Something of an aside, but this is sort of equivalent to asking "how does that work for the $50 dollars the casino gave me to gamble with for free"? I once made 50 dollars exactly in that way by taking the casino's free tokens and putting them all on black in a single roulette spin. People like that are not the ones companies like that make money off of.
Google Gemini (2.0 Flash, free online version) handled this rather okay; it gave me an arguably unneccessary calculation of the individual prices of ball and bat, but then ended with "However with the information given, we can't determine exactly how many balls and bats Sally stole. The fact that she has $20 tells us she could have stolen some, but we don't know how many she did steal." While "the fact that she has $20" has no bearing on this - and the model seems to wrongly imply that it does - the fact that we have insufficient information to determine an answer is correct, and the model got the answer essentially right.
IME it is rarely productive to ask an LLM to fix code it has just given you as part of the same session context. It can work but I find that the second version often introduces at least as many errors as it fixes, or at least changes unrelated bits of code for no apparent reason.
Therefore I tend to work on a one-shot prompt, and restart the session entirely each time, making tweaks to the prompt based on each output hoping to get a better result (I've found it helpful to point out the AI's past errors as "common mistakes to be avoided").
Doing the prompting in this way also vastly reduces the context size sent with individual requests (asking it to fix something it just made in conversation tends to resubmit a huge chunk of context and use up allowance quotas). Then, if there are bits the AI never quite got correct, I'll go in bit by bit and ask it to fix an individual function or two, with a new session and heavily pruned context.
The KoboldCpp UI I was using had a pretty straightforward way of setting up a chatbot-style dialogue interface that would send the dialogue context so far with each message input, along with some supporting prompt infrastructure reminding the model of how it's supposed to reply based on that context. It also allowed editing the context directly; but I didn't often do that except to correct simple one-off inaccuracies, instead usually opting to restart from the beginning when things would go off-track.
When the model would reply as multiple characters at once, or interleave narration with the dialogue, I'd usually just play along - most commonly it would decide for 2 or 3 steps to carry on the dialogue just hallucinating the things I would say.
I did a lot of experimenting with different ways to frame the dialogue, from a Zork-like adventure game, to an instant messenger chat, to just paragraphs of prose with dialogue mixed in as if from a novel. I found all of these methods to have different strengths or weaknesses, but the model also tended to blur between them after a few messages so I'd just play along as far as I could each time.
It's just one pretty trivial example, but many months ago I had to develop some backstory for my character in a D&D campaign. I had some character background and some paragraphs I wrote to start off with, but I wasn't sure what direction to go exactly and was trying to come up with different ideas. I dropped what I had into a prompt built to setup a pseudo-text-adventure-game using a local WizardLM model running through KoboldCpp, then I ran through it interactively to the point of getting incoherent about 100 times in a row. This yielded some new ideas, some old ones, and overall an interesting distribution of outcomes; where 2 or 3 happened most commonly, and all 100 were basically variations on about 10.
I won't argue that any of what the LLM came up with would stand on its own as particularly interesting - often quite the opposite - but it showed me examples of all the "obvious directions to go" as well as some other RPG cliches that I could either adopt or choose to avoid, and ultimately served as an excellent brainstorming assistant with interesting ideas and the ability to carry through and embellish them with enough detail to get a sense for either what works or what doesn't.
For me, searching "animals made from circles", your comment put this HN thread as the #1 result while the #2 result was a syndicated article about the linked post. When I get more specific and search "animals drawn only from circles" it turns up the linked post as the first result. But my results may be more specific partly because I don't use ad blockers.
I had the same experience. I played for a few minutes, and the only repetitions I got were countries I already identified on the first guess. I didn't get repeats of any country that I missed initially.
"Document ranking reduces to these hard problems" would imply that document ranking is itself an instance of a certain group of hard problems. That's not what the article is saying.
A key point of Rorschach tests is that they are always the same plates, so there is a large body of literature and case studies specific to interpretations of each one. I think that like any sort of psychological inventory the results are extremely context dependent and require a trained professional to interpret. The trending stuff on social media as mentioned in the article strikes me as more akin to the pseudoscientific misappropriation of things like the MBTI to compare different people in a non-clinical setting, and AFAIU there is really no scientific basis for any such thing.
For me the most conspicuous thing missing is dynamics... particularly when there are "ghost notes" in between much louder ones during a fast passage, it seems like something is missing.
That said, however, I find it to be oddly satisfying to watch. Curious if experience playing different instruments has anything to do with it. To me, something like a xylophone or a steelpan feels pretty analogous to this.
My concern with having the LLM's write tests is that it's hard to be convinced that they've written the right tests. Coupling human TDD with a genetic algorithm of some sort that uses LLM's to generate candidate populations of solutions, one could be assured that once a solution gets far enough through the tests [assuming one ever does], it is guaranteed to have the correct behavior (as far as "correct" has been defined in the tests).
It feels to me that there may be even more potential in flipping this idea around - human coders write tests to exact specifications, then an llm-using coding system evolves code until it passes the tests.
Common Lisp allows one to use arrays or other structures, and can even be used to inline assembly. So despite lists being a major part of the ecosystem and code representation itself, they are not mandatory for implementing an algorithm, or necessarily a performance drawback to using lisp. By the same token, it's easy for people to accidentally use lists and pointers to implement algorithms just as inefficiently in other languages - Python comes to mind. A standard approach in Python is to lean on external libraries for high performance computation, and this can be done just the same way in Lisp - but Lisp can also be used directly to write efficient low level algorithms by making a conscious effort not to use the list- and pointer-based functions to do so.
I've seen a ton of pages in my feed with synthesized photos of things like log cabins, and unlike what the article mentions (comments "ranging from bot-like praise to snarky remarks aimed at users who seem to believe the images are real"), they are entirely met with praise, and never [as far as I can tell] called out by anyone as being AI generated. I assume the negative reactions are being hidden by the comment-ranking algorithm because of their sentiment. But I also wonder about the legitimacy of most of the accounts leaving comments. The whole thing seems like bots.
One thing I noticed is if you set the graph axes to "Speed" and "handling" and the weight slider about 3/4 toward "handling" then Koopa becomes the only member of that 2D slice of pareto front. I suppose this could be a benefit depending on the course design. At least, it's an example of how looking at only 2 dimensions can be limiting.
LLM's-at-home are in a vastly different place than they were 10, 5, or even 2 years ago, so I'm not sure the assumptions about typical traffic from then hold true now.
Though I'm not a particularly big fan of them, the Pholcidae spiders are very nonintrusive, nonaggressive, nearly weightless and paper-thin, and do an excellent job eating venomous spiders and unwanted insects. I tend to leave them undisturbed in out-of-the-way corners when possible.
You can do each variable on its own line with stuff like "defvar" and "setf", but that will add them to the current package namespace - and convention is to avoid polluting namespaces with names that will be unused later. So, if the variables are only used within a given scope, the scope is explicitly delineated with a "let" block. It's not a general requirement.
I'm not sure what you mean about the difference between LET and LET* (the latter simply lets subsequent variable declarations refer to previously declared variables in the same block), and LETREC is not a builtin part of Common Lisp.
If one ever tried to make edits to the latents prior to decoding them with a VAE in SD1.5 and then in SDXL, it could be seen that that local changes had somewhat unpredictable and global effects on the image in SD1.5, while in SDXL the changes have more predictable impacts to the output image and some of the different latent channels end up corresponding more directly to the resulting image channels.
Definitely a fascinating write-up. I have been curious about these differences for a while, though I had never considered this a "problem" per se.
I find speaking to an LLM as if it were a human to be far more effective than barking orders at it. Although this is less true of some models compared to others if they've been fine tuned to respond to that sort of thing. Even still in general I don't find barking orders to be very effective.
I'd have to agree, my only point was that people may think they remember having watched it a certain way simply because they read the viral headlines so many times
I can only speculate, but I have seen this linked on social media dozens of times, often including that detail of
"shutting down an extremely hostile senator" in the caption. So whether or not one interprets the senator as rude or hostile, people are getting exposed to that idea pretty regularly.
I had the idea to wait for a few letters so I could type them all at once, but by then, it was too hard to figure out which one had come first (I'd assumed it didn't matter and was just trying left-to-right) before they all reached the bottom and I couldn't catch up.
I have attempted this sort of thing with GPT 3.5 many times and never been successful, although I've still never been taken off of the GPT4 waiting list that I signed up for months ago and I'm not going to subscribe without trying it first. I [and presumably many thousands of others] have tried things like this with many LLMs and image generating models, but to my knowledge we've come up rather short. I've never managed to recreate anything verbatim and have struggled to get anything resembling a copyright infringement out of stable diffusion with the sole exception of a meme image of Willy Wonka.
That said, the meme image of Willy Wonka comes out of stable diffusion 1.5 almost perfectly with surprising frequency. Then again, this is probably because it appeared hundreds or thousands of times in the training set in all sorts of contexts because it's such a popular meme. There is a tension between its status as an integral part of language and its nature as a copyrighted screen grab.
Agreed; the suggestion that people just expect random shuffle to work differently from that seems well-meaning and rooted in some kind of truth but quite disconnected from the fact you stated. I'm not sure why media players that offer something other than a true "shuffle" can't just provide the option to have either functionality.
In the context of Shakespeare, I'd agree that there may be some competitive potential in the product. But in the context of news, something that evolves and relies on timely and accurate information, I don't see how something like that turns into competition for the NYT by being trained on past NYT outputs.
If the argument is that people can use ChatGPT to get old NYT content for free, that can be illustrated simply enough, but as another commenter pointed out, it doesn't really seem to be that simple.
I'm not sure if the verbatim content isn't more of a "stopped clock is right twice a day" or "monkeys typewriting shakespeare" situation. As I see it, most of the value in something like the NYT is as a trusted and curated source of information with at least some vetting. The content regurgitated from an LLM would be intermixed with false information and all sorts of other things, none of which are actually news from a trusted source - the main reason people subscribe to the NYT (?) and something at which ChatGPT cannot directly compete against NYT writers.
Brake lights being the color of stop lights is helpful because it's a reminder to stop, just as turn signals being the color of caution lights is helpful as a reminder to be cautious of a turning vehicle. Green lights don't fall into either category, being generally a cue to step on the gas.