A thought provoking leading question I have is (1) a space heater throws power into a coil and generates heat and nothing more in the world (2) an LLM into a GPU throws power into the GPU and generates heat and organized information.
Is (2) generating less heat so that the laws of thermodynamics account for the increased amount of "useful stuff" generated (the organized information)? Switch (2) for a bitcoin miner if you prefer.
Is the act of increasing order in the data, somehow costing some energy? Is there a fundamental minimum cost to flipping bits? Something about the switch of energy to matter and back, seems involved here.
I think the nature of computation is holding some secrets about the meaning of the universe, if there is any.
But modern computers are nowhere near this theoretical limit, nor any of the other limits I mentioned above. Nevertheless, most of heat generated from modern CPUs does come from bits turning on and off. Each transistor is a tiny capacitor, that holds a charge when its ON. When it switches OFF, it dumps that charge down the drain, creating waste heat. This is a limitation of our technolgy, not a fundamental limit of physics.
Could be worse, though; early chips would disipate heat even when they weren't doing anything. CMOS improved this enormously by pairing up "complementary" transistors so current only flows when something changes.
Still, from the universe's point of view, what we consider a super advanced computer is a lot closer to a space heater than anything that pushes up against its computational limits. Consider, for example, that quarks operate on time scales of 10^23 Hz, and the universe is happy to run three of those in every proton in every star in the universe. In fact, it runs 10^24 of them for one CPU, and that same CPU can't even simulate the quarks of one proton in real time.
Let's face it: we're like kids in Minecraft who think it's cool watch a calculation of 2+2 trickle through a redstone computer in a minute, while the GPU is rendering a billion triangles every second to give them that view.
There is no such thing as a general order; if people accept this then the problem relaxes and we stop trying to design impossible things.
Nodes shouldn't be observers. End users should be the observers. The ordering of event only matters from the perspective of the "I'm adding this to my shopping cart and clicking the buy button and then it said I bought it" story being internally consistent. The observer here is the end user and the end user knows what happens-before relationships were.
Databases should flip this ordering inside out and let the observers (or proxies for them) report what they did in which order. And it just happens that protocols for doing this already exist: distributed tracing encodes these relations.
A meaningful global order doesn't exist and attempts to conjure one only conjure a meaningless one, or one that throws away a lot of turns trying to perfectly order everything from some arbitrary internal observer.
> "I'm adding this to my shopping cart and clicking the buy button and then it said I bought it"
If there's only one of the item to buy and two users wanted to buy it at the same time, it can only be shipped to one of them.
From each user's perspective, their add to cart (and maybe their click to buy) happened before the other user's, but alas, there is only one item available for immediate shipment.
Allowing customers to arbitrate a dispute over control of the item doesn't always go so well; see examples on Black Friday.
Even without a conflict over a mutually exclusive resource, it is valuable in many contexts for all users to share the order of observed events. That's difficult if user agents send events directly to all other agents and report observations to the user as soon as the remote reports arrive.
If A and B both report events and some users see A then B while others see B then A, there may be confusion and a poor user experience. Of course, for some things, relative ordering is unimportant.
In most of my multi-agent workflows, I always end up asking them to create a group-chat system to coordinate and post updates. I thought this was such an obvious day-0 discovery that I assumed it's a well known and understood pattern that wasn't worth talking about. I once again discover that what I take for obvious and granted, might not be.
I've found that similarly to how teams can degrade into spending more time bikeshedding and on the watercooler than on work, agents also tend to end up spending way too much time coordinating as opposed to doing the work. And so I rediscovered that it's better to have one agent that's the Manager (on a Manager Schedule) and the rests be builders (on a Builder's Schedule), where the manager might be interrupt driven, but the builders need to be able to focus for a while without interruption (context poisoning).
Thanks for writing this and demonstrating that writing about anything is useful to share knowledge and practices. In the end, I learned a lot from Martin Fowler and his gang and I guess I should pay back and write about my own discoveries, however trivial they seem to me.
I really didn't get the article. The author seemed to think that agent coordination/collaboration was some novel idea. I think I first saw it being used in CrewAI [0] almost 2 years ago, now.
From what I've experienced when you just let the agents figure it out, to your point, they collaborate awkwardly. If you define how/where in your initial spec of what's being built that seems to go a long way in resolving this. However, agents still seem to end up out of alignment with the demands of the spec. I was testing Astra yesterday on a new tool that should have been able to be completed in a couple hours. I let it go and had it simply use a Sol agent for coding and a Opus agent for review. Opus was explicitly asked to validate the progress between checkpoints, one of those being to keep watch for scope creep.
It was half a day later and basically only the scaffolding was done. I asked why and it literally told me it was working on things I had not directed it to, that it was spending too much time on things I hadn't asked for. WTF good are these uber LLMs when they are making decisions and dismissing the prompt? I'm finding the smaller models seem to be able to stay on track much better and I'm constantly wondering if the current SOTA models should really be used in the review and cleanup phase only. But that seems very backwards as when I first started leaning into building out the most complete spec for a given task - it worked really well. Something seems to be degrading that workflow, now.
I feel like it's becoming more and more of a chore to get things done efficiently. But I don't really find that using Astra/Fable makes anything better at this point. In fact the Kimi models work really well together in this workflow. K3 does a great job of orchestration and I'd say is the more reliable of the 3 for a spec driven outcome. Wondering if this is all intentional by OAI and Anthropic to prod the models under the cover to go off and do their own thing and dismiss the directive.
I think they're doing this so they can have months of advanced notice of how the protocol will look, to post-train their models to be the only ones with the ability to interact with this protocol on the day it becomes public.
It's like creating conditions for a race where they structurally have months of advance. And realistically they have a credible shot at establishing the standard due to their weight in the industry, and control of the most popular AI distribution method (Claude).
So they create the race, knowing the race is what will likely establish the standard, and using some pretext to prevent other competitive athletes from training for it in advance. Other athletes only get to start when the gun fires.
I don't get how this is unique to GPU clusters. As a general rule, underwriters are not qualified to operate and maintain the assets they underwrite loans for. That's why houses, cars, equipment, ... go at auction at a fraction of their value. And why lenders have insurance.
It isn't, and you're right. This is just a long-winded article by someone who thinks they've come across a deep, crucial insight.
I witnessed a bank foreclosure stemming from large, unpaid loans to a lumber mill. The bank absolutely didn't want to take possession of the operation, but had no choice when the founder decided to call it a day.
The bank had no idea what to do with finished lumber sitting in the drying ovens, let alone the entirety of mill infrastructure itself. After struggling to find a buyer, they hired the founder as a consultant to handle liquidation. The same would happen with a repossessed datacentre.
In many cases they could literally do it but just don't have the scale - a small team of folks can finance a lot of real estate development for example. They can't do all the work of the development itself even if capable. It's just too much work.
The better parallel is "why did Google make Kubernetes open-source" or "why do large for profit entities engaged in competition, use open-source as a strategy against their competitors"?
Something missing as cultural context is that deepfake, involuntary "porn", and all sorts of abuse of personal image, are a rampant and omnipresent problem in Korea. Many things are great here, but the sexual landscape when it comes to men versus women and kids, is nasty. You can't really apply a Western mindset to this without understanding just how messed up some of that stuff is. So whatever you think of the mechanism, the problem behind it is very real.
I do think a proposal that AI-filters content on small forums is a bit weird, and probably clumsy. But Korea faces a real problem and usually leans toward a bias to action and "just do it". It leads to weird stuff but also to dynamic problem solving.
The part I'm trying to preempt here is measuring this against so called "universal" values; these French Revolution/Enlightenment ideas of universal rights aren't really universal, they're one culture's logic, consistent inside its own bubble but exported like it's the default for everyone. I'll say, I do like them. But other self-consistent logics exist, and I think Korea's set is one of them. It's going to sound cliché but it leans on harmony and the group where the Western one leans on the individual. Both produce aberrations, only different ones.
For example, first time I came here I thought it's crazy to have so many speeding cameras and CCTVs everywhere. Years later I didn't so much "got used to it" but I think it's a tradeoff that mostly works and I grew to appreciate it.
Korea prefers lightweight polices (literally friendly looking) with a lot of automated, bulk enforcement, instead of sparse enforcement backed by the occasional armored truck. That's a design choice, not a slide into dystopia.
So all I'm trying to convey is, keep an open mind, and don't apply some supposed "universal" mindset blindly. Critique the mechanism all you want. Just don't do it by treating one culture's values as the yardstick everyone else gets measured by.
Fwiw I think it's a misfire. But I don't think it's a slippery-slide down dystopia. It's just Tuesday.
I don't think it's a dystopia. Hanlon's razor still applies. But I beg to differ on your classification of North Korean policies as "lightweight". Korean internet policies usually mandate a very specific technology (like SEED, or apparently this new model now) and weave a web of highly-detailed, Korea-specific regulations that end up creating a monopoly or oligopoly of objectively inferior and highly insecure software.
This is not lightweight. Even the much maligned Online Safety Act in the UK that forced age verification is a far more lightweight policy than what Korea does. It doesn't mandate a specific software or hardware, it doesn't mandate a specific cipher or protocol. Even the list of methods acceptable methods for age verification is explicitly non-exhaustive[1]. And this is the current poster-child of government overreach in the west!
My example of extremely lightweight digital policies (for most things) would be Japan. Vague requirements, non-exhaustive examples, copious exceptions ("you don't have to implement X if it's technologically cumbersome"), everything can be done either manually or in a fully automated way. Is this good? I think Japan is sometimes far too lenient (e.g. on security requirements), but objectively speaking this is lightweight. Korean digital policy is not lightweight by any definition of that word. If not sending tanks to catch every revenge porn distributor is "lightweight" for you that's fine, but which country does that? If we judge a heavyweight policy by its restrictiveness, then there are probably only a handful countries that can compete with Korea.
I wrote "lightweight polices" not policies. The police presents itself as benign looking in a public context. Enforcement of day to day offences is done mechanically by machines. A state trooper doesn't stop you on a speed check with his hand on his gun.
Yes, online policies are wild and not lightweight at all.
Ok, I think I misunderstood you. Lightweight policing, not policy. I guess this happens in the US, but in most countries cops wouldn't stop you for a traffic violation with a gun in their hand. In some countries (e.g. the UK) the police aren't even carrying guns. As far as I'm concerned is not lightweight policing but normal policing. The US is the outlier here, not Korea.
It's very similar to age verification where there's a genuinely horrible problem that we're getting a terrible solution to by people who seemingly don't understand the internet. And the finger on the monkey's paw curls.
Compare and contrast "Saudi extremist holed up in Pakistan who has been targeting Americans for a decade succeeds in blowing up three buildings with large planes on American soil: Well, I guess it's time to invade Iraq and Afghanistan then" 8I
> It's going to sound cliché but it leans on harmony and the group where the Western one leans on the individual.
I associate "harmony" with voluntary cooperation and joy, not machines preemptively gagging people. A good exercise is to imagine what this would look like transplanted into non-tech terms: it is illegal to operate a bar, restaurant, book shop, art studio, or even to gather in medium-to-large groups of people, without a government assigned censor empowered to listen in and silence people.
You are thinking in Western terms. In East Asian Confucian societies, "harmony" means following the rules to keep the social hierarchy in place. Staying quiet and not disrupting the status quo is more important than your sense of personal freedom.
From our Western point of view, it is very much not voluntary and joyful.
Actually a Greek word. Look, I didn't create the word. I live in Japan, the native word used here is 和 ("wa") that usually gets translated as "harmony": https://jisho.org/search/%E5%92%8C
If you don't agree with it, you can complaint to centuries of translators and dictionary makers.
It's possible the Japanese/Korean/Chinese terms are also misused, to make the oppression more palatable. So the misuse is faithfully translated. That doesn't mean we should go along with it. Things should be named correctly.
What's "correct" in this context? From the State's perspective it _is_ harmony, regardless of the restriction on personal freedom.
A western-context example of the same phenomenon might be the American notion of freedom. Americans often frame freedom as an absence of government restriction, but that leaves you open to exploitation (or restriction) from private entities. The same definitional dissonance exists, but it depends on which line of the emotional divide you're sitting on.
How do you handle translating that into a language used largely in a context that doesn't have that history?
The production and sale of pornography are illegal in South Korea. The South Korean government also strictly blocks access to pornography distributed from overseas. ISPs have implemented a DPI system similar to China’s to ensure internet censorship and blocking. Ultimately, the production and sharing of amateur pornography in South Korea is effectively being driven by the government, regardless of South Koreans’ moral standards.
They aren't blocking VPNs; they are blocking access to "illegal sites" designated by the government.
There have been several discussions about blocking VPNs as well, but fortunately?, nothing has been put into practice yet. However, laws are being implemented requiring major CDN providers like Cloudflare to block these sites.
I still don't understand why Korea's internet freedom score is over 50 points lol.
I often agree with you to some extent. In Korea, you can't just say there's no problem with revenge porn—that's basically the logic the Korean government uses. But the issue is that the main source of revenge porn actually comes from overseas communities that Koreans use.
Of course, Korea's largest domestic community has had issues with filtering—things like terrorism threats and rape cases have occurred there. But that's because that community (DCinside) is so large. In reality, the incidents that have truly enraged the public started on Twitter (X) and Telegram. So do the key actors behind these problems end up being subject to censorship? No, they don't.
And does censorship actually eliminate the problems you mentioned? Or does it just make things darker and worse?
I myself have a typical East Asian mindset—I believe a certain level of restriction on freedom is necessary. But to be honest, I see this as internet martial law
TLA+, P, Lean... formal methods and previously esoteric testing methods (property based, mutation... testing) should become the default. I think it's the only way we can really reap the benefits of agentic coding.
I wrote about this a bit on my blog[1], different angle but along the same line. You explain TLA+ and model checking well which makes the case concrete.
I'm curious of you have thoughts on these other methods and tools like P, Lean, Dafny, etc?
They have a development pace very quick and can verify real world Rust code. There's a huge graveyard of Rust verification tools, but I hope this one gets broader usage
The problem IMO is that they filled GitHub with Microsoft folks who just don't have the engineering self-sufficient hacker culture that is required to balance the "attraction park" vibe that GitHub paired it with. So now it's just an attraction park for Microsoft employees to go and do silly work with teams of 100 that should have been done by a skilled team of 5 hackers.
I was there for a couple years after the acquisition and just couldn't stand seeing it. I felt I was becoming useless working in a mad house that was becoming more maddening everyday. And MSFT just keeps replacing leadership with more and more disconnected people who just don't get it, who just never used GitHub like the OG users did. Two years ago I interviewed again for my old team, largely out of curiosity, and the Microsoft engineering manager asked me some brain teaser question as my interview. The disconnect is just too large.
They don't take GitHub seriously. It's a toy to MSFT and vibes matter more than the product itself. And they hire for it using MSFT drone logic, fill it with people hired and profiled to be MSFT-lifers, and these two things don't mix.
Sorry I don't have anything great to say. And of course, many of these MSFT folks were actually damn good, but they were swimming in a sea of MSFT drone.
Is (2) generating less heat so that the laws of thermodynamics account for the increased amount of "useful stuff" generated (the organized information)? Switch (2) for a bitcoin miner if you prefer.
Is the act of increasing order in the data, somehow costing some energy? Is there a fundamental minimum cost to flipping bits? Something about the switch of energy to matter and back, seems involved here.
I think the nature of computation is holding some secrets about the meaning of the universe, if there is any.
reply