I see Munroe's work as filling the same role in society as did Socrates in his time. Not only in commentary about current events, government, society, etc but also in expressing his viewpoints in a fashion accessible to society. Socrates paved the way to bring philosophy to the masses. Munroe uses a popular medium and comedy to the same effect.
The Gaussian Processes underpinning this work are hardly a product of the 'AI Hype Machine' - they've been around for decades, have strong statistical underpinnings, and are being widely explored for experimental design across many disciplines. Reflexive and poorly-informed backlash to any variety of machine learning is no more productive than blindly hyping up LLMs.
Meta Platforms, Inc featuring this technology with a title announcing “AI for American-produced cement and concrete” is, on the other hand, 1000% a product of the AI Hype Machine.
Sure, it's clearly marketing. I think a private company pursuing marketing via open research with open source code (including datasets) is a good trade. A hypey blogpost + research is better than no blogpost and no research.
A sidenote along these lines - I've recently done an MSc, and found that the default approach to lectures is now to present slide decks. One of the profs, however, delivers a more traditional lecture, writing everything on a blackboard. I've found the second style far more effective, largely because writing caps the rate at which information can be conveyed. Because slides have no such bottleneck, I've found they're often misused and overladen with information which is skipped over too quickly.
Do you have any evidence that inference revenue is growing faster than training costs? RLVR is significantly less compute-efficient than token-prediction pretraining - especially as labs are trying to train models to achieve agentic tasks which take tens of minutes per rollout.
It's definitely true that they've increased their revenue rapidly. But at the same time the 'scaling laws' that the labs were first built around require exponentially-scaling cost (10x flops for a fixed reduction in training loss).
If anything, a better look at the economics is a reason to look forward to one of them IPO-ing. I suspect the labs probably could cut R&D and turn a profit, but that might only work for one generation, until they get superseded by the competition.
There is no doubt that competition is what is driving unprofitability. So when people say AI can't be monetized, I laugh. Right now, foundational AI is unprofitable because of competition, not because they can't make money.
But this is exactly the problem - we have to take it on faith that inference is profitable because nobody actually knows. It’s hard to even define what that would mean, and while I am suspicious of claims that frontier lab CEOs are just out-and-out liars or bad people, defining and calculating the real cost of inference would be time- and labor-intensive in its own right and there is no strong incentive to do it other than “tech reporters are curious.” Until the IPO, we just won’t know.
A lot of people know. A lot of insiders have been saying tokens are profitable. Is there a conspiracy theory for everyone to lie? Including OpenAI, Anthropic CEOs, employees, Cursor management, inference providers of Chinese models?
Profitable on what basis? They generate more revenue than the cost of electricity? Does that factor in the cost to service the massive, multi-layer cake of debt that was necessary to even begin to serve inference in the first place - not from a training perspective but from a hardware and facilities perspective?
I’m not talking about training costs. I’m talking about startup costs. You have to pay for GPUs (or to rent data centers). You have to pay for the electricity that runs those data centers, and in a lot of cases these frontier labs are building the data centers on credit, so you need to pay for the construction, the materials, etc. If it was as simple as “running the GPUs costs less than we charge for it,” I might be inclined to agree. But the GPUs don’t just appear by magic.
Right now, the demand is far more than supply for GPUs. Every cloud company is saying they're leaving money on the table because they don't have enough compute to serve the demand.
It seems like you're arguing that the bubble is going to collapse soon, like the author? How can it collapse when the demand is so much bigger than supply? Do you think the demand is fake? Or that AI will stop making progress from here on out?
The demand is real. The tech is real. The economics are completely unsustainable. Switching costs and barriers to entry are too low, operating costs are too high. And if the tech improves, it actually makes it even easier for competitors to swoop in and take market share. Not long ago, an agent that was 80% as good as SOTA was not usable. A year from now, an agent that is 80% as good as SOTA will be better than the best agent is today. We have it on good authority that today’s agents are very good, very useful. Why bother paying full price?
This is deeply ironic in a way. Because the whole premise of AI labor replacement is that AI does not need to be better than human labor, it just needs to be cheaper with acceptable performance. But the same is true one step down: discount AI doesn’t need to be better than bleeding-edge AI, it just needs to be cheaper with acceptable performance.
That really means a lot, ainch. I hope it makes your late-night sessions a little more bearable. If you find anything that doesn't work well with the papers you read, keep me posted
On that note, Terence Tao gave a good interview to Dwarkesh Patel talking about Kepler. He pointed out that the previous geocentric models were actually more accurate than Kepler's at the time, in part because they'd had so much complexity piled on to solve minor errors. Kepler's theory was more elegant, but at the time it wasn't necessarily a better model.
I think important paradigm shifts can often look like this - there's not necessarily a reason to expect them to be instantly optimal. Deep Learning vs 'good old-fashioned AI' is another example of this dichotomy; it took a long time for deep learning to establish itself.
I think it depends whether you can leverage some knowledge. It's possible for a person/LLM to look at a loss curve and say "oh that's undertraining, let's bump the lr" - whereas a Bayesian method doesn't necessarily have deeper understanding, so it'll waste a lot of time exploring the search space on poor options.
If you're resource unconstrained then BO should ofc do very well though.
Yah, I'm a bit skeptical - ime humans tend to under explore due to incorrect assumptions. Often this is due to forming a narrative to explain some result, and then over attaching to it. Also, agents aren't actually good at reasoning yet.
Good Bayesian exploration is much, much better than grid search, and does indeed learn to avoid low value regions of the parameter space. If we're talking about five minute experiments (as in the blog post), Bayesian optimization should chew through the task no problem.
I'd like see a system like this take more inspiration from the ES literature, similar to AlphaEvolve. Let's see an archive of solutions, novelty scoring and some crossover rather than purely mutating the same file in a linear fashion.
I was a little surprised to see a Telegram integration rather than Slack or Teams, given Anthropic's enterprise-first posture. But then I looked it up, and it turns out Telegram dwarfs both, at around 1bn MAUs, vs 50m and 300m respectively! I had no idea - reminds me of the time I found out Snapchat has 2x the userbase of Twitter.
I’ve been using Telegram for about 10 years, and it’s one of the few products that has consistently felt great the entire time. It’s fast everywhere: backend, mobile app, desktop app, all of it. Everything just works. Its sync is out of this world—fluid, fast, and seamless across devices. You can use it on your phone, then move to your PC or laptop and continue instantly without friction. Unlimited message history and file storage are fantastic, and the bot platform is absurdly powerful. It’s boring in the best way, which is exactly what you want from a channel for interacting with your agents everywhere.
Instead of Telegram, go self-hosted for company related communication activity. See Gamers Nexus recent video on self-hosted discord alternatives https://www.youtube.com/watch?v=kpjcmXbmMVM
For encrypted group conversation over third-party networks use Signal, or Matrix which try to keep your conversations private.
The fact that Telegram has their support playbooks online, is interesting. Though I would call the following claim a stretch
> As a result, we have disclosed 0 bytes of user data to third parties, including governments, to this day.
Telegram had always impressed me for the same reasons. They have constantly gotten worse since about 2022/2023 though. Dark patterns, pay gate, they lost chat history for some of my closest contacts including 15k+ lost photos, no support at all. Something changed in their product direction and I started moving all my chats to Signal.
They need to stop cramming in useless features like Stories and NFT's and they recently redesigned the Android app to be like iOS but it broke my muscle memory, moved ALL items around (like "Saved" being in 1st spot when sharing, after the update it was buried, then it was at the top again and now it's under "Create a Story") and trashed the performance with the chat list not even loading at start (switching the tabs at the bottom "fixes" this, speaking of tabs, who even needs these there? The hamburger menu was fine). At least the desktop program is good AND native and not web slop.
Since we are apparently giving messaging platform reviews here, I feel exactly the same way about Microsoft Teams. It works great. It does everything I want. It doesn’t get in my way. 10 out of 10 keep up the great work guys!
> Since we are apparently giving messaging platform reviews here, I feel exactly the same way about Microsoft Teams. It works great. It does everything I want. It doesn’t get in my way. 10 out of 10 keep up the great work guys!
It looks like we found a high executive using company money to buy a product no one wants to use.
It's easy to promote Teams if your secretary is handling it for you and you don't need to suffer yourself.
The other possibility: Microsoft started an astroturfing campaign on HN.
Teams network connectivity is a plain joke. If you use suspend, or frequently change network, the thing will just never reconnect, even though you have VPN alive and all network applications perfectly running.
And the thing is just absurbdly sluggish, only display blurred grey lines instead of text in a meager attempt to look snappy.
I only use Teams for meetings and the calendar, and the occasional chat during a meeting. I find it totally fine and I don't really think about it much one way or the other. For reference I have a 2021 M1 Max with 64 GB.
Probably all managers and engineers working on Teams have similar copious amounts of memory and powerful CPUs on their devices and hardly use their own product. That would explain a lot
It honestly wasn't much different on my 2018 i5 Mini with 32 GB.
Maybe what sucks here is the experience of running it on Windows. Or maybe it sucks for large meetings? But I never have Teams meetings with > 40 people at this company.
When it is online, I agree with things asides from the "fast" part, actually. But many companies have a secondary service for async comms/chat when being Teams cannot be online, and compared to Slack.
Honestly can't tell if this is not sarcasm/rage bait.
Teams that has 3 different UI frameworks on every platform (but your best bet is the web)? With the Microsoft login that tends to loop forever redirecting to God knows where?
Back in the day, when I used to play pokemon go, there was a small local community and we would struggle to decide where to meet up for the daily raids because people would basically not respond (so as t not commit), or not know which gym each other meant exactly, nor give live updates when people moved around, etc. etc.
Then I joined a group from a bigger city where I commuted for work. They had a telegram group chat with two "channels", one for talking, one for bot posts. The telegram bot could be sent a single screenshot of a raid, and it would use OCR to automatically generate an interactive UI for that raid for everyone to see, with all the relevant info, and it would also clear itself up when the raid is no longer relevant. You could press buttons to say you were going, that you MAYBE were going, if you were late, and if you already started/done it, all in single clicks. Tons of options, tons of information, all live updated.
I was bedazzled. That feature singlehandedly removed all attrition from urban social gaming. And it was entirely grassroots. It made me try out making my own telegram bots, and yeah, you basically have the power to make a little app in chat form, even some that feel like CLI commands.
It's been OVER HALF A DECADE and I have yet to see a single other chat application have that degree of freedom where it comes to applications and bots. Some like discord even did whole ass 100% reworks of their bot AP to support the likes of slash commands, and still fall short. And there's none worse than Teams. Teams hates you. Teams spent the prior 2 years before this one basically pointing a gun to our heads telling us they were removing webhooks and pushing back on it whenever they repeatedly get told that's the most insane and dogshit idea ever. And they still did it. There's just no spark in Teams UX. No self-respect. It's a soulless product made entirely as a dumping place of "synergy" with other M$ products. It's reciprocal, I hate it too.
Oh and my local group never go into telegram because they didn't want a new app. It died, but I still kept playing after work without problem. It makes me wonder how fast Teams would die if it wasn't proped up by 365 and Azure subscriptions.
> we would struggle to decide where to meet up for the daily raids because people would basically not respond (so as t not commit), or not know which gym each other meant exactly, nor give live updates when people moved around, etc. etc.
This kind of thing is so common in groups of people, it's one of my pet peeves. My own family does this in our group messages when trying to make big decisions like who should host thanksgiving or where we should go for a family vacation.
I make it a point to just take charge and tell people that we're doing XYZ now. It usually either results in a decision, or gets the discussion going enough that I can do it again with new information.
That has roughly been my MO as well and it works great for groups where identities have settled.
But one has to keep in mind that, in our currwnt "more woke" times, if you go this way in a new group you run the risk of being labeled an array of things. So tread carefully there.
I wonder if Teams hates you, because they are doing the bidding of their actual customers (corporate decisions makers and purse holders), and those people's interests are not exactly aligned with the users'.
The problem is that these people holding the actual purse don't care enough about their subordinates' experience. They care about the price tag, and about compliance. Apparently the makers of Teams think about the same. None of them thinks in terms of lost productivity.
Yes, compliance is a big one. And it's not so much that they are actively hostile to user productivity (and quality of life), they just don't care enough.
Odious is one of the most reserved words you could use to describe Telegram, which is primarily a host for scams that the influencers and other bottom feeders aren't allowed to monetize on the big social networks.
I feel the opposite. I'm on Teams all day at work and have reluctantly opened Telegram recently to try a Claw despite having an account for years.
I've been surprised how little support there has been for Teams in the whole AI ecosystem. It seems all developers assume that the whole world is at startups working on Slack when most businesses are on Microsoft 365.
Compared to operating on text files (which is relatively very simple and something Claude Code is great for), I have a feeling it's kind of a disaster dealing with Microsoft integrations and the different file formats
Just the fact how much Microsoft lies when you click the "keep me logged in" button should tell you why nobody bothers with Teams integration with anything.
The main reason is just how hard it is to actually create anything that integrates with Teams. You have to jump through so meany hoops, wade through so many deprecated APIs, guess through so many half-way-wrong-by-now documentation pages.
After building a proof of concept, we decided that we will only continue Teams integration if anyone is going to pay serious money for it.
Telegram's bot API is literally one of the friendliest APIs (of any kind) I've ever seen. It's the first thing I reach for when server-to-mobile notifications are concerned.
It's just as easy to set up as ntfy.sh, except that it doesn't break every other week on iOS.
This is so true. I don’t like Telegram for a host of reasons, but the bot architecture is second to none. Try creating a bot in Slack. You’ll pull your hair out for hours. Same goes for Discord. Utter nightmare. Telegram? You send a DM and it is basically done.
Discord webhooks aren’t too bad… but the proper bot thing is ridiculous. They really lack a development mode server, having to know everything about oauth and token permissions before even starting is bonkers and why do I even need an app is beyond me. I’d probably have my bot completely implemented in telegram in the same time I figured out what an app is in discord and how to even add a new app to my server.
Slack (and Discord) webhooks are good for just shooting one-sided data into channels, but for interactive bots Telegram is so far ahead of anyone else it's crazy.
Signal specifically is missing any kind of official bot support, cutting off massive audiences from even considering it as an option.
I think it might because telegram integration it's just easy to do, I don't use telegram for actually messaging, I use it just to deploy my bots, it's a simple way to build simple tools, in a few lines you can get something working, you can have commands that work like buttons, accept images, respond with images and don't need anything else than your telegram account
Surprisingly large number of businesses run on whatsapp, as a consultant in Asia it's prob around half the businesses I've worked with prefer it over teams/slack. If Meta had been sensible about API access Telegram wouldn't have even got a foothold.
WhatsApp is actually more popular than Slack, isn’t it? In my country, almost everyone uses Slack, and I’ve hardly ever heard of any companies using WhatsApp, so that was surprising to me.
Living in Japan has made me realize how different cultures can be — even down to the apps and services people use.
It honestly surprises me, and now I kind of want to try WhatsApp too.
Honestly Whatsapp is nothing special. It works well, just like many other chat apps nowadays. The interest is that in some large parts of the world, everyone uses it already.
The only thing it's that creating bots for whatsapp it's not as easy as for telegram and it cost money. Actually that is the business plan for whatsapp making money from whatsapp business
Getting an agent working via an existing Slack setup is fairly effortless and the control over output format is useful
I had a look at getting same agent up in WhatsApp but it seems to need FB business acc to even start process, to get FB business you need an FB personal etc. ... looked like too much effort
It's not even funny how a multibillion-dollar company with thousands of employees having unlimited access to the "world's best coding models" lags behind a small one-man [1] open source project that already had multiple plugins for the same feature [2] for months.
Pi already has 700+ third-party packages [2] for various purposes of various quality. But it doesn't matter, since creating a new working Pi extension to suit your needs is just a prompt away, and you don't even have to restart your coding session.
Telegram has the best programmatic integration. Trivial to get working. You can be up and running in minutes. I use it to talk to a claw-style agent and it's truly unbelievable what you get for free.
Apples and oranges comparison, one is a messaging app, the other two are used for communication and collaboration across teams in a workspace. I have worked in 5+ companies who used either Slack and Teams, none used Telegram for any comms.
Telegram is 'bot friendly' since the beginning, gaining a lot of users with crypto boom a decade ago with coin drops and things like that, so it is very good to develop for, but I have your initial sentiment first - shame this hasn't launched with tools people actually use for work.
One issue is that 95% of the integrations will be fine with the default configuration. The others including some with high profit potential will have weird configs that will frustrate your customers the first time they try if not well tested/documented. It's better to take time and get it right. Enterprise customers love piloting and spending time, so best to approach that the right way too. Going with less complex options, that arguably have better APIs, makes it easier to develop your core product too and get real feedback from users.
I'm bullish on Claude. It will see a surge in users, and will likely surpass Perplexity this year. However I don't think it will catch up to even Meta AI (which had 10x the number of users) this year.
I use Claude. I use Codex. I've never heard of or used Meta AI. Nor do I have a Facebook account. Never have, never will.
I am also a software developer. So while the numbers of "people" that use one AI or another may be higher than either of these, it's not a useful metric for myself.
That's fine. I'm not making a value judgement about which LLMs you should use, if any.
I'm only pushing back against someone thinking "oh HN talks about Claude a lot, therefore Claude must be extremely popular". The information bubble is a real problem.
It's probably true that Anthropic's revenue is booming. But we need massive grains of salt:
a) they are private and revenue numbers for private companies are hopelessly unreliable, and
b) they are planning an IPO, so there's an extra incentive to big up the numbers. Anthropic always brings up ARR, which is very gameable when the year hasn't ended yet
Talk about a bubble. No one outside of programmers know what the heck is Claude. In Asia, ChatGPT and Gemini dominates LLM usage, followed by Perplexity.
Microsoft released a report with some numbers on Deepseek adoption globally. They say it's got ~90% market share in China, and is growing in popularity across Africa.
Telegram has a major issue with bots and bad actors though. They paywalled privacy features making it truly a terrible experience for users. 3-10 per day random messaging you.
Telegram is more popular among "normal people", and it also has a laissez-faire attitude towards bots and bot development. Making a bot that you, or even other people, could add to their contact list and use is pretty easy.
It's wild, but "people who want to build and run their own one-off bot for something like home automation" are almost treated by Telegram like first class citizens.
You're telling me that Anthropic, one of the hottest companies on the planet right now couldn't field four teams of developers to integrate with Discord,
Slack, Telegram, and Teams? AI being such a productivity multiplier, seems like they could just choose to do it all. I mean, mythical man month and all that, but do it three times and have a retrospective and use Claude to refactor the pain points and centralize the learnings.
Not really a meaningful comparison. Telegram is a personal messenger while Slack and Teams are for work. Telegram should be put alongside WhatsApp, iMessage, WeChat etc., which all have user bases in the billions.
The human genome contains around 1.5GB of information and DeepSeek v3 weighs in at around 800GB, so it's a bit apples-to-oranges. As you say, what's been evolved over hundreds of millions of years is the learning apparatus and architecture, but we largely learn online from there (with some built-in behaviours like reflexes). It's a testament to the robustness of our brains that the overwhelming majority of humans learn pretty effectively. I suspect LLM training runs are substantially more volatile (as well as suffering from the obvious data efficiency issues).
If you'd like an unsolicited recommendation, 'A Brief History of Intelligence' by Max Bennett is a good, accessible book on this topic. It explicitly draws parallels between the brain's evolution and modern AI.
The comparison is weird as we don't think with the Genome. There are something like ~100 billion neurons with ~100 trillion connections in an adult human brain . I don't know how many bytes of sourcecode deepseek has, but I don't think it helps in determining the amount of reasoning it can do.
> The comparison is weird as we don't think with the Genome
The genome determines how your brain learns, so yeah we do. We don't solve short easy tasks via learning, no, but longer tasks that involves learning involves our DNA.
Learning also happens on the species level. The species "learns" (thru natural selection) which genes produces brain structures that lead to survival and reproduction.
The human genome isn't its own thing, the genome as a static sequence is really just an abstraction. What actually functions as the heritable unit includes epigenetic marks, non-coding RNA regulation, 3D chromatin structure, and mitochondrial DNA. In the real biological world there are very few sharp edges - systems bleed into one another and trying to define something like 'the number of bits in the human genome' is very difficult, but it's undoubtedly way bigger than you posit here.
> The human genome contains around 1.5GB of information and DeepSeek v3 weighs in at around 800GB, so it's a bit apples-to-oranges.
The apples-to-apples comparison is comparing the human genome to the code behind a particular LLM. The genome defines the structure that learns and thinks, just like the code for the LLM.
And that same information contained in an LLM is a compression of how many terabytes of training data? Maybe in the future there will be models an order of magnitude smaller and still better performing.
What I'm saying is you can't judge the data in the genome by purely counting the bytes of data.
reply