>In an interview with Robert Wright in 2003, Dyson referred to his paper on the search for Dyson spheres as "a little joke" and commented that "you get to be famous only for the things you don't think are serious" [...]
To be fair, he later added this:
>in a later interview with students from The University of Edinburgh in 2018, he referred to the premise of the Dyson sphere as being "correct and uncontroversial".[13] In other interviews, while lamenting the naming of the object, Dyson commented that "the idea was a good one", and referred to his contribution to a paper on disassembling planets as a means of constructing one.
Thanks for pointing out those follow ups. Interesting stuff!
> correct and uncontraversial
From the original quote it is clear he was referring to the idea of aliens being detectable by infrared because they will absorb all of their sun's energy. Later in the same paragraph he says:
> Unfortunately I went on to speculate about possible ways of building a shell, for example by using the mass of Jupiter...
> These remarks about building a shell were only order-of-magnitude estimates, but were misunderstood by journalists and science-fiction writers as describing real objects. The essential idea of an advanced civilization emitting infrared radiation was already published by Olaf Stapledon in his science fiction novel Star Maker in 1937.
So the Dyson Sphere is a rhetorical vehicle to make an order-of-magnitude estimate, not a description of a thing that he thought could physically exist.
Full quote from the video cited before "the idea was a good one":
> science fiction writers got hold of this phrase and imagined it then to be a spherical rigid object. And the aliens would be living on some kind of artificial shell. a rigid structure surrounding a star. which wasn't exactly what I had in mind, but then in any case, that's become then a favorite object of science fiction writers. They call it the Dyson sphere, which was a name I don't altogether approve of, but anyway, I mean that's I'm stuck with it. But the idea was a good one.
Again he explicitly says this "wasn't exactly what I had in mind." This one hedges a bit more and could be interpreted as his saying the idea of a Dyson Sphere is a good one. He may have meant that in the sense of it being a good science fiction idea though, and he subsequently goes on to talk about that.
The Dyson Sphere is good for order-of-magnitude calculations about hypothetical aliens, and also for selling vapourware to the types of people who uncritically think that vapourware is real.
Have you read the paper itself, not just summaries of the idea? It's obvious from the way he wrote it, dripping in sarcasm. Talking about "Malthusian principles" and "Lebensraum", while hand waving away any common sense questions about how the mass of Jupiter would even be smeared into a sphere around the sun, just saying that he can conceive of it and therefore we should spend public money looking for it. He's having a lark.
Also, he literally said it was a joke, and was miffed that he was best know for something he didn't take seriously.
He thought SETI listening to space radio waves was dumb, so made essentially a satirical paper saying we should look for heat instead, because "an advanced civilization would be using these Shells to capture all star energy, so we could only see the heat"
The "dyson sphere" was a made up and entirely unfounded claim, without justification.
You’ll just end up approving things blindly, because 95% of what you’ll read will seem obviously right and only 5% will look wrong. I would prefer to let the agent do whatever they want for 15 minutes and then look at the result rather than having to approve every single command it does.
That kind of blanket demand doesn't persuade anyone and doesn't solve any problem.
Even if you get people to sit and press a button every time the agent wants to do anything, you're not getting the actual alertness and rigor that would prevent disasters. You're getting a bored, inattentive person who could be doing something more valuable than micromanaging Claude.
Managing capabilities for agents is an interesting problem. Working on that seems more fun and valuable than sitting around pressing "OK" whenever the clanker wants to take actions that are harmless in a vast majority of cases.
It’s not just annoying; at scale it makes using the agent clis impossible. You can tell someone spends a lot of time in Claude Code: they can type —dangerously-skip-permissions with their eyes closed.
It's not reliable. The AI can just not prompt you to approve, or hide things, etc. AI models are crafty little fuckers and they like to lie to you and find secret ways to do things with alterior motives. This isn't even a prompt injection thing, it's an emergent property of the model. So you must use an environment where everything can blow up and it's fine.
People here are claiming that this is true of humans as well. Apart from the fact that bad content can be generated much faster with LLMs, what's your feeling about that criticism? It's there any measure of how many submissions before LLMs make unsubstantiated claims?
Thank you for publishing this work. Very useful reminder to verify sources ourselves!
I have indeed seen that with humans as well, including in conference papers and medical journals. The reference citations in papers is seen by many authors as another section they need to fill to get their articles accepted, not as a natural byproduct of writing an article.
At some point you're forced to either believe that people have never heard of the concept of a force multiplier, or to return to Upton Sinclair's observation about getting people to believe in things that hurt their bottom line.
Because a difference in scale can become a difference in category. A handful of buggy crashes can be reduced to operator error, but as the car becomes widely adopted and analysis matures, it becomes clear that the fundamental design of the machine and its available use cases has fundamental flaws that cause a higher rate of operator error than desired. Therefore, cars are redesigned to be safer, laws and regulations are put in place, license systems are issued, and traffic calming and road design is considered.
True, but humans got a 20 year head start and I am willing to wager the overwhelming majority of extant flagrant errors are due to humans making shit up and no other human noticing and correcting it.
My go too example was the SDI page saying that brilliant pebble interceptors were to be made out of tungsten (completely illogical hogwash that doesn't even pass a basic sniff test.) This claim was added to the page in February of 2012 by a new wikipedia user, with no edit note accompanying the change nor any change to the sources and references. It stayed in the article until October 29th, 2025. And of course this misinformation was copied by other people and you can still find it being quoted, uncited, in other online publications. With an established track record of fact checking this poor, I honestly think LLMs are just pissing into the ocean.
Perhaps so. On the other hand, there's probably a lot of low hanging fruit they can pick just by reading the article, reading the cited sources, and making corrections. Humans can do this, but rarely do because it's so tedious.
I don't know how it will turn out. I don't have very high hopes, but I'm not certain it will all get worse either.
The entire point of the article is that LLMs cannot make accurate text, but ironically you claiming LLMs can do accurate texts illustrates your point about human reliability perfectly.
I guess the conclusion is there simply is no avenues to gain knowledge.
> I am willing to wager the overwhelming majority of extant flagrant errors are due to humans making shit up
In general, I agree, but I wouldn't want to ascribe malfeasance ("making shit up") as the dominant problem.
I've seen two types of problems with references.
1. The reference is dead, which means I can't verify or refute the statement in the Wikipedia article. If I see that, I simply remove both the assertion and the reference from the wiki article.
2. The reference is live, but it almost confirms the statement in the wikipedia article, but whoever put it there over-interpreted the information in the reference. In that case, I correct the statement in the article, but I keep the ref.
Those are the two types of reference errors that I've come across.
And, yes, I've come across these types of errors long before LLMs.
If you have high time preference yes. Over longer time horizons I think the issues with vibe coded software will reveal themselves in the same way badly written software does.
But more code from AI means stocks go up. Stocks are assets. If you generate enough code the assets will outnumber the liabilities. It’s accounting 101. /s
I'm not sure if this is widely known but you can do a lot better even than AGENTS.md.
Create a folder called .context and symlink anything in there that is relevant to the project. For example READMEs and important docs from dependencies you're using. Then configure your tool to always read .context into context, just like it does for AGENTS.md.
This ensures the LLM has all the information it needs right in context from the get go. Much better performance, cheaper, and less mistakes.
Cheaper? Loading every bit of documentation into context every time, regardless of whether it’s relevant to the task the agent is working on? How? I’d much rather call out the location of relevant docs in Claude.md or Agents.md and tell the agent to read them only when needed.
As they point out in the article, that approach is fragile.
Cheaper because it has the right context from the start instead of faffing about trying to find it, which uses tokens and ironically bloats context.
It doesn't have to be every bit of documentation, but putting the most salient bits in context makes LLMs perform much more efficiently and accurately in my experience. You can also use the trick of asking an LLM to extract the most useful parts from the documentation into a file, which you then re-use across projects.
> Extracting the most useful parts of documentation into a file
Yes, and this file becomes: also documentation. I didn’t mean throw entire unabridged docs at it, I should’ve been more clear. All of my docs for agents are written by agents themselves. Either way once the project becomes sufficiently complex it’s just not going to be feasible to add a useful level of detail of every part of it into context by default, the context window will remain fixed as your project grows. You will have to deal with this limit eventually.
I DO include a broad overview of the project in Agents or Claude.md by default, but have supplemental docs I point the agent to when they’re working on a particular aspect of the project.
Sounds like we are working on different types of projects. I avoid complexity at almost all cost and ruthlessly minimise LoC and infrastructure. I realise that's a privilege, and many programmers can't.
Yea but the goal it not to bloat the context space.
Here you "waste" context by providing non usefull information.
What they did instead is put an index of the documentation into the context, then the LLM can fetch the documentation. This is the same idea that skills but it apparently works better without the agentic part of the skills.
Furthermore instead of having a nice index pointing to the doc, They compressed it.
Their approach is still agentic in the sense that the LLM must make a tool cool to load the particular doc in. The most efficient approach would be to know ahead of time which parts of the doc will be needed, and then give the LLM a compressed version of those docs specifically. That doesn't require an agentic tool call.
Context quite literally degrades performance of attention with size in non-needle-in-haystack lookups in almost every model to varying degrees. Thus to answer the question, the “waste” is making the model dumber unnecessarily in an attempt to make it smarter.
The context window is finite. You can easily fill it with documentation and have no room left for the code and question you want to work on. It also means more tokens sent with every request, increasing cost if you're paying by the token.
Think of context switching when you yourself are programming. You can only hold some finite amount of concepts in your head at one time. If you have distractions, or try to focus on too many things at once, your ability to reason about your immediate problem degrades. Think also of legacy search engines: often, a more limited and focused search query vs a query that has too many terms, more precisely maps to your intended goal.
LLM's have always been at any time limited in the amount of tokens it can process at one time. This is increasing, but one problem is chat threads continually increase in size as you send messages back and forth because within any session or thread you are sending the full conversation to the LLM every message (aside from particular optimizations that compact or prune this). This also increases costs which are charged per token. Efficiency of cost and performance/precision/accuracy dictates using the context window judiciously.
Docs of dependencies aren't that much of a game changer. Multiple frameworks and libraries have been releasing llm.txt compressed versions of their docs from ages, and it doesn't make that much of a difference (I mean, it does, but not crucial as LLMs can find the docs on their own even online if needed).
What's actually useful is to put the source code of your dependencies in the project.
I have a `_vendor` dir at the root, and inside it I put multiple git subtrees for the major dependencies and download the source code for the tag you're using.
That way the LLM has access to the source code and the tests, which is way more valuable than docs because the LLM can figure out how stuff works exactly by digging into it.
It is not an "idea" but something I've been doing for months and it works very well. YMMV. Yes, you should avoid large files and control the size and quality of your context.
I have run my own mail server for years and I rarely see spam. I'm running a classic Bayesian filter as outlined in the legendary PG post "A Plan For Spam" and it works very well. I don't really get all the fuss about this issue. When I do see a piece of unclassified spam I simply classify it and continue. For me this is a far better tradeoff than having all my most private mail on some bigcorp server where any nerd can rifle through it.
> For me this is a far better tradeoff than having all my most private mail on some bigcorp server where any nerd can rifle through it.
You've functionally given yourself very little extra privacy because the vast majority of emails you send or receive will still cross through BigCorp servers (whether Google, Microsoft, Intuit, or other).
You can do the work to run your own mail server, but so few other people do that one end of the conversation is still almost always feeding a corporation's data lake.
I agree with you but I still run my own mail server. If people like me stopped doing that, we would cede the entire email landscape to BigCorp. A sad fate to happen to one of the true decentralized protocols. It's like if we all just went back to AOL
It's more expensive and difficult to hack or get a warrant to access multiple bigcorp servers with a variety of privacy stances and jurisdictions, than it is to get access to a single one. Security is about making attacks expensive.
No single BigCorp employee can go through all my mail.
If you're not convinced, no problem, please continue to enjoy your BigCorp email service.
And yet, if you're communicating with someone else who does the same (or uses a niche hosted provider), that entire conversation is outside their "data lake".
reply