gojomo

7 ms·

61.
by gojomo 1y ago
Pre-opening BART tubes were definitely used for George Lucas's first feature film, THX-1138: https://www.sfgate.com/streaming/article/bart-transbay-tube-...
62.
by gojomo 1y ago
Grandparent testimony of success, & parent testimony of frustration, are both just wispy random gossip when they don't specify which LLMs delivered the reported experiences. The quality varies wildly across models & versions.
63.
by gojomo 1y ago
Given that other work shows that models often converge on similar internal representations, I'd not be surprised if there were close analogues of 'subliminal learning' that don't require shared-ancestor-base-model, just
64.
by gojomo 1y ago
I'd expect a noticeable delay with current local LLMs - especially visiting a site for the 1st time. But then they could potentially memoize their heuristics for certain designs, including recognzing when some "deeper thought&quot
65.
by gojomo 1y ago
~simonw's demo of a quickie customized HN front-end is great. But ultimately, your browser should have a local, open-source, user-loyal LLM that's able to accept human-language descriptions of how you'd like your view of some
66.
by gojomo 1y ago
https://en.wikipedia.org/wiki/A_Void
67.
by gojomo 1y ago
Thanks, that's great! But: if all developers did 136 AI-assisted issues, why only analyze excluding the 1st 8, rather than, say, the first 68 (half)?
68.
by gojomo 1y ago
Did each developer do a large enough mix of AI/non-AI tasks, in varying orders, that you have any hints in your data whether the "AI penalty" grew or shrunk over time?
69.
by gojomo 1y ago
It doesn't. Signal opts its data out of even encrypted local iPhone backups.
70.
by gojomo 1y ago
And note: Signal has been misleading users about options for iOS backups for a decade! In September 2015, users were told it was "on the roadmap": https://github.com/signalapp/Signal-iOS/issues/905#i
71.
by gojomo 1y ago
Do such backups retain Signal message history - unlike Apple's native local backups to MacOS?
72.
by gojomo 1y ago
News to me! Where is this option described, ideally by Signal itself? If you are alleging that Apple's own local Finder/Itunes backup of an iPhone includes Signal messages, that's not true, against reasonable user expectation
73.
by gojomo 1y ago
The post claims with regard to the cost of a cloud backup that "Local backups still exist" - but that's a lie, there's no local backup option on iOS.
74.
by gojomo 1y ago
> far from novel Techniques can be arbitrarily old & common in industry, but still be a novel academic paper, first to document & evaluate key aspects in that separate (& often lagging) canon.
75.
by gojomo 1y ago
No - and you can compare the style & written tics for continuity with my 18y of posts here. I used 'delving' in an HN comment more than a decade before LLMs became a thing! https://news.ycombinator.com/item?id=
76.
by gojomo 1y ago
Yes, but: for humans, even without an expert-over-the-shoulder providing fresh feedback, drilling/practice works – with the right caveats. And, counter to much intuition & forum folklore, it works for AI models, too – with analogou
77.
by gojomo 1y ago
Of course, training on synthetic data can't do everything! My main point is: it's been doing a bunch of surprisingly-beneficial things, contra the obsolete beliefs about model-output-worthlessness (or deleteriousness!) for further
78.
by gojomo 1y ago
OK, sure, there are gradations. The new encoding can contain a FLOAT32 side channel on every character, to represent its proportional "AI-ness" – kinda like the 'alpha' transparency channel on pixels.
79.
by gojomo 1y ago
Perhaps. But these models can already clearly write about the world, in useful ways, without such 'qualia' or 'biological underpinnings'.
80.
by gojomo 1y ago
No, new more-capable and/or efficient models have been forged using bulk outputs of other models as training data. These inproved models do some valuable things better & cheaper than the models, or ensembles of models, that generat
81.
by gojomo 1y ago
The training sets can already include direct data series about the world, where the "work of human beings" is just setting up the the collection devices. So models can absolutely "experience the world". But I'm not
82.
by gojomo 1y ago
Less than you might think! Some of the frontier-advancing training-on-model-outputs ('synthetic data') work just uses other models & automated-checkers to select suitable prompts and desirable subsets of generations. I find
83.
by gojomo 1y ago
This was an intuitively-appealing belief, even with some qualified experimental support, as of a few years ago. However, since then, a bunch of capability breakthroughs from (well-curated) AI generations has definitively disproven it.
84.
by gojomo 1y ago
Look, we just need to add some new 'planes' to Unicode - that mirror all communicatively-useful characters, but with extra state bits for... guaranteed human output - anyone who emits text in these ranges that was AI generated,
85.
by gojomo 1y ago
Lacks context & examples to know what they're concerned about. Has a righteous, bossy tone that doesn't seem earned by case particulars or its (anonymous) author. "Mozilla: Improve your messaging. Now."
86.
by gojomo 1y ago
There is no de jure legal requirement that the RIAA, Disney, Nintendo, or the government be "pleased to hear" about new technology. And, while copyright prohibits some sorts of reproduction of copyrighted materials, it doesn'
87.
by gojomo 1y ago
The giveaway that Marcus is generating slop for consumption by unsophisticated AI-haters is that he doesn't bother to mention what models/options involved.
88.
by gojomo 1y ago
After I saw third-party "10 year battery enclosure" offerings for AirTags, was wondering when other workaround customizations like this might appear. Other impactful variants might be: * senses whether another 'sibling'
89.
by gojomo 1y ago
Is there a frozen client that someone could audit for assurance, then repeatedly use with your TEE-hosted backend? If instead users must use your web-served client code each time, you could subtly alter that over time or per-user, in ways u
90.
by gojomo 1y ago
Prior discussion when the paper was 1st reported in February: https://news.ycombinator.com/item?id=43176553

More ›