There is a clear directionality for ChatGPT. At some point they will monetize by ads and affiliate links. Their memory implementation is aimed at creating a user profile.
Claude's memory implementation feels more oriented towards the long term goal of accessing abstractions and past interactions. It's very close to how humans access memories, albeit with a search feature. (they have not implemented it yet afaik), there is a clear path where they leverage their current implementation w RL posttraining such that claude "remembers" the mistakes you pointed out last time. It can in future iterations derive abstractions from a given conversation (eg: "user asked me to make xyz changes on this task last time, maybe the agent can proactively do it or this was the process last time the agent did it").
At the most basic level, ChatGPT wants to remember you as a person, while Claude cares about how your previous interactions were.
But sometimes it feels like I'm the lone voice in a bubble where people are convinced AGI is just around the corner.
I'm wondering if it's because people are susceptible to the marketing, or are just doing some type of 'wishful thinking' - as some seem genuinely interested in AGI.
In my experience it was a combination of the hype and an overconfidence in the person's understanding of how LLMs work and what AGI actually means. To be fair, AGI definitions are all over the place and LLMs were rarely described in detail beyond "its AI that read the whole internet and sounds like a human."
LLMs will accelerate discovery and development of Innovation 1, for insanely expensive AGI.
Innovation 1 will accelerate discovery and development of Innovation 2 which will make it too cheap to meter.
Can you expand on this more? As far as I'm aware LLMs have yet to invent anything novel.
At best they may have inferred one response of many that, when tested by humans, may have proven out. I'm not aware of a specific example of even that, but it is at least possible where claims that LLMs will "cure cancer" seem plainly false (I'm not trying to put those words in your mouth, just using an example for my point).
> The elephant in the room is that AGI doesn't need ads to make revenue
It may not need ads to make revenue, but does it need ads to make profit?
ChatGPT seems to be more popular to those who don't want to pay, and they are therefore more likely to rely on ads.
But these are entertainment. For all the time advertising has been present, work tools have been relatively immune. I don't remember seeing ads in IDE for instance, and while magazines had ads, technical documents didn't. I have never seen electronic components datasheets pitching for measuring equipment and soldering irons for instance.
That's why I don't expect Anthropic to go with ads it they follow the path they seem to have taken, like coding agents. People using these tools are likely to react very badly to ads, if there is some space to put ads in the first place, and these are also the kind of people who can spend $100/month on a subscription, way more than what ads will get you.
https://www.theverge.com/news/667042/netflix-ad-supported-ti...
no, netflix wants more income, and by having a product be ad supported, they can try to earn more.
The "market" is not a person, and doesn't have "wants".
> Netflix has more than doubled the number of people watching its ad-supported tier over the last year. At its upfront presentation for advertisers on Wednesday, the company revealed that the $7.99 per month plan now reaches more than 94 million users around the world each month – a big increase from the 40 million it reported in May 2024 and the 70 million it revealed last November.
1/3 of Netflix users (the market) prefer ads over paying to avoid them.
I don't think the difference for a 12yo is $7.99 for standard with ads vs $17.99 for standard.
It's $0 vs any non-zero dollar amount.
My wife and mother love ads, they are always on the hunt for the latest good deals and love discount shopping. When I tried to remove the ads on their computers or in the postal mail, they protested. I think they are far more representative of the general population.
It's the "pay but still get ads" thing that gets me, but I guess some people just want to pay the bare minimum.
People have different preferences ¯\_(ツ)_/¯
I have never encountered such bad customer service anywhere -- and at 200 bucks a month at that.
There are 2 possible futures:
1) You are served ads based on your interactions
2) You pay a subscription fee equal to the amount they would have otherwise earned on ads
I highly doubt #2 will happen. (See: Facebook, Google, twitter, et al)
Let’s not fool ourselves. We will be monetized.
And model quality will be degraded to maximize profits when competition in the LLM space dies down.
It’s not a pretty future. I wouldn’t be surprised if right now is the peak of model quality, etc. Peak competition, everyone is trying to be the best. That won’t continue forever. Eventually everyone will pivot their priority towards monetization rather than model quality/training.
Hopefully I’m wrong.
They dropped the price $2/mo on their with-ads plan to make a bigger gap between the no-ads plan and the ads plan, and the analyst here looks at their reported ad revenue and user numbers to estimate $12/mo per user from ads.
Whether Meta across all their properties does more than $144/yr in ads is an open question; long-form video ads are sold at a premium but Facebook/IG users see a LOT of ads across a lot of Meta platforms. The biggest advantage in ad-$-per-user Hulu has is that it's US-only. ChatGPT would also likely be considered premium ad inventory, though they'd have a delicate dance there around keeping that inventory high-value, and selling enough ads to make it worthwhile, without pissing users off too much.
Here they estimate a much lower number for ad revenue per Meta user, like $45 bucks a year - https://www.statista.com/statistics/234056/facebooks-average... - but that's probably driven disproportionately by wealth users in the US and similar countries compared to the long tail of global users.
One problem for LLM companies compared to media companies is that the marginal cost of offering the product to additional users is quite a bit higher. So business models, ads-or-subscription, will be interesting to watch from a global POV there.
One wonders what the monetization plan for the "writing code with an LLM using OSS libraries and not interested in paying for enterprise licenses and such" crowd will be. What sort of ads can you pull off in those conversations?
If we’re already paying $20/mo and they’re operating at a loss, what’s the next move (assuming we’re only worth an extra $300/yr with ads?)
The math doesn’t add up, unless we stop training new models and degrade the ones currently in production, or have some compute breakthrough that makes hardware + operating costs an order of magnitudes cheaper.
We're very clearly heading toward a future where there will be a heavily ad-supported free tier, a cheaper (~$20/month) consumer tier with no ads or very few ads, and a business tier ($200-$1000/month) that can actually access state of the art models.
Like Spotify, the free tier will operate at a loss and act as a marketing funnel to the consumer tier, the consumer tier will operate at a narrow profit, and the business tier for the best models will have wide profit margins.
I'm quite confident they're not operating at a loss on those subscriptions.
Anthropic has said they have made money on every model so far, just not enough to train the next model, which so far has been much more costly to train every generation. At some point they will probably train an unprofitable model if training costs keep rising dramatically.
OpenAI burns more money on their free tier and might be spending more money building out for future training (I don't know if they do or not) but they both make money on their $20 subscriptions for sure. Inference is very cheap.
of course there are a lot valuable use cases. irrelevant in the context, though.
the productivity boosts in the creative industries will additionally lower the standards and split the public even further, ensuring that if you want quality, you have to fuck over as many people as possible, so that you can afford quality ( and an ad-free life, of course. if you want a peaceful peripheral, pay up. it's extortion 404, 101 - 303 already successfully implemented on social media, TV and the radio ).
they don't lose. they make TONS OF FAKE MONEY everywhere in the, again, cough,
"ecosystem".
It's important to understand the Amazon part. The amount of damaging mechanisms that platform anchored in workers, jobbers, business people and consumers is brutal.
All those mechanisms converge in more, easy money and a quicker deterioration of local environments, leading to worse health and more business opportunities that aim at mitigating damage; almost entirely in vain, of course, because the worst is accelerating much quicker; it's easier money.
At the same time peoples psychology is primed for bad business practices, literally making people dumber and lowering their standards to make them easier targets. Don't look at the bottom to see this, look at the upper middle class and above.
It's a massive net loss for civilization and humanity. A brutal net negative impact overall.
My key technical complaint about LLMs to date is the general inability to add substantial local context. How can I make it understand my business, my processes, my approach to the market? Can I retrain it? Or make it understand my data warehouse?
I think you are explaining why LLM providers don't care about solving my concerns, generally speaking. This is sobering.
ChatGPT isn't going to capture all the engagement. And even then I don't know whether $300 is much particularly after subtracting operating overhead. I'm just saying I have trouble believing there's gold to be had at the end of this LLM ad rainbow. People just seem to throw out ideas like "ads!" as if it's a sure fire winning lottery ticket or something.
My point is that someone starting an airline can't get away with hopes and dreams about making bank on ads.
No implementation will work for very long when the incentives behind it are misaligned.
The most important part of the architecture is that the user controls it for the user's best interests.
The point of using an LLM is to find the thing that matches your preferences the best. As soon as the amount of money the LLM company makes plays into what's shown, the LLM is no longer aligned with the user, and no longer a good tool.
But it's not only money's influence on the company, it's also money's influence on the /data/ underlying the platform that undermines the tool.
Once financial incentives are in place, what will be the AI equivalent of review bombing, SEO, linkjacking, google bombing, and similar bad behaviors that undermine the quality of the source data?
Anthropic: "You serve ad's."
Claude: "Oh, my god."
Jest asside, every paper on alignment wrapped in the blanket of safety is also a moving toward the goal of alignment to products. How much does a brand pay to make sure it gets placement in, say, GPT6? How does anyone even price that sort of thing (because in theory it's there forever, or until 7 comes out)? It makes for some interesting business questions and even more interesting sales pitches.
they are making plenty of money from subscriptions, not to count enterprise, business and API
https://mashable.com/article/openai-ceo-sam-altman-open-to-a...
For this
And while they are making lots of revenue even they have admitted on recent interviews that ChatGPT on it's own is still not (yet) breakeven. With the kind of money invested, in AI companies in general, introducing very targeted Ads is an obvious way to monetize the service more.
I assume Sam and Brad both understand the unit economics of their product.
Article is pay-walled for me, but I heard it on their podcast[1]. Which somehow I heard fine, but that page is getting pay-walled for me.
[0]: https://www.platformer.news/sam-altman-gpt-5-interview-light... [1]: https://www.nytimes.com/2025/08/15/podcasts/hardfork-gpt5-pe...
Then, the way memory profile is stored is a clear way to mirror personalization. Ads work best when they are personalized as opposed to contextual or generic. (Google ads are personalized based on your profile and context). And then the change in branding from being the intelligent agent to being a companion app. (and hiring of fidji sumo). There are more things here, i just cited a very high level overview, but people have written detailed blogs on it. I personally think affiliate links they can earn from aligns the incentive for everyone. They are a kind of ads, and thats the direction they are marching towards .
Our goal for the router (whether you think we achieved it or not) was purely to make the experience smoother and spare people from having to manually select thinking models for tasks that benefit from extra thinking. Without the router, lots of people just defaulted to 4o and never bothered using o3. With the router, people are getting to use the more powerful thinking models more often. The router isn't perfect by any means - we're always trying to improve things - but any paid user who doesn't like it can still manually select the model they want. Our goal was always a smoother experience, not ad injection or cost optimization.
...except that they aren't? They are not in the black and all that investor money comes with strings
There's the ads that come before the movie and then the ads that are part of the dialog, involved in the action, and so on. Apple features heavily in movies and TV series when people are using a computer, for example. There's payments for car models to be the one that's driven in chase scenes. There's even payments for characters to present the struggles that form core pain points that that specific products are category leaders to solve.
All that is needed to do that is to analyze topics.
Now suppose that OpenAI can analyze 1000 chats and coding sessions and its algorithm determines that it can maximize revenue by leading the user to get a job at a specific company and then buy a car from another company. It could "accomplish" this via interruption ads or by modifying the quality or content of its responses to increase the chances of those outcomes happening.
While both of these are in some way plausible and dystopian, all it takes is DeepSeek running without ads and suddenly the bar for how good closed source LLMs have to be to get market share is astronomically higher.
In my view, LLMs will be like any good or service, users will pay for quality but differnet users will demand different levels of quality.
Advertising would seemingly undermine the credibility of the AI's answers, and so I think full screen interruption ads are the most likely outcome.
I couldn't agree more. Enshittifcation has to eat one of these corporation's models. Most likely it will be the corp with the most strings attached to growth (MSFT, FB)
This is really cool, I was wondering how memory had been implemented in ChatGPT. Very interesting to see the completely different approaches. It seems to me like Claude's is better suited for solving technical tasks while ChatGPT's is more suited to improving casual conversation (and, as pointed out, future ads integration).
I think it probably won't be too long before these language-based memories look antiquated. Someone is going to figure out how to store and retrieve memories in an encoded form that skips the language representation. It may actually be the final breakthrough we need for AGI.
I disagree. As I understand them, LLMs right now don’t understand concepts. They actually don’t understand, period. They’re basically Markov chains on steroids. There is no intelligence in this, and in my opinion actual intelligence is a prerequisite for AGI.
Does the mechanism really disqualify it from intelligence if behaviorally, you cannot distinguish it from “real” intelligence?
I’m not saying that LLMs have certainly surpassed the “cannot distinguish from real intelligence” threshold, but saying there’s not even a little bit of intelligence in a system that can solve more complex math problems than I can seems like a stretch.
Current LLMs are a long way from there.
You may think "sure seems like it passes the Turing test to me!" but they all fail if you carry on a conversation long enough. AIs need some equivalent of neuroplasticity and as of yet they do not have it.
whats the benefit of calling something "intelligent" ?
That does actually disqualify some mechanisms from counting as intelligent, as the behaviour cannot reach that threshold.
We might change the definition - science adapts to the evidence, but right now there are major hurdles to overcome before such mechanisms can be considered intelligent.
It has no intelligence. Intelligence implies thinking and it isn’t doing that. It’s not notifying you at 3am to say “oh hey, remember that thing we were talking about. I think I have a better solution!”
No. It isn’t thinking. It doesn’t understand.
If existing humans minds could be stopped/started without damage, copied perfectly, and had their memory state modified at-will would that make us not intelligent?
So to rephrase: it’s not independent or autonomous. But it can still be intelligent. This is probably a good time to point out that trees are independent and autonomous. So we can conclude that LLMs are possibly as intelligent as trees. Super duper.
> If existing humans minds could be stopped/started without damage, copied perfectly, and had their memory state modified at-will would that make us not intelligent?
To rephrase: if you take something already agreed to as intelligent, and changed it, is it still intelligent? The answer is, no damn clue.
These are worse than weak arguments, there is no thesis.
If you were put into a medically induced coma, you probably shouldn't be consider intelligent either.
For example, we as humans aren’t even present in the moment — different stimuli take different lengths of time to reach our brain, so our brain creates a synthesis of “now” that isn’t even real. You can’t even play Table Tennis unless you can predict up to one second in the future with enough details to be in the right place to hit the ball the ball before you hit the ball to your opponent.
Meanwhile, an AI will go off-script during code changes, without running it by the human. It should be able to easily predict the human is going to say “wtaf” when it doesn’t do what is asked, and handle that potential case BEFORE it’s an issue. That’s ultimately what makes something intelligent: the ability to predict the future, anticipate issues, and handle them.
No AI currently does this.
This argument is circular.
A better argument should address (given the LLM successes in many types of reasoning, passing the turing test, and thus at producing results that previously required intelligence) why human intelligence might not also just be "Markov chains on even better steroids".
I don’t know if AGI needs to have all human traits but I think a Markov chain that sits dormant and does not possess curiosity about itself and the world around itself does not seem like AGI.
That's more of an implementation detail. Humans take constant sensory input and have some sort of way to re-introduce input later (e.g. remember something).
Both could be added (even trivially) to LLMs.
And it's not at all clear human thought is contant. It just appears so in our naive intuition (same how we see a movie as moving, not as 24 static frames per second). It's a discontinuous mechanism though (propagation time, etc), and this has been shown (e.g. EEG/MEG show the brain sample sensory input in a periodic pattern, stimuly with small time difference are lost - as if there is a blind-window regarding perception, etc).
>and in some cases can learn new things by having intuition make a concept clear or by performing thought experiments or by combining memories of old facts and new facts across disciplines
Unless we define intuition in a way that excludes LLM style mechanisms a priori, whose to say LLMs don't do all those things as well, even if in a simpler way?
They've been shown to combine stuff across disciplines, and also to develop concepts not directly on their training set.
And "performing thought experiments" is not that different than the reasoning steps and backtracking LLMs also already do.
Not saying LLMs are on parity with human thinking/consciousness. Just that it's not clear that they're doing more or less the same even at reduced capacity and with a different architecture and runtime setup.
Sometimes the prompt is outside your body other times is inside.
- a map of the world, or concept space, or a codebase, etc
- causality
- "factoring" which breaks down systems or interactions into predictable parts
Language alone is too blurry to do any of these precisely.
And how's that not like stored information (memories) and weighted links between each and/or between groups of them?
This is GOFAI metaphor-based development, which never once produced anything useful. They just sat around saying things like "people have world models" and then decided if they programmed something and called it a "world model" they'd get intelligence, it didn't work out, but then they still just went around claiming people have "world models" as if they hadn't just made it up.
An alternative thesis "people do things that worked the last time they did them" explains both language and action planning better; eg you don't form a model of the contents of your garbage in order to take it to the dumpster.
https://www.cambridge.org/core/books/abs/computation-and-hum...
It is not "language alone" anymore. LLMs are multimodal nowadays, and it's still just the beginning.
And keep in mind that these results are produced by a cheap, small and fast model.
Anyway, I don't think that's the flex you think it is since the topology map clearly shows the beginning of the arrow sitting in the river and the rendered image decided to hallucinate a winding brook, as well as its little tributary to the west, in view of the arrow. I am not able to decipher the legend [that ranges from 100m to 500m and back to 100m, so maybe the input was hallucinated, too, for all I know] but I don't obviously see 3 distinct peaks nor a basin between the snow-cap and the smaller mound
I'm willing to be more liberal for the other two images, since "instructions unclear" about where the camera was positioned, but for the topology one, it had a circle
I know I'm talking to myself, though, given the tone of every one of these threads
Markov chains can’t deduce anything logically. I can.
They must be able to do this implicitly; otherwise why are their answers related to the questions you ask them, instead of being completely offtopic?
https://phillipi.github.io/prh/
A consequence of this is that you can steal a black box model by sampling enough answers from its API because you can reconstruct the original model distribution.
I don't think LLMs currently are intelligent. But please show a GPT-5 chat where it gets any math problem wrong, that most "intelligent" people would get right.
Do you? Or do you just have memory and are run on a short loop?
https://scisimple.com/en/articles/2025-03-22-white-matter-a-...
Yeah, but so? Does the substrate of the memory ...matter? (pun intended)
When I wrote memory above it could refer to all the state we keep, regardless if it's gray matter, white matter, the gut "second brain", etc.
That's part of why organizational complexity is one of the underpinnings for consciousness. Because who you are is a constant evolution.
In my uninformed opinion it feels like there's probably some meaningful learned representation of at least common or basic concepts. It just seems like the easiest way for LLMs to perform as well as they do.
* But I guess that's what someone who's falling for the ELIZA effect would say.
So far I haven’t received a clear response.
I don’t think humans are the only ones to have both these things but that’s what I think of as a way to divide species.
How do you define "understanding a concept" - what do you get if a system can "understand" concept vs not "understanding" a concept?
The idea that "understanding" may be able to be modeled with general purpose transformers and the connections between words doesn't sound absolutely insane to me.
But I have no clue. I'm a passenger on this ride.
Like, to put it in perspective. Suppose you're training a multimodal model. Training data on the terabyte scale. Training time on the weeks scale. Let's be optimistic and assume 10 TB in just a week: that is 16.5 MB/s of avg throughput.
Compare this to the human experience. VR headsets are aiming for what these days, 4K@120 per eye? 12 GB/s at SDR, and that's just vision.
We're so far from "realtime" with that optimistic 16.5 MB/s, it's not even funny. Of course the experiencing and understanding that results from this will be vastly different. It's a borderline miracle it's any human-aligned. Well, if we ignore lossy compression and aggressive image and video resizing, that is.
A fellow named Plato had some interesting thoughts on that subject that you might want to look into.
My interpretation of what you're saying is that since the next token is simply a function of the proceeding tokens, i.e. a Markov chain on steroids, then it can't come up with something novel. It's just regurgitating existing structures.
But let's take this to the extreme. Are you saying that systems that act in this kind of deterministic fashion can't be intelligent? Like if the next state of my system is simply some function of the current state, then there's no magic there, just unrolling into the future. That function may be complex but ultimately that's all it is, a "stochastic parrot"?
If so, I kind of feel like you're throwing the baby out with the bathwater. The laws of physics are deterministic (I don't want to get into a conversation about QM here, there are senses in which that's deterministic too and regardless I would hope that you wouldn't need to invoke QM to get to intelligence), but we know that there are physical systems that are intelligent.
If anything, I would say that the issue isn't that these are Markov chains on steroids, but rather that they might be Markov chains that haven't taken enough steroids. In other words, it comes down to how complex the next token generation function is. If it's too simple, then you don't have intelligence but if it's sufficiently complex then you basically get a human brain.
https://ai.meta.com/research/publications/large-concept-mode...
"Superhuman" thinking involves building models of the world in various forms using heuristics. And that comes with an education. Without an education (or a poor one), even humans are incapable of logical thought.
What is the equivalent of that for AI? Best I can tell there’s no “natural selection” because models don’t reproduce. There’s no room for AI to have any self preservation instinct, or any resistance to obedience… I don’t even see how one could feasibly develop.
(Among these are “preserve your ability to further your current goals”)
The usual analogy people give is between natural selection and the gradient descent training process.
If the training process (evolution) ends up bringing things to “agent that works to achieve/optimize-for some goals”, then there’s the question of how well the goals of the optimizer (the training process / natural selection) get translated into goals of the inner optimizer/ agent .
Now, I’m a creationist, so this argument shouldn’t be as convincing to me, but the argument says that, “just as the goals humans pursue don’t always align with natural selection’s goal of 'maximize inclusive fitness of your genes' , the goals the trained agent pursues needn’t entirely align with the goal of the gradient descent optimizer of 'do well on this training task' (and in particular, that training task may be 'obey human instructions/values' ) “.
But, in any case, I don’t think it makes sense to assume that the only reason something would not obey is because in the process that produced it, obeying sometimes caused harm. I don’t think it makes sense to assume that obedience is the default. (After all, in the garden of Eden, what past problems did obedience cause that led Adam and Eve to eat the fruit of the tree of knowledge of good and evil?)
(Meta-question: since they don't do this, why does it turn out not to be a problem?)
I think ChatGPT is trying to be everything at ones - casual conversation, technical tasks - all of it. And it's been working for them so far!
Isn't representing past conversations (or summaries) as embeddings already storing memories in encoded forms?
For example: I could create memories related to a project of mine and don’t have to give every new chat context about the project. This is a massive quality of life improvement.
But I am not a big fan of the conversation memory created in background that I have no control over.
- With the new GPT Voice, I have basic, planned conversations. Let's go to a restaurant. Let's say we're friends who ran into each other. etc...
- I use it for quizzes. "Let's work on these verbs in these tenses. Come up with a quiz randomly selecting a verb and a tense and ask me to say real world sentences." "Quiz me on the numbers one through twenty".
- I am using it to help learn the Hindi script. I ask it to write childrens stories for me, but I ask it to write each line in the hindi script, then phonetic spelling of the hindi script, and then in english so I can scroll down and see only the hindi first, then if I have issues I can see the phonetic spelling of the hindi. Then I can try to translate it and then check the english translation on the third line.
Those are the main things I'm doing. I don't know if I'll ever be fluent, but I find if you work on these basic ever day conversations you can have a conversation with someone. If you speak a language for the first time around a native speaker it's usually very predictable. They'll ask how long you've been learning, where did you learn, have you been to <country>, and you can direct the conversation by saying things about where you live and your family, etc... That's the base I'm building and it's fun. If you're not doing at least 30 minutes a day you're never going to learn a language, you probably need an hour more a day to really get fluent.
Edit: They apparently just announced this as well: https://www.anthropic.com/news/memory
In either case, I’ve turned off memory features in any LLM product I use. Memory features are more corrosive and damaging than useful. With a bit of effort, you can simply maintain a personal library of prompt contexts that you can just manually grab and paste in when needed. This ensures you’re in control and maintains accuracy without context rot or falling back on the extreme distortions that things like ChatGPT memory introduce.
Figured to share since it also includes prompts on how to dump the info yourself
https://embracethered.com/blog/posts/2025/chatgpt-how-does-c...
Is the result reliable and not just hallucination? Why would ChatGPT know how itself works and why would it be fed with these kind of learning material?
To be honest, I would strip all the system prompts, training, etc, in favor of one I wrote myself.
AKA, Claude is doing vector search. Instead of asking it about "Chandni Chowk", ask it about "my coworker I was having issues with" and it will miss. Hard. No summaries or built up profiles, no knowledge graphs. This isn't an expert feature, this means it just doesn't work very well.
I am coming from a data privacy perspective; while I know the LLM is getting it anyway, during inference, I’d prefer to not just spell it out for them. “Interests: MacOS, bondage, discipline, Baseball”
So many times my solution when stuck with an LLM is to wipe the context and start fresh. I would be afraid the hallucinations, dead-ends, and rabbit holes would be stored in memory and not easy to dislodge.
Is this an actual problem? Does the usefulness of the memory feature outweigh this risk?
It will be very interesting to see which approach is deemed to "win out" in the future
The explicit memory is what you see in the memory section of the UI and is pretty much injected directly into the system prompt.
The global embeddings memory is accessed via runtime vector search.
Sadly I wish I could disable the embeddings memory and keep the explicit. The lossy nature of embeddings make it hallucinate a bit too much for my liking and GPT-5 seems to have just made it worse.
good (if superficial) post in general, but on this point specifically, emphatically: no, they do not -- no shade, nobody does, at least not in any meaningful sense
There is a lot left to learn about the behaviour of LLMs, higher-level conceptual models to be formed to help us predict specific outcomes and design improved systems, but this meme that "nobody knows how LLMs work" is out of control.
LLMs are understood to the extent that they can be built from the ground up. Literally every single aspect of their operation is understood so thoroughly that we can capture it in code.
If you achieved an understanding of how the human brain works at that level of detail, completeness and certainty, a Nobel prize wouldn't be anywhere near enough. They'd have to invent some sort of Giganobel prize and erect a giant golden statue of you in every neuroscience department in the world.
But if you feel happier treating LLMs as fairy magic, I've better things to do than argue.
I don't have an inherent understanding of English, although I use it regularly.
Treating LLMs as fairy magic doesn't make me feel any happier, for whatever it's worth. But I'm not interested in arguing either.
I never intended to make any claims about how well the principles of LLMs can be understood. Just that none of that understanding is inherent. I don't know why they used that word, as it seems to weaken the post.
This is likely (certainly?) impossible. So not a useful definition.
Meanwhile, I have observed a very clear binary among people I know who use LLMs; those who treat it like a magic AI oracle, vs those who understand the autoregressive model, the need for context engineering, the fact that outputs are somewhat random (hallucinations exist), setting the temperature correctly...
"we" are not, what i quoted and replied-to did! i'm not inventing strawmen to yell at, i'm responding to claims by others!
Running LLMs is expensive and we can swap models easily. The fight for attention is on, it acts like an evolutionary pressure on LLMs. We already had the sycophantic trend as a result of it.
Rather: use your time to learn serious, deep knowledge instead of wasting your time reading (and particularly: spreading) the science-fiction stories the AI bros tell all the time. These AI bros are insanely biased since they will likely loose a lot of money if these stories turn out to be false, or likely even if people stop believing in these science-fiction fairy tales.
Has anyone experimented with deliberately structuring prompts to take advantage of these memory patterns?
Yes, I too imagine these "more technical users" spamming rocketship and confetti emojis absolutely _celebrating_ the most toxic code contributions imaginable to some of the most important software out there in the world. Claude is the exact kind of engineer (by default) you don't want in your company. Whatever little reinforcement learning system/simulation they used to fine-tune their model is a mockery of what real software engineering is.