krackers

17 ms·

331.
by krackers 7mo ago
We often talk about "aligning models" or training them, little attention is paid to how models align/train _us_ as we interact with them. The reward functions they're trained under get "backpropagated" into our
332.
by krackers 7mo ago
There's one difference that if a program is run as tool call, the internal states and control flow are not visible to the LLM. You can imagine this being useful for "debugging" in a meta-sense, the same way humans can use deb
333.
by krackers 7mo ago
What level is copy pasting snippets into the chatgpt window? Grug brained level 0? I sort of prefer it that way (using it as an amped up stackoverflow) since it forces me to decompose things in terms of natural boundaries (manual context ma
334.
by krackers 7mo ago
This isn't even their tour de force, try https://aresluna.org/frame-of-preference/
335.
by krackers 7mo ago
Catalyst was already sort of a death knell, since it's an admission that it's ok to port over iPhone/iPad HIG to mac. Maybe swiftUI too, since it's replacing appkit and all its various affordances.
336.
by krackers 7mo ago
Expert systems are basically decision trees which are "gofai" (good old fashioned ai) as opposed to deep learning. I've never really seen a good definition for what counts as "gofai" (is all statistical learning&#x2
337.
by krackers 7mo ago
I remember reading somewhere that to target gen-z/alpha advertising shouldn't be direct or in your face but instead present things in an "organic" way that makes it seem cool, without the appearance of trying too hard to
338.
by krackers 7mo ago
you should have an indicator for "foe of foe" as well
339.
by krackers 7mo ago
See https://news.ycombinator.com/item?id=47301241 particularly the tiktok shorts.
340.
by krackers 7mo ago
And beyond 20/20/20 rule, try to physically go outdoors in sunlight if possible. You can go down by 0.5 diopters if you do this consistently enough.
341.
by krackers 7mo ago
>I do not think this is the last we’ve seen of Lil Finder Guy… https://www.macrumors.com/guide/siri-chatbot/ > Apple is planning to make visual design changes. It's not quite clear what that will entail
342.
by krackers 7mo ago
> best nutrition, daycare, early childhood learning, classes, tuition In the past you didn't have to do any of this to be able to have a decent "middle class" living, now if one doesn't do all of that then they are al
343.
by krackers 7mo ago
Where else would they park their wealth? Stock market is all imaginary money. Maybe gold/silver. But housing works as an easy way to park wealth in a tangible way that's guarded against inflation. This fact is probably also why ho
344.
by krackers 7mo ago
>designer is that this skill is learnable Why do you think this? Being a designer is ultimately a matter of "good taste" and intuition for HIG (that you learn to systematize and formalize) and not everyone has this to start off
345.
by krackers 7mo ago
I'm reminded of the notion of "komolgorov complexity" here. There might be some tasks for which a short natural language description is sufficient, and others for which a sufficiently formal description is needed to the point
346.
by krackers 7mo ago
I'm guessing there's a very strong prior to "just keep generating more tokens" as opposed to deleting code that needs to be overcome. Maybe this is done already but since every git project comes with its own history, you
347.
by krackers 7mo ago
No, I think it really did allow it. You can see the source for those emoji enabler apps in [1]. I think that various AppKit APIs do end up writing to that file so r/w access to that file may have been required. And in those days sandbo
348.
by krackers 7mo ago
>wasn't X, it was Y >It’s like... The guy is clearly talented and it's a great story, why would you let an LLM tell it for you? >And by extracting session keys, they could silently monitor supposedly private communication
349.
by krackers 7mo ago
I was curious as well, https://blog.mozilla.org/security/2021/01/07/encrypted-clien... has some info >analysis has shown that encrypting only the SNI extension provides incomplete protection. As just
350.
by krackers 7mo ago
I think you typo'd, "sixteen (n^2) different combinations of points" should be 2^n instead.
351.
by krackers 7mo ago
>A 180° arc cannot straddle a 180° gap This can be the case if the other point is exactly 180 degrees from the anchor point that works though? But I think this case occurs with probability 0 ("almost never") so can basically be
352.
by krackers 7mo ago
>can't meaningfully see and interact with the page like the end user will Isn't this a great use case for LLM tests? Have a "computer use agent" and then describe the parameters of the test as "load the page, the
353.
by krackers 7mo ago
What features do you think it needs, that wouldn't spoil the "elegance" of the language? I think one good feature would be higher order messaging, in fact there's already a PL paper discussing how it looks like in object
354.
by krackers 7mo ago
Well they used strings of < 800 chars, you probably run into context window and training limits at some point (they mention some result that you need at least something of GPT-2 size to begin recognizing more intricate CFGs (their synthe
355.
by krackers 7mo ago
No, this statement is not true for anything except a base model. Benchmaxxing during RL phase is how you get the advertisement style "punchy" writing, because even though people don't usually write that way it is eye catching
356.
by krackers 7mo ago
Figure 12 shows probabilities I think, it actually does seem to be 100% at temperature 0.1 for certain pretraining runs.
357.
by krackers 7mo ago
This already happens, user vs system prompts are delimited in this manner, and most good frontends will treat any user input as "needing to be escaped" so you can never "prompt inject" your way into emitting a system rol
358.
by krackers 7mo ago
This is very interesting since there is another notable paper which shows LLMs can recognize and generate CFGs https://arxiv.org/abs/2305.13673 and of course a^n b^n is also classic CFG, so it's not clear why one
359.
by krackers 7mo ago
All system prompts are already wrapped in specific role markers (each LLM has its own unique format), so I'm sure every lab is familiar with the concept of delimters, in-band vs out-of-band signalling and such. It'd not clear why
360.
by krackers 7mo ago
Unless I've misunderstood the math myself, I don't think GPs comment is quite right if taken literally since "predict the next 2 tokens" would literally mean predict index t+1, t+2 off of the same hidden state at index t

More ›