Not really. Google IPOd at a market cap of ~$41B after adjusting for inflation.
In different terms, Google IPOd at a market cap ~ 10% smaller than Ford’s at the time.
Anthropic is at 50x Fords market cap. It’s over half of Googles market cap, and Google is a competitor in the same market with a dozen already successful lines of business.
They are incredibly highly valued in a way where I don’t even understand mechanically how they could grow like that without the entire financial system crumbling.
Google is up ~ 6,600% since IPO. Apply that to Anthropic and they’ll be worth something like $132T in 2048. As a point of reference, the Congressional Budget Office is predicting the US national debt to be around $102T in 2050 (paper was from 2020, might be outdated).
Anthropic, the singular company, will have to grow to be worth 130% of the entirety of the US’ debt. Napkin math says their revenue would probably have to exceed the US governments tax income. Like genuinely, Anthropic would have to become a world superpower all on its own.
They would be so big that they’re an existential threat to whatever country they call home. They would be such a large portion of the economy that central banks would lose any kind of control (presuming Anthropic doesn’t issue their own currency, they’d be a larger economy than most countries).
I can’t see how that could happen mechanically. The world would burn to the ground even trying to make it happen. If you thought Standard Oil was bad, this would be that but for every product and service in existence and all run by the same company.
I don’t believe there’s a central coordinator (it wouldn’t be distributed then).
I think you’re missing that rooms are attached to servers, and you can join rooms that aren’t on your server.
If I go join a room with 50k people on another server, those 50k people need to pull an online status from my server and my server needs to pull all of theirs.
Multiply that by all the rooms 50k people have joined and you get a _lot_ of messages bouncing around.
It’s sort of hard to optimize without either a central and implicitly trusted cache, or giving up on trust entirely. Central cache is sort of obvious. You could go peer to peer but then you have to deal with “what if somebody modified their server to issue forged updates”. You could maybe sign them but I’m not sure how anyone would validate that without getting a pubkey from your server and now we’re back to scaling problems and have a new “storing a million pubkeys” storage issue.
It’s wild because the ability to identify your citizens is one of the only core competencies that spans basically every form and manner of governance, from a huge country’s government to city council.
Like virtually every other function of government first requires establishing who someone is and whether they’re a citizen.
It’s shocking to give that up. If ID.me decides tomorrow they don’t actually like the US government (crazy, but crazy happens sometimes), think about how many US systems just stop working because they can’t tell if you are you or even a citizen.
(ignore that some of these should be chained and some should be arguments, I haven’t read the proposal fully)
They’re doubly nice in Clojure because of the lazy evaluation. You can map a pipeline over a list and it will evaluate one at a time rather than pre-evaluating and allocating memory for the whole thing.
This is the first time “open source” has meant anything other than “probably free” to the average person.
I constantly hear people complain about UIs. You know who probably makes a great UI for non-technical retirees? A non-technical retiree. They’ll probably start from forking an OSS project, though.
I think the really clever folks will move up to building vim/emacs/Firefox esque apps that are heavily scriptable, and your weekend OSS projects will be done by non-technical folks who just want something really specific.
> I constantly hear people complain about UIs. You know who probably makes a great UI for non-technical retirees? A non-technical retiree. They’ll probably start from forking an OSS project, though.
You know what makes a great UI for non-technical people? Codex or Claude.
GUIs were a response to the limited resources available on old computers. And, as we're increasingly finding out, text is FINE. People like Teh Pretty(tm), but they'll take plain text if it solves their problem.
But, congratulations, you DID identify the primary bottleneck--customer support. And, no, I am NOT going to be unpaid customer support for non-technical people. Bleah! Get back to me when Claude/Codex don't suck for that.
Sensible people largely invest in index funds or retirement accounts, which this wouldn’t prevent.
Practically every single one of these proposals only restricts your ability to buy and trade a particular stock. You can still trade stocks, you can still have an account manager, you can still tell them you want more exposure to solar, you just can’t tell them you want to buy RandomSolarCo by 4pm Thursday.
My understanding (which may be wrong) is that English is best but most of the Romance languages do okay because English has borrowed from them so much.
I think you only start getting seriously bad impacts in non-Romance languages where they have their own alphabet, or use the English alphabet phonetically causing letters to appear next to each other that never would in English. E.g. in English "qu" is a pretty obvious token, you rarely see a Q without a U after it. Happens all the time in anglicized Middle Eastern words, though. Quran is like a whole other thing, you'd almost always expect another vowel after qu, not a consonant.
I've heard English is one of the most expressive languages, but is complex to learn due to absorbing words and ideas from so many other languages. Lots of contradictions, situational rules, and unwritten rules (like the order of adjectives: the big, brown chair instead of the brown, big chair).
I don't know how true that is, and I can't imagine knowing more than one language. I'm impressed by anyone who knows more than one....
I’m actually genuinely curious about this as well. I’m saturated at low single digits where I run out of things I actually want.
Like I could definitely find ways to do it if someone is paying me to burn tokens (have 5 of the agents come up with busywork for the rest).
I had my agent add OpenRouter support and provider toggling and it took 15 minutes-ish on a single agent.
Some of me wonders if these are just deeply unoptimized projects and agent harnesses burning tokens because it sounds fancy. I can’t imagine why you would ever need 200 agents for a normal codebase unless 200 agents was the goal or the codebase is an absolute nightmare.
Subscriptions from providers (Google for carbon, Openai for Sol 6.1 mainly, and Claude) are about 1k$ usd a month combined and provide about ~40k usd in available inference.
I save Claude Fable for being the 'architect/tech lead' and use Chatgpt which is IMO least compentent in general, using Sol 6.1 agents which are very cheap (Check Artificial Analysis for good writeups/cost/benefit leaderboards).
Google Antigravity is surprisingly competent, so it is going to be taking on the sub orchestrator roles when other usage is too low.
Ask yourself how you would plan to make use of 1000 free junior human software engineers and 100 free human senior software engineers (aka Fable) and you are asking an equivalent question.
Below is copy/paste of my reply to parent comment
Have a large spec for the task. Then fable head-orch reviews and is instructed to use git for work tracking, and also to assign well scoped independent 'modules' with their own dedicated repositories to sub orchestrators. These orchestrators get to work on the spec for each module, they spawn sub sub workers.
This scales really easily. Add more spec requirements or run on a separate project. So far I'm releasing a direct to consumer shop and a saas by the end of the year and trying to use up extra inference/usage on making some steam games I can release to help self-fund future work/server needs. Right now using Hetzner for servers as my gaming pc is no longer powerful enough, and using Hetzner for the ci/cd build server since speed of test suites are a major blocker on progress as Fable loves massive test suites (which seem worth it as it helps it work).
As a test run I was able to burn $40 usd via DeepSeek api with low cache miss using Deepseek v4.1 flash in about 2 hours, and if you know its cost, you will understand how insane that is.
But is it useful? Worth anything or am I just burning tokens? Well, it appears quite useful. I have my own Ci/Cd and Git so I can save on the massive GH actions, my own custom MCP so that the swarm can communicate with each other and use harnesses from each provider I have usage on. (Api keys would be better but subscription TOS usually requires harness to be used with the sub). I've already built a steam game with custom rust engine that seems to work well enough that I'm about to release. And my shop is solving a hard problem with my own codebase using determinisitc code that has been heavily tested and iterated on. The SaaS is going to be a chatgpt/claude plugin/mcp and is for job app analysis and applications (relegated for now to chrome extensions due to TOS/legal because using AI to retrieve 2fa codes and solve captchas is a big no no still, so customers will have to handle that on their own machine/browser/phone if they want auto-apply + other automated feature set.
It will, but it does so by burning tokens. I'm sure you've also seen the output of an overly eager grep by an LLM.
I think we're verging on the point where people will care more about the most token-efficient ways to do a task than the lowest human effort version. Humans are expensive, humans using half as many tokens are cheaper.
I can't read the article because of the paywall, but from what I can grasp it's sort of a "generic GPIO controller".
Offhand, I'd be a little worried that it could be replicated for way cheaper with any run-of-the-mill LLM agent and an esp32. I've done some esp32 dev with LLMs and they don't really struggle with it, and I'm suspicious this is going to be way more expensive than "esp32 + random GPIO peripheral from 30 years ago".
The examples in the article feel like something any semi-modern LLM could one-shot, but I can't read it so maybe I'm underestimating the complexity.
I bought it because it’s turnkey and to support the Home Assistant foundation. I want support, not a project I have to maintain myself, that’s what the money buys.
If you have tokens and a bunch of free time, your proposal is an option. I just want it to work with my Home Assistant, and so I’m happy to give them my money. $70 is a steal for the value proposition. “Price is what you pay, value is what you get.”
TLDR turnkey for some, diy for others, based on what you’re optimizing for.
From the piece:
> The Aux-2 doesn’t do anything you couldn’t do already with an ESP32 board and various accessories, or even other ESP-based devices. The Apollo Automations Cast-1, for example, costs $35 and supports the same music-casting and Bluetooth proxy features I’m actually using from the Aux-2 at the moment. You could bodge together something even cheaper with DIY parts.
> That’s not a knock against the Aux-2. That’s the whole point. The ecosystem is open by design, which I find endlessly refreshing. Both Nabu Casa and Apollo fund the Open Home Foundation, which develops Home Assistant and ESPHome (and Sendspin and Music Assistant), and both provide hardware that’s a little easier to use than DIY options. But there’s nothing stopping you from making your own devices. It’s actively encouraged.
Fair enough, I’m not here to tell anyone how to spend their time and money. I have 2 of their voice satellites because I was too lazy to solder that much haha.
In different terms, Google IPOd at a market cap ~ 10% smaller than Ford’s at the time.
Anthropic is at 50x Fords market cap. It’s over half of Googles market cap, and Google is a competitor in the same market with a dozen already successful lines of business.
They are incredibly highly valued in a way where I don’t even understand mechanically how they could grow like that without the entire financial system crumbling.
Google is up ~ 6,600% since IPO. Apply that to Anthropic and they’ll be worth something like $132T in 2048. As a point of reference, the Congressional Budget Office is predicting the US national debt to be around $102T in 2050 (paper was from 2020, might be outdated).
Anthropic, the singular company, will have to grow to be worth 130% of the entirety of the US’ debt. Napkin math says their revenue would probably have to exceed the US governments tax income. Like genuinely, Anthropic would have to become a world superpower all on its own.
They would be so big that they’re an existential threat to whatever country they call home. They would be such a large portion of the economy that central banks would lose any kind of control (presuming Anthropic doesn’t issue their own currency, they’d be a larger economy than most countries).
I can’t see how that could happen mechanically. The world would burn to the ground even trying to make it happen. If you thought Standard Oil was bad, this would be that but for every product and service in existence and all run by the same company.
reply