everforward

No, the purpose of the company is to employ people. That's _why_ it has a definition; people like the efficiency of specialized roles that let them buy more stuff for cheaper via wages _from their job_. They're a legal fiction we all engage in because of the benefits. They can choose not to provide jobs, just as we can all choose to stop believing that "Amazon" is a real entity that can own things.

There's a give and take here that your analysis refuses to acknowledge. "Meta" is not a physical entity that can protect its property, it relies on humans engaging in the ruse (for money to be clear, ie a job).

> Regarding predictable patterns, note that managers are incentivized to do busy work rather than report that software is "done".

This is an abject failure of the executive levels then, given that managing that sort of thing is their role. More or less brings us back to the same point of "why do we all have to suffer because the C-suite can't do their jobs?".

I’m not actually convinced they’re a good hedge because the market cap seems to presuppose a lot of success.

Phrased differently, I don’t see how these companies could grow their equity like Google or Amazon did way back when without having full AGI and changing the entire world. They would need to become worth more than the entire world’s current GDP added up.

There’s a big range on the spectrum where AI does well enough to mess up your life but not well enough to make you even moderately wealthy.

Super cool! One thought that would make this way more useful to me (and probably a thousand times more complicated, sorry) is if it had a way to split out “GPIO” subroutines from the main app, and schedule them separately. So GPIO would become an addressable unit like pods or load balancers in k8s.

The idea is I could specify something like “I have this app that needs to pull data from a GPIO labeled as a ds18b20 and outdoors” and it takes care of scheduling my app, starting a ds18b20 reader, and networks them together. I never need to know if they’re on the same esp32, separate ones, or in a crazy alternate world the compute bit goes on a VM on a laptop with no GPIOs and the sensors go on a real esp32.

Having GPIOs be an addressable resource would be so cool. Like if nothing else, I could start a new esp32 project without wiring up a breadboard. Most of my stuff is some combination of the same 10 chips anyways.

And understanding that the marginal cost of their product rounds to $0. Refunds are more or less free, especially if people will take Steam credit back.

I also work in the space.

I think you’d be surprised if you look at the younger (post-2000-ish) companies. Specifically about this part:

> If you run simple workloads, it’s probably not necessary but then you’ll probably be paid for that.

You would be surprised how many workloads either are simple, or can be ham-fistedly shoved into a “simple” architecture.

Eg I’ve seen a lot of Postgres-on-k8s solutions that are basically just Postgres on bare metal with a bunch of useless abstractions. It writes to local disk, moving the pod to a new host is basically restoring from backup. Wrap it in some tool that ingests YAML and call it a day.

k8s is sort of the king of this. I hear people say they understand k8s and I can’t help but furrow my eyebrows. I don’t even pretend to, there are so many insanely complicated edge cases.

Eg if I’m sending UDP packets to a service and scale the deployment from 2 pods to 4 and the host for one of the new pods crashes and comes back up with a bad iptables rule, what are the range of potential behaviors k8s would show? What’s the difference between blocking say all traffic vs only traffic to one or two of the etcd instances? What if it’s blocking half the hosts in the cluster?

Storage has its own weird vagaries around single vs multi mounts and race conditions therein.

k8s is the king of “I don’t have to know that, the platform will do it”. I’m not sure I’ve ever heard anyone outside networking talk about whether a pod will saturate a NIC. Just chuck it in the cloud and let it figure itself out.

I’m personally using it to augment my coding more than replace it. I find joy in it.

I use my agent to prototype a rough version that mostly works, and then I iterate through (with or without my LLM) to clean it up. Then back to the top with a new feature until I’m satisfied with the project and code.

I do tend to like it. My real selling point is that I’m not “locked in” to any choices until I’m happy with how it looks. I used to have to build the thing to see how it looks and acts, and going back on those choices was painful.

I do think there will be value in it, but I’ve been wrong before. My agent occasionally makes stupid choices that make every prompt after that 4x as hard to fix. Like I screwed up a pydantic model in an unreleased library and it supported a “legacy” path no one would ever use for like the next 5k LoC. By the time I noticed, legacy handlers were like 25% of the codebase.

That’s gonna burn a looot of tokens if you scale it up to a million line of code app.

That’s what companies hire for. I’ve learned those skills and no one ever really wants them.

If you build tools with that context, they’ll have to replace you with someone that has that context. There’s few of them, and they’re expensive.

It’s easier to hire folks that should be interns or juniors that know that one tool, and then replace the whole department every decade.

It’s why everything is YAML now, despite it being horribly verbose and clunky. Can’t have our people needing to know how load balancing works, just add another 95 lines of YAML boilerplate so they don’t need to learn 5 lines of nginx config.

Coming from the Ops side, “full stack” usually translates to “our entire infrastructure stack is managed AWS services, we don’t know how they work”.

It turns out about as well as if you set the Ops team to shipping features.

Eg I’ve seen more than one full stack dev team build a synchronous RPC-over-message-bus framework because they can’t figure out how to make HTTP or actual RPC calls reliable. It is hands down the most poorly and overengineered thing I have ever seen on a service that takes single digit qps.

The ops teams fight that shit tooth and nail because if it blows up, they’re the ones who get called. And it is 100% too late to fix it by then. Ripping it out will take months, and the dev team will say they’ll do it, but they never will because the VP is leaning on them about some feature.

You either say no early, or you are forever on the hook for whatever abomination the intern has cooked up.

I don’t disagree there are people doing good things, but they’re doing them for bad people. Doing good things for an evil dictator doesn’t remove that they were done for an evil dictator.

Also, practically and sadly, they’re precisely the people you’d lean on when you want change because they’re more likely to funnel that pressure upwards in the form of quitting or demanding raises. The ICE folks are “true believers” and I suspect their faculties are immune to the power of logic.

It seems reasonable to me to ask, since otherwise they have no way to ask you to stop breaking the site.

There are better ways to do this, they just haven’t been built yet.

Eg this may not need actual contact info. Wikipedia just wants a way to ask you to stop. Anthropic could add an opaque agent/session identifier that Wikipedia could flag, and Anthropic just forwards that on to you in the UI.

If you don’t have free text fields it’s usually not super useful for spam. There are ways to get a binary “good/bad” signal back without that though.

That’s why esp-idf has components. I have a bunch of components that are basically “compile if this SDK config is set, otherwise pretend you’re an empty file” using macros.

Then I just pull in all my components. The ones I use will get compiled, the ones I don’t won’t.

Mostly works fine. Eventually you will run out of NVRAM to save binaries on, but another esp-32 is like $3.

I’ve said this other places but an x64 processor is probably better value these days unless you need video and GPIOs on the same device for some reason.

Even then, you can get a USB to GPIO adapter. I don’t use RPis much these days and it’s kind of a shame because RPi + Python usually has a quicker dev loop than C + esp-idf. I need 5 of some of them though, and I’m not buying a dozen RPis these days.

My not is that having a slant is perfectly acceptable, having tech companies so large that their comment policies affect the Overton Window isn’t.

I don’t have to care if the Washington Post or Financial Times have a political leaning, there are alternatives. Less so with Instagram. You can pry Facebook out of my mothers dead hands when that day comes.

Agreed, but it's worth calling out that some niches more or less require resin.

Anything small and aesthetic is going to basically require resin. Warhammer minis and jewelry (tiny filigree) come to mind.

FDMs have gotten better, but you're still not going to get those fine sub-millimeter details like you can with resin.

A lot of the extant harnesses are built around interaction, which adds a bunch of complexity.

Eg an agent that only runs off a webhook trigger doesn’t have a use for all the UI stuff Pi does (and you may not want the local state like Pi does by default).

I’m still sort of with you, I’d generally do ACP to an open source agent. You do get much finer grained control if you use one of these agent SDKs (there’s a bajillion of these as well).

It’s fine and safe, it just feels a bit silly. Yes, I could have an LLM read a ton of stuff and generate a list of suggestions on the fly.

Or I could have an honest to god farmer make that list now, ahead of time. Regular search could absolutely surface that.

But also, as someone who does hydroponics, you’re going to end up reading the papers anyways unless you’ve done this a lot. It’s not easy to pick out something like low iron from too little light.

Some movies and some games (or parts of games) really are just for fun. Eg a lot of Bethesda games would improve if you could do a randomly generated storyline on your 4th playthrough, and WoW has 0 storyline to their dailies. It could be kind of fun like a fortune cookie to get a random story. I don't expect art, but it'd be funny to see if it made up a meme-able story.

An awful lot of mainstream movies do have fairly trite plots, with a predictable film style. That's not entirely hate, sometimes I really want to see Michael Bay blow up half of LA. I don't think the next Fast and Furious movie would suffer horribly from AI writing, it's never really been an "art" kind of series anyways.

> For that to be made stable, you really need a whole 3D model of the scene, and the cost increase there might be huge.

I suspect it is, because the 2D -> 3D models I've seen are terrible, but more importantly the 2D -> 3d models are terrible. Noteworthy because I suspect the number of people who have 3D scans laying around is tiny.

What you're describing is vaguely how video games work, to my understanding. There may be interesting parallels if that becomes a technique (like how distant objects are sometimes still rendered, but at a lower resolution/polygon count).

Their infringement is different, and would really be a trademark issue if we hadn’t just invented the idea of copyrighting a character separately from the story they are a part of in the ‘30s (they are not individually copyrightable under the original text).

There’s a substantive difference between ingesting other people’s work and hosting copies of fans knockoff models of their favorite characters.

It’s even weirder because in other mediums this wouldn’t even be a thing. Nobody gets nearly as up in arms about fans doing their own cuts of movies, or re-using characters in written fan fiction, or etc. Very odd to me how up in arms people get about this specifically for visual art.

I think you’re misunderstanding me; I’m not saying places will use an LLM instead (whether cloud-based or on-premise).

I’m talking about them training their own models with a much more specific purpose.

You brought up choosing a menu option in a UI. Jev is really good at something like “find the menu option that best fits the users prompt”. That’s general, and requires a big complicated model.

Companies are usually more interested in something less variable. Ie the menu option they’re looking for is fixed. “Take an arbitrary website and find the Login button”. That’s a _way_ simpler model, and well within the abilities of many large companies to train.

You only need to pay Jev if there are so many variables the model basically has to be general.

Eg classifiers are a separate thing from LLMs and are literally only capable of structured output. It will output one of the X options it was trained to recognize.

I think some of this is that the Venn diagram of people who like AI and the people who like machine learning isn’t a circle. There’s a lot of “intelligences” out there that aren’t an LLM or decision model.

I suspect they’re ignoring it because there isn’t money in it.

Jev is super cool, and it’s super cool that it can handle novel categorization. Most organizations aren’t spending the big bucks on novel classification, though. They’re doing the same thing a billion times, which means Jev has to compete with having a custom model built and doing your own inference.

It’s very hard to build something generic like Jev, but it’s fairly manageable to build something more specific that handles a single kind of task. Especially at the scale where OpenAI would care about your attention.

I don’t really understand neural networks and I managed to train one to do hand gesture recognition on an old Leap Motion. It’s a pretty limited domain, but it trains on the CPU of my decade old laptop in under 3 minutes. The accuracy is quite good too, especially when you discount errors from the Leap (like misidentifying which finger is extended).

It almost definitely won’t. I don’t think I can ever recall a corporate contract getting cancelled by a vendor because of a rogue employee.

They only ever get cancelled over money

Lately I’ve been starting to wonder if the world has changed in a way where long term planning doesn’t work anymore. If information and the reaction to it propagates so quickly that the world changes faster than your plans can adapt.

A long term plan for a company is _at least_ a decade. If your plan ended today, it started in 2016 so it has had to survive COVID, inflationary periods, several trade wars, the rise of AI and now the side effects of the war in Iran.

If your long term plan has to adapt to those in real time, is it even a long term plan? Or is it really a series of long term plans cut so short they became short term plans? I can see companies adapting to just drop the long term premise entirely.

I agree about the rationalization, I tried to imply that with air quotes.

> municipalities paying the full cost of policing

Since we're talking realpolitik, this will never happen. In a first principles sense, it won't because it's unpredictable at the locality level (how do you set tax rates for FY27 without knowing what you'll pay out for this?). In a realpolitik sense, it won't because a) people don't like their taxes going up and would rather bank on it not happening to them, and b) the police will throw a shit fit and refuse to work when they realize they have to consider property damage while doing their jobs, and that the public will be _pissed_ if they crash a cruiser into a parked car while LARPing John Rambo.

> Ultimately we need out of touch voters to realize that unaccountable government can easily turn its sights on them, as well.

Good luck and godspeed. The latest stats say 56% of US adults read at or below a 6th grade level, so say it slow and in monosyllabic words lol

Not in a criminal court, or really anything where the government is the counterparty. Philosophically it’s because those things are part of your “civic duty” as a citizen like taxes. Realistically it’s because the demographics of people getting the money would be unpalatable to the public. Recidivism is high, some large portion of it would go to people with prior convictions.

The absurd end of this spectrum is what happens when police damage a random persons property while chasing a suspect. They’re generally not liable for it, however most homeowners insurance won’t cover it.

I remember an early GoFundMe from a lady that went to work and a guy broke into her house and had a shootout with the police. She came home to being homeless (house got condemned because plumbing, gas, etc is bullet-riddled), and it’s a really disconcerting story to hear. I don’t recall if the city settled to quiet the story down or not.

Yes, this kind of thing happens a lot. The incentives are misaligned with the goal to a degree where you almost wonder if it’s intentional.

This happens a lot because at the individual level, it’s sort of in everyone’s best interest to take the deal. The prosecutor wants $2,000 and 0 jail time and it all goes away. If you don’t take the deal, even assuming you win in the end, it will probably cost you way more than $2,000 and however long the probation check ins take. You will recover neither your time, nor the money you spent on lawyers/transportation/days off/etc.

At the macro level it happens because prosecutors get judged on their conviction/case closure rate, because we implicitly assume they found the right person. Browbeating people into accepting token punishments that technically close out the case counts for the metrics voters actually watch.

If I might lean a little conspiratorial for a moment, it also behooves the prosecutors office to flood the defense attorneys with cases. People will vote to raise the prosecutors funding to deal with criminals, no one will vote to raise defense attorney funding. Last I saw the average prosecutors office gets something around double the funding the defense gets.

It got so bad in some city in Oregon there was (is, really), a near constitutional crisis because the city got trapped between the provisions of the 6th amendment. Defense attorneys were so flooded they said they couldn’t take any new cases. The 6th amendment says the accused has a right to counsel, which is an issue. It also says they have a right to a speedy trial. The city was going to have to allocate money to the defense and get them lawyers quick, or drop a bunch of cases.

That never happened, the judge basically just said “Nuh uh” and I believe threatened the defense office with sanctions if they didn’t just assign the cases to lawyers who were already saying they had more cases than they could imagine.

I’m a little shocked we haven’t seen a huge wave of successful “ineffective assistance of counsel” appeals. Pointing to a signed, public statement your lawyer was overworked before you even talked to them seems like a decent starting point.

This doesn’t particularly make sense to me, can someone help elucidate why this would impact Chinese stocks more than an equivalent company in different geography/government?

Is this related to the US govt banning internal DeepSeek use? A currency thing and prices are USD denominated? Chinese AI stocks were used to hedge and now aren’t?

I don’t see a purported reason, and it makes me feel like the article doesn’t want me to ask that.

H100s are renting out at a few dollars per hour. An awful lot of people could run them if they wanted.

OpenRouter's DeepSeek provider is basically this. Instead of renting H100's for dollars per hour, you pay someone that arbitrages the $/hr to $/token conversion because you can't actually saturate an H100 for hours. It's still effectively the same as running it yourself because of the ability to change providers to someone else arbitraging the same thing, and that there's always the escape hatch of switching to actually renting the H100s yourself.

When it solves your problem. I don’t think it makes sense to globally qualify a model as “good enough” any more than it does with programming languages. Python is good enough for a web app, but not for an OS kernel.

To your point, I think we’re past the hump on the Pareto distribution. The models are good enough for most things, and the rate they’re improving at the things they’re bad at is dropping. The odds a new model is good at something the old one was bad it is okay, the odds it’s something you care about is dropping. It makes a lot of sense to discriminate on price rather than features in that environment.

I’ve never heard anyone say raw inference was unprofitable. It wouldn’t even make sense, I can’t see a rug pull for renting out commodity GPUs.

People are saying _training + inference_ is unprofitable (I don’t have a dog in the race and don’t really care either way). That’s a lot of GPUs and salaries that aren’t generating revenue at any moment.

Personally, I would guess Anthropic and OAI are profitable. I just suspect that they’re a pretty normal level of profitable and not the absurd, near-alchemical “can turn dirt into gold” level of profitable their valuation seems to have priced in.