I've spent a few weeks building and using a terminal LLM client based on that RLM paper that was floating around a little while ago. It's single-conversation, with a tiny, sliding context window, and then a tool that basically fuzzy searches across our full interaction history. It's memory is 'better' than mine - but anything that is essentially RAG inherently will be.
My learning so far, to your point on memory being a limiting factor, is that the system is able to build on ideas over time. I'm not sure you'd classify that as 'self-learning', and I haven't really pushed it in the direction of 'introspection' at all.
Memory itself (in this form) does not seem to be a silver bullet, though, by any means. However, as I add more 'tools', or 'agents', its ability to make 'leaps of discovery' does improve.
For example, I've been (very cautiously) allowing cron jobs to review a day's conversation, then spawn headless Claude Code instances to explore ideas or produce research on topics that I've been thinking about in the chat history.
That's not much different from the 'regular tasks' that Perplexity (and I think OpenAI) offer, but it definitely feels more like a singular entity. It's absolutely limited by how smart the conversation history is, at this time, though.
The Memento analogy you used does feel quite apt - there is a distinct sense of personhood available to something with memory that is inherently unavailable to a fresh context window.
I think a hidden problem even if we solve memory is the curation of what gets into memory and how it is weighted. Even humans struggle with this, as it's easy to store things and forget the credibility (or misjudge the credibility) of the source.
I can envision LLMs getting worse upon being given a memory, until they can figure out how to properly curate it.
Roughly, we had Cursor software engineers record real questions they were asking models, and then had them record the PR that they made that contained the result. We then cleaned these up. That is the benchmark.
Are you able to give a sense of how many questions, which domains they were split over, and how that split looked in % terms?
As a user, I want to know - when an improvement is claimed - whether it’s relevant to the work I do or not. And whether that claim was tested in a reasonable way.
These products aren’t just expensive - it requires switching your whole workflow. Which is becoming an increasingly big ask in this space.
It’s pretty important for me to be able to understand, and subsequently, believe a benchmark - I find it really hard not to read it as ad copy where this information isn’t present.
> Like some Newcomer men. They don't feel truly masculine
until after they've given birth.
> I'm afraid, George, that giving
birth doesn't quite cut it. You ever see movies? Remember Sylvester Stallone? That beefy fellow with the headband, always had a big gun? Remember that scene in First Blood when Stallone falls off a cliff? He has this huge gash in his arm and he sews himself up. See, that's considered being a man.
> Tell you the truth, Matt, I find his movies simplistic. Why does everything have to be so complicated with you?
Later in the script the extraterrestrial references this in an unintentionally hilarious way, provoking a concerned response from IIRC his wife:
> If I wanted I could fall off
a cliff and sew myself up.
> George, have you had your
lead supplements today?
Aside from its lampshaded effect on popular US conceptions of masculinity in general, the Rambo fantasy seems to have been so popular among, uh, boys who like to cosplay as soldiers, that the knife featured in the movie became the dominant form of cosplay knife for many years, if we believe https://www.youtube.com/watch?v=-n3QiP5LNDE. Some poorly-thought-out regulation here in Argentina has criminalized the possession of knives made to look similar, specifically having a sawblade on the back.
https://tvtropes.org/pmwiki/pmwiki.php/Film/FirstBlood1982 discusses some of the popular literary tropes that appear in it, including "Action Film, Quiet Drama Scene" (which affected the popular perception of Vietnam veterans such as the fictional protagonist); "Affably Evil", in a context that some people think of whenever they hear about a police manhunt on the news; "Asshole Victim", in which the most unpleasant person coincidentally suffers great misfortune; "Break the Haughty", in which the arrogant sheriff turns out to be a coward; "Trauma Button", whose shallow depiction of PTSD was the pattern for the popular understanding of PTSD for many years; and of course "Invincible Hero".
A lot of these are not "near-universal" in the sense of "applicable in nearly every situation", but they are "near-universal" in the sense that everybody has either seen the movie, or seen other movies made by people who were influenced by the movie, or heard stories from people who were influenced by one of those movies, etc.
Some of them are applicable in nearly every situation. Whenever someone thinks that bad things won't happen to them because they're a nice person, for example, they're unconsciously believing in the puddle of ideas around "Asshole Victim", and Rambo's instance is just one drop of blood in that puddle. More insidiously, when people learn that someone has suffered misfortune, "Asshole Victim" subconsciously prompts them to search for reasons they deserved it.
Of course it's easiest for me to identify the thought-patterns that result from tropes I dissent from, not the ones that reflect (as I misunderstand it) Reality.
One of my favourite things a book (or video games, paintings - every single type of art) can do is: On first read - hint that there are stories hidden under the surface, available only to those who have already experienced the art's end point.
I think this is an act of a generosity. It shows that an artist is not only competent enough to execute something that complex, but also thinking deeply of their audience's time and money. It's hospitable: One of more virtuous virtues, IMO.
The best art should reveal itself to you over years, as you change as a human, and your sense of yourself - and the world you live in - changes. I think that should always be the aim of the artist.
You can't swim the same river twice; the same river can't be swum twice by the same person. Some rivers are much better for swimming :-)
> The best art should reveal itself to you over years, as you change as a human, and your sense of yourself - and the world you live in - changes.
Or: a single book that you enjoy different parts of it depending on your life situation is not better at all than if it were several different books instead, one for each life situation. A book for all ages is no better than a kid + an adult book. Maybe in a practical standpoint.
I don't think books are things you can benchmark & create rankings of (fun as it would be to try) which I guess my initial comment implied.
But I think I might disagree - my sister recently had a daughter, and there is so much beauty in the books I (am rarely able to) read to her. My appreciate for Mog's (to pick one example) artistic achievement has only grown by looking at it from a different time in my life.
It's not effective in the same way as it was - I don't find the narrative quite as suspenseful as I remember, for example - but this is sort of exactly what I'm talking about, condensed into a very basic analogy.
If anything, I'm enjoying reading Mog far more! And certainly more than I would enjoy a spin-off 'X-Rated Mog for Adults' series.
I'm being a little facetious, obviously. But I'm going to stick on this one.
> I think this is an act of a generosity. It shows that an artist is not only competent enough to execute something that complex, but also thinking deeply of their audience's time and money.
If the artist is thinking deeply of her audience's time then she should not make the book (or art) worth rereading. Putting things at the beginning that only make sense at the second reading is rude and inconsiderate to the audience.
Well, this is my view of the books or shows that have rewatchability anyway. Fuck them, I won't. Also screw the artist for messing with my time.
Haha. It's lucky that art is subjective, so I'm definitely right :-)
I wasn't saying a piece shouldn't stand on its own on first viewing - that is not hospitable, IMO. I think you misinterpreted what I was trying to say: That I like art that has prismatic qualities, ie, revealing different things when looked at from a different time and place.
This isn't mutually exclusive with the first angle you come at it from being beautiful. It's just additive.
This is sort of a strange reply. You don’t have to spend any time on art at all really. For many people, the more they can spend time enjoying the art they like, the better. If you don’t like the art, that’s one thing, but if you do like it, why must it be shorter?
The video game community is often pretty explicit about this. They want their favorite games to be longer, not shorter, because they want to spend more time enjoying it. I don’t think it’s so strange that people may apply the same mentality, to books, movies, etc as well.
We already kind of have a solution for this with SLAs. Humans, being (probably) non-deterministic, also fuck up. An expectation of a level of service is, I think, reasonable. It's not "zero mistakes ever", just as it can't be "zero bugs ever".
We're firmly in the realms of 'this thing is kind of smarter / faster at a task compared to me my employees, so I am contracting it to do that task'.
That doesn't mean 'if it fails, no payment'.
But I think it's too analogous to non-tech-products to hide behind a 'no refunds' policy. It's that good - there are consequences for it, I think.
I think that sometimes we forget the fact that famous people have regular people brains under their famous skin.
I knew a guy who, hitting fifty or so, with more time on his hands, got really into collecting the wrappers of "Tunnock's Tea Cakes", which are a confectionary product housed in elegant gold packaging. Other people get into horse racing, or MLM scams, or - very, very frequently - history.
It is interesting that Thiel has become more interested in 'the antichrist', given a lot of people would cast him as such. But if I was a billionaire, and my new hobby was religious history - particularly the idea of "the antichrist" - there is absolutely no way I wouldn't be using my money to hold conferences about it and shit like that. I mean, wouldn't you, about your thing, were you a billionaire? He's a gigantic nerd like the rest of us.
Maybe it's an edgy, mid-life-crisis, "I'm so scary" emo pose - in a kind of Hot Topic way. Or maybe Thiel is actually the antichrist, or trying to summon him. Or whatever. But, usually, when a person does this sort of thing, it's usually just because they're interested in it and have the time.
You can take the bad-faith, more exciting view of it, obviously. Or - and this is arguably more embarrassing for Thiel, I think - see it as what it likely is: The hobby a middle-aged man might choose when his obscene wealth meant you would need a really, really, really large amount of Tunnock's Tea Cake wrappers - perhaps the whole company - to scratch the same relative itch.
A real classic. As an author who was asked by my publisher to perform my own audiobook:
1. There is a reason 'reader of audiobooks' is a profession - it is stupid difficult. I will never do it again.
2. I loved this tape so much. It does such interesting things with its soundscape (from memory - if it actually is just Gibson reading it, then he must have embedded those memories through the sheer brilliance of his performance.
3. My fiancee is partially-sighted (I see her as an investment that will appreciate as biohacking becomes more and more prevalent) and she reads mostly by audiobook.
It's not really how I prefer to read - I get distracted too easily - but I've been appalled at the production quality of what I've overheard. While Gibson's work is a special case, an audiobook is only one dimension away from a film adaptation.
4. Literally all my millennial-Gen-Z-cusp friends who are non-readers opted for the audiobook of my book, not the book-book. Anecdata, but interesting. They would just switch Rogan or whatever out during their commute until they felt they'd listened to (what I assume as) enough of it to be socially acceptable.
5. I have no market knowledge other than that I signed my audiobook rights away to my publishers in the industry-standard way.
6. I'm sure it'd be very easy to procure data that made a case for audio fiction that was well-produced and incorporated soundscape-like elements, being incredibly commercially successful. It strikes me as a form that is ripe for innovation. And everyone loves books on tape.
7. There has been so much really interesting innovation in 'aural mood amendment' over the last decade or two. Some of it seems like pseudoscience, some of it seems legit - I wish I had sources to share. Apologies that I don't.
8. I assume someone has already built this concept - well-produced, soundscape-driven longform audio fiction - I'm not a consumer of that market well enough to know it. It'd be a really, really fun project. I'm sure it'd be very tough to get profitable, but it's almost too fun to care. This could be another reason it doesn't exist.
9. Gibson's 2003(?) novel Pattern Recognition is insanely underrated - probably not by people here - but I think the prose is better, and in a decade or two it will feel just as (if not more) prescient. It's a really, really good example of a literary classic that didn't get attention from book dweebs because it's from a 'genre' guy. If you like Neuromancer, and want to think about the next couple of decades in a similar way, you will really love it. I always thought it'd make a great double-bill with the movie Children of Men .
I would second Pattern Recognition. Something about it struck a chord with my similar inclination towards clothing and accessories without massive labels. Never realised it was so such a big thing that would spawn sites like cool hunting, or the current gen-z styling that tends towards all black/all white.
the best author-read audiobook I've listened to recently has been Service Model by Adrian Tchaikovsky[0].
He did a really good job and I would never have guessed it was the author himself. As you say, it's a very different and difficult job.
Absolutely! I listened to it just a couple weeks ago - I was amazed at how good his narration was! He had various voices, accents, great pacing, etc. Tchaikovsky is as good of a voice actor as he is an author. (Actually, maybe a bit better.)
The best author narrated books, by far, are narrated by Douglas Adams. He's recorded all his books, and they're all great. There's something special hearing the words coming straight from the genius himself.
This is cool, but, man, I felt like such a pathetic excuse for a human being when, brutally craving nicotine, with my vape empty of the fruit-flavoured juice that I am literally addicted to like the stupid pathetic baby that I am, and stuck with the cravings because all the shops are closed until morning, and so, in need of a distraction, I opened Hacker News. FFS.
Sometimes the only option is to laugh at your own expense! Clearly this is a sign. I should buy more juice next time. And maybe start smoking more actual cigs.
Quitting smoking was the best decision I made and I wish I did it earlier. If you feel ready Allen Carr's book worked for me. I wasted so much time smoking.
Thanks man, appreciate it, and kudos for stopping.
I might quit, I guess, when my fiancee has to for pregnancy, should it eventually stop being funny to tell her how wonderful each individual cigarette I smoke is. But I really, really love smoking - more than any of my other hobbies, wholesome or otherwise.
I'm sure that'll change, and I'll look that book up if and when it does. Thanks again.
Motivation has to come from you, of course, but you might find it once you see your future baby and realize you have a shot at really seeing how he or she turns out as an adult. That was the impetus for me to start taking my health more seriously.
I've been (I am?) addicted to many substances, from fruit-flavoured nicotine juice through to heroin.
I find self-deprecating humour useful, personally. It helps me not wallow, to take the cravings less seriously. I of course wouldn't say the same about anyone who isn't me.
Because, as you say, someone who has an addiction isn't lesser than anyone else. It's a state of being that requires an awful lot of strength.
That said, having to use that strength on 'mango e-liquid' is, I think, funny in an absurdist way. We live in strange times!
No - that's what the drugs are for. Only joking (for real).
I appreciate that, thank you. And I can totally see how my comment above might read patronising to people dealing with shit that is functionally much, much worse than the situation I described, so I appreciated the opportunity to right that, also.
One thing I'd find frustrating if using this when working on music: "When beat is playing, get new sample for drum plays as soon as it generates".
If "this sample is programmed to play in playing beat", then it'd be good to do a silent replace, or have the option to, so that you don't get a weird unquantized snare sound, for example. I'm finding it puts me off the generated sound more if I initially hear it in a context where it doesn't fit.
Would it actually set any kind of legal precedent, or just establish a sort of cultural vibe baseline? I know Anthropic doesn't have to admit fault, and I don't know if that establishes anything in either direction. But I'm not from the US, so I wouldn't want to pretend to have intimate knowledge of its system.
The number of bizarre, contradictory inferences this settlement asks you to make - no matter your stance on the wider question - is wild.
The settlement doesn't set any kind of precedent at all.
The existing ruling in the case establish "persuasive" (i.e. future cases are entirely free to disagree and rule to the contrary) precedent - notably including the part about training on legally acquired copies of books (e.g. from a book store) being fair use.
Only appeals courts establish binding precedent in the US (and only for the courts under them). A result of this case settling is that it won't be appealed, and thus won't establish any binding precedent one way or another.
> The number of bizarre, contradictory inferences this settlement asks you to make - no matter your stance on the wider question - is wild.
> What contradictions do you see? I don't see any.
I guess us seeing very different things is also what a settlement might be for :-).
But I think I was wrong.
I think others in the thread are debating the contradictions I saw. I tried typing them out when I made my earlier comment, but couldn't get them to fix to any kind of logic that made sense to me. They just seemed contradictory, at the time.
I think the same arguments have now been made much more clearly by others - specifically around whether a corporation downloading this work is the same as a human downloading it - and the responses have been very clear also.
The settlement figure was tied implicitly to Anthropic's valuation in the Ars article [0] where I think I originally posted my comment. Those comments were moved here, so I've linked below.
Specifically linking the settlement sum to the valuation of a corporation is what caught me in a loop - that valuation assumes that Anthropic will do certain things in the future. I was thinking too much, maybe, about things like:
"Would a teenager get the same treatment? What about a teenager with a private company? What about a teenager who seemed dumber than that teenager to the person deciding their company's valuation? What about a teenager who had not opened the files themselves, but had spun up a model from them? What about a teenager who had done both?"
Etc. I think I was getting fixated on the idea that the valuation assumes future performance, and downloading the files was possibly necessary for that performance, but I was missing the obvious answers to some of my questions because of that.
I do think that some of the more anthropomorphising language - "training data" is an example - trips people up a lot in the same way. And I think that if the settlement sum reflects anything to do with the valuation of that corporation, that does create some interesting questions, but maybe not contradictions.
Hmm my huge concern was that if the settlement were to get approved, it would set a legal precedent for other "settlement approvals" like this one, setting back AI research in the US, paving way for China to win the race.
My learning so far, to your point on memory being a limiting factor, is that the system is able to build on ideas over time. I'm not sure you'd classify that as 'self-learning', and I haven't really pushed it in the direction of 'introspection' at all.
Memory itself (in this form) does not seem to be a silver bullet, though, by any means. However, as I add more 'tools', or 'agents', its ability to make 'leaps of discovery' does improve.
For example, I've been (very cautiously) allowing cron jobs to review a day's conversation, then spawn headless Claude Code instances to explore ideas or produce research on topics that I've been thinking about in the chat history.
That's not much different from the 'regular tasks' that Perplexity (and I think OpenAI) offer, but it definitely feels more like a singular entity. It's absolutely limited by how smart the conversation history is, at this time, though.
The Memento analogy you used does feel quite apt - there is a distinct sense of personhood available to something with memory that is inherently unavailable to a fresh context window.