima of probability, at least within set
378.
by pona-a 2y ago
We already do beam search, which allows LLMs to backtrack on branches until they find the highest probability path. Discounting things like Toolformer, this is about as good as it's going to get.
379.
by pona-a 2y ago
Yes, but I believe there are some more objective traits that define good writing beyond just style. It's the use of metaphor, more precise choice of wording, overall speaking less in logical predicates and more so in powerful impressio
380.
by pona-a 2y ago
I've actually done that same concept a while back when whisper.cpp came out. A significant challenge is sane paragraph segmentation, as even humans don't often agree on the best place for a line break. I wonder what approach you&#
381.
by pona-a 2y ago
So instead we need something that is fun and dysfunctional? Let's be honest, even with all the LLM tech, voice interfaces could never be more than a stop-gap before the user does reach for a screen. That's why smart speakers may h
382.
by pona-a 3y ago
I don't have the history stored, but lesser-resource used translation models like English-Ukrainian seem to sometimes produce completely out-of-left-field translations for mispronounced, poorly formatted, or incomplete words. I think t
383.
by pona-a 3y ago
> Congress shall make no law __respecting an establishment of religion, or prohibiting the free exercise thereof__; or abridging the freedom of speech, or of the press; or the right of the people peaceably to assemble, and to petition th
384.
by pona-a 3y ago
I wonder if one can take the English Wikipedia ZIM archive and process it to an efficiently-stored offline map of names to URLs, hopefully even with fuzzy matching.
385.
by pona-a 3y ago
For me, QML, while expressive, had a lot of rough edges. If most of your logic is inside the C++ backend, its JS engine is just useless overhead. If you do something like lite data wrangling in QML, or worse, try to adapt existing JS librar
386.
by pona-a 3y ago
That looks really nice! A healthy balance of reactive UI, expressiveness, and sanity. I wonder if there's a similar Swift binding for Qt... Ubuntu Touch and KDE could really benefit from that, as QML proved to be a total typing nightma
387.
by pona-a 3y ago
"Nobody uses"... Do you understand what you just said? Raw GPS is universally unusable, requiring up to half an hour of waving your phone under clear sky, hoping it a) will actually work; b) will not loose signal; c) will not drai
388.
by pona-a 3y ago
I believe cutting out the whole GPT craze would be best. OpenAI bills are always anything but negligible and I think most users would agree lowering the subscription cost or using it for things like Wolfram integration may be best for the c
389.
by pona-a 3y ago
Fuzz testing seems surprisingly powerful as part of an IDE and doesn't require any new annotation but seems to waste a ton of resources... I wonder if someone developed an algorithm for most efficiently breaking your assertions, perhap
390.
by pona-a 3y ago
But what if we use predicate-based typing? Then your "type" will be something like `(x) ↦ (0 < x < ∞)`, holding true only if the value matches it. Then, given the type function is pure, the hypothetical compiler may somehow

More ›