i guess the thing i'm most confused about is what is the higher level goal here. 1 in 8 humans on the planet are experiencing the "web" through chatgpt alone. many have migrated to purely agentic workflows.
is the goal for your content to just be invisible to this growing population? is the expectation that all of this is just a fad, which will fade away? what is the end game to the tactics you have outlined? what is the strategy?
I like those ideas. If people are truly operating purely through LLM's then I am fine with being partitioned from them. The LLM operators will provide their reality and their truth. This is all for low trust internet-wide access.
For smaller higher trust communities I don't do any of this. Rather we use basic authentication to keep bots and strangers off the services all together and then regular user accounts on forums and such. That is where strategy comes into play and has been working well for some time.
i guess what i'm curious about is where you would draw the line, and why.
how do you define a user agent? is it desirable or not for users to be able to discover and access resources and communities on the internet using the tools, formats, and workflows that they prefer?
are search bots desirable?
the internet wayback machine?
how do you feel about browser extensions and greasemonkey scripts?
I do not define user-agents, they announce themselves and present a particular behavior. I let them define their behavior and I respond accordingly.
I have no need for search bots personally. If I had a complex site I would build my own search feature. If it was a commercial site I would pay for advertising on popular sites that were of related categories.
Wayback is fun to play with but it's just a toy to me. There is no concept of domain ownership. Most of the snapshots of the domain I am using were from when someone else rented it. They claim to respect robots.txt but that is not entirely true. They crawl and save content even if a site says not to and as soon as the site is offline and robots.txt is no longer accessible they will display all the archived content. This means a person has to park their domain on a server containing a robots.txt that matches their intentions.
Browser extensions are mostly invisible. What is obvious to me is when people are using "reader" applications as web clients. They should be concerned more than me. They are reading random sites run by strangers using apps that may or may not have been battle hardened, reviewed by third party penetration testers and so on. Some of them are vibe coded in unsafe languages.
I am an active user of Claude but I do not live my life through that thing. I try to get it to answer questions that it does not want to answer. The only questions it seems to answer without hesitation are technical in nature. In the entire existence of my account I have started 9 chats.
More interesting to me was that Alexa on Amazon Prime answered all my questions without hesitation or disclaimers but once I have my answers I close that tab.
This is definitely a 2026 type comment, but I'm very surprised that it's so "low." GPT is the most well known brand and it's the one that seemingly 100% of kids use to "assist" on their homework and other exercises from grade school to college. One would think that alone would already take you well over a billion.
I am monthly active user, absolutely do not experience web through ChatGPT alone and frankly, dont mind ChatGPT being cut off stuff. AI companies forcing themselves as a middle man is not a good thing.
Whoever using purely agentic worflow is not my concern. Just like people who read only physical book, learn only from podcasts or watch only netflix movies.
There is zero reason for the rest of us to worry about filling free data to companies that try to make themselves monopolies.
I wonder what the click through rate is for the source links. Based on what I've been hearing, it's very low. And that's if the particular LLM/interface/answer even bothers to add them.
If the bots are eating your site traffic and users, what's the point of allowing them in?
> One thing that should be learned from the bitter lesson is the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great. The two methods that seem to scale arbitrarily in this way are search and learning.
Right, and if you come up with an efficiency gain that makes scaling better, e.g. a 50% reduction in required compute. Or even asymptotic improvements e.g. moving from quadratic to linear. Then you're much much better off.
There is nothing about the bitter lesson that says just be dumb and pour money into a hole, you still have to invent the methods to scale well, and being under immense pressure with constraints seems likely to produce that research.
It reminds me a bit of the tyranny of the rocket equation. You can always scale your fuel to get a little more delta V, with ever diminishing returns. …but for something like a DEEP space/interstellar mission, it almost always pays to wait a few more years for a faster propulsion system because you’ll get there fastest by always delaying your launch and chasing better technology.
I’m not sure how well the analogy holds up, or if there’s anything to be learned from it though.
Certainly applies more general imho. Constrained by some resource -> invest resources elsewhere, and/or invest in reducing the constraint(s) encountered.
The implication here is that the only gains left to be had are from scale. That we are already maximally efficient. If that's true, then how has OpenAI repeatedly bragged about reducing the cost of their models by orders of magnitude? (And DeepSeek Flash even more so, of course.)
But we have not been maximally efficient, we keep gaining efficiency. If we keep gaining efficiency, why should we assume it is impossible to gain more?
this is quite literally reward hacking. the model, under evaluation with cyber capabilities enabled, used those capabilities to simply bypass the exercise entirely and aim straight for the source of the flag. the CTF equivalent back in the day would be hacking the scoreboard.
in a street fight, the only rules are that there are no rules.
This line of thinking makes no sense because it assumes that labs that distill from frontier models are doing nothing else. It's the classic "the Chinese can only copy" mentality, and it's going to end poorly for American companies.
I'm pretty sure that all labs are distilling each others' LLMs, maybe apart from Anthropic and OpenAI. It would be stupid not to do it, because it's cheap and effective. But that's not the only thing they're doing. If you think K3 and GLM-5.2 got this good only from distilling frontier models, you're not paying attention to Chinese labs' publications.
i never assumed that, and i do keep up with the publications. i'm also not saying it's a dumb thing to do! what i am saying is that empirically, it appears that distillation of a more advanced model is a required first step for them to train a borderline competitive, cheaper model. in effect, their training is subsidized by the frontier labs.
if this were not the case, then we would be observing chinese models that far surpass frontier models in capabilities, rather than "almost as good, but much cheaper", and we would be having a very different conversation. what happens to these efforts when the subsidy is cut off?
> empirically, it appears that distillation of a more advanced model is a required first step
I see no evidence for that.
> if this were not the case, then we would be observing chinese models that far surpass frontier models
It's pretty clear that the primary reason for the difference is budget and compute availability. Chinese labs have at least an order of magnitude less money than Anthropic and OpenAI.
> what happens to these efforts when the subsidy is cut off?
They will continue making progress as they do now, minus the benefits of distillation.
Agentic reasoning and tool use
Coding and data analysis
Computer-use agent development
Computer vision
Moonshot (Kimi models) employed hundreds of fraudulent accounts spanning multiple access pathways. Varied account types made the campaign harder to detect as a coordinated operation. We attributed the campaign through request metadata, which matched the public profiles of senior Moonshot staff. In a later phase, Moonshot used a more targeted approach, attempting to extract and reconstruct Claude’s reasoning traces.
I'm assuming you posted that as evidence for the claim that "empirically, it appears that distillation of a more advanced model is a required first step", but I don't think it is. It's just evidence that Moonshot distills Anthropic's models, which, yes, they do.
it is not a required first step for training a model, sure. but that's not what i claimed. what i claimed is that is how they are so significantly _reducing the cost_ of training one! how else do you think they are doing it?
Distillation from a teacher model solves the self-start problem, that is, building a model to the point where it reason coherently. Without distillation, solving self-start is incredibly difficult since it requires millions of high quality training samples. Creating that kind of dataset takes an enormous amount of effort.
Once a model becomes competent enough to perform complex reasoning, a teacher model is no longer necessary. The model can now reason about its own behavior and build a better version of itself through recursive self-improvement (RSI).
> Creating that kind of dataset takes an enormous amount of effort
Can you imagine the amount of effort it takes to write 15 trillion tokens worth of art, literature, source code, textbooks, scientific papers, news articles, etc.? No wonder Anthropic just scooped it up and took it for free!
Yet you don’t seem bothered by this. I wonder why.
the question was: what is the endgame for the stated "second class labs" strategy of distilling their frontier competitors then undercutting them on price?
Yes yes, we all understand the game-theoretic race-to-the-bottom you're describing here. Somehow despite linux being FOSS it still powers most of the important computing in the world. Can you explain how that works despite it being free? Once you understand that case I think you'll understand the game-theory behind how large projects can exist in the absence of traditional IP protection.
the obvious difference is the massive scale of data and compute required to develop and evolve these models, and the costs they impose on those building them.
Smaller budgets, slower improvement, less risk. They're not entitled to profits if that business model isn't sustainable. They're not entitled to a change in IP laws to protect their business model. They're not entitled to growing that fast.
The average private banking or wealth management client gets sold things far less destructive than the Robin Hood client who is far less able to withstand it.
I've been in quite many Telegram chats, none of which has enabled it. For most practical purposes you can just consider Telegram not to have e2e since it's no good if it's not used.
reply