everforward

I'm doubtful they can, based on current inference prices. An H100 draws about 50W while idling, which is ~$8/month at average US electricity prices. They also draw ~400W while active.

The electricity prices are relevant because if you paid $0 for your H100 and didn't use it a single time, you could buy millions of tokens in inference just on the electricity it draws while idling. If you can't keep that thing saturated through the night, you're probably underwater overnight. Likewise, it's too small to run even the frontier open source models so you need to be able to live with worse models.

Max power matters because you aren't going to run many of those H100s before you blow breakers in most houses. Newer houses in the US are 15A service to non-kitchen breakers, so 1650W (that might be peak rather than continuous, not sure). If you're plugging that into an existing run, you could maybe run 2 before you start blowing breakers? You can't just plug 4 H100s into the wall in a normal house.

Maybe I'm wrong, though. I'd be curious, it'd be neat to run my own inference for something more than what'll run on a 3080.

Land is finite but not meaningfully so at this point. If my early morning napkin math is right, we could give every person on earth about 6.4 acres of land.

Land where people want to live is scarce. Obviously a variety of factors but one of them is that land is bifurcating the way income is. Small amounts are growing in value rapidly and the rest is declining (or growing more slowly). Urban property is skyrocketing, rural property in coal country is struggling.

For big commercial rental properties they probably don’t care because it doesn’t really push the property value.

I don’t think it really changes the property value because those interior walls are usually cosmetic and not load bearing. If the bank has to foreclose, the next buyer will be well aware they can strip those walls out for nothing relative to the cost of buying the building.

I’m doubtful it pushes the tax revenue for the same reason, everyone is already evaluating based on what it could be.

Smaller units are probably just not worth the effort to manage. It’s a lot of over head to start auditing old duplexes where the owner has a property or two and isn’t a giant company with a whole compliance department.

Oh, I also don’t think I’ve ever seen that but I’m not surprised. Even if you could steal a huge amount of storage, filling it with data would take ages and the cat and mouse game of moving the data as hacks get uncovered would be untenable.

I have seen things get hacked for bandwidth, back in the days before you could rent a gbps uplink from the cloud for $0.12. Some scene release groups would hack into universities or companies to do the initial seeding over their super fast links. It used storage, but that wasn’t really the goal.

YMMV but I wouldn’t personally reach for either. They’re both out of band so you have to deal with pulling input back into the context and what to do if one side won’t communicate right.

File edits are just tool calls under the hood. If you’re using a decent agent then you should be able to override or extend the filesystem tools. If you’re on ACP, file reads/writes get proxied to your ACP client and you can inject your hooks there.

It’s pretty trivial to implement, this is well within the bounds of things most agents implement (for open source agents anyways, no idea how to extend Claude Code or Codex these days).

Huh, that is odd. Working backwards, that would be ~ $0.40 originally. Wonder if that’s also flat out wrong or if they’re doing some kind of currency handling that breaks when you start dealing with huge multipliers.

Storage-based billing is huge, unless you mean something other than “places that make you pay for storage separately”.

Also many places I’ve worked, storage is a huge part of the spend but that depends a lot on what you do. e-commerce doesn’t use a ton of it, but if you handle user-generated content or do any kind of training (LLM, computer vision, etc) then you can very much end up in a place where storage becomes a top line number for infra spend.

GitHub pre-Copilot was probably like that. They host a shitload of data, most of which is just at rest the majority of the time. Storage and networking are probably the majority of their infra costs.

Yeah, this one is bad because it’s off by so much I’m shocked it wasn’t caught by tests, alerts about unusual changes in the billing system, or even accounting. Like surely the P&L reports look all kinds of wrong right now, they have to be showing like 6M% profit margins and revenue measured in quadrillions.

I’m also a little surprised this didn’t trip a circuit breaker. For something as non-real-time as billing, I’m surprised they don’t have an automated kill switch that pauses the billing system and fires a page if variance in bills spikes. Naively some kind of “if the standard deviation of customer bills for this year changes by more than 50%, pause the billing system”. At that number of customers, those numbers should be pretty stable beyond internal billing changes they could normalize for.

Someone said the numbers are all off by 2^30 because they screwed up and are charging the per GB price for each byte.

It’s probably an artifact of them all being currency multiples of 2^30

This seems like an odd request to me, but maybe I'm misunderstanding what you want or what Juggler does.

To my understanding, Juggler is a frontend to your agent and would replace your editor. That's the selling point I got, a better UI than text for orchestrating your agents.

ACP servers are agents (like if you look at the spec you have to implement an entire agent's worth of endpoints, from creating sessions and sending messages to handling starting MCP servers the client specifies).

I don't get the sense that the point of Juggler is to make the world's best agent (rather it's to make the world's best interface to agents), but OP would know way better than me.

What you're describing sounds like an ACP proxy, which probably exists but is a different class of thing.

If you just want to drive OpenCode via Juggler rather than your text editor, you also want Juggler to be an ACP client (it can be both an HTTP server and ACP client, I have some stuff at home that does that).

You can drive agents via ACP these days, which I think is the layer you would want for what you’re talking about.

The API is basically what you see as a user of Claude Code or Pi or whatever. You can make new sessions, send messages to sessions, configure which MCPs get started, etc.

I’ve been poking at something similar to what you’re talking about via that route. My client prompts the agent to do a thing, and then afterwards launches deterministic things to check it which can either re-prompt the original session or start a new session.

Eg it automatically runs the tests afterwards, and will send a new prompt in the original chat to fix them if they fail. I also briefly poked at a security analyzer that gets changed files via git and makes a new session to check whether there are security issues and propose a fix that then gets sent to the original session.

If you want a circular loop where the LLM can adjust its own workflow while keeping it deterministic, you can let the agent modify the ACP client that drives it.

ACP support would be a big deal and would make this a real contender for me. It looks super nice, but the idea of re-writing my Pi plugins is a real blocker.

I’m generally happy with my agent and want to keep that, but I do think the UI could be better and this looks like a neat step that way.

Those things are less static than you seem to believe. Eg recurring prescriptions; at least once a year my insurance gets mad about my scripts and requires a pre authorization that occasionally contains a stipulation that I try their “preferred” (read: cheaper for them) alternatives.

My planning is now wrong. Likewise I can only do FSA for planned procedures if I know close to a year ahead of time, which isn’t super likely. What portion of procedures are serious enough that insurance won’t call them elective, but also tame enough that you can put it off for a year for your FSA?

I don’t think I’ve ever had my FSA be completely correct. I’m always either over-stocked and losing money, or under-stocked and paying with post-tax dollars when I don’t have to.

It seems hard to cure because a lot of this is stuff that just probably shouldn’t be done. Ie the structure of the products veers so close to anti-competitive practices that it’s just untenable in the face of regulatory enforcement.

Google runs the dominant search engine, which they control the rankings of and sells ads on, while also competing against companies that buy ads from them and fight to maintain a spot on the index is almost immediately suspicious. The potential for abuse is incredibly high, and at one point would probably have been concerning enough to invoke regulators without even acting on the potential for misconduct.

It’s like taking the babysitter out to a fancy dinner alone. It could be something totally normal, but it looks bad enough that you probably wouldn’t do it.

The real answer is that Google would probably need to sell off that arm. There is no configuration where Google retains the control or benefits of the Shopping product without being locked in conflicts of interest around the index. It’s always going to look like the way Standard Oil was setup, because it is set up the way Standard Oil was. They own infrastructure, and they compete upstream against other companies forced to use that infrastructure. There’s no way to resolve that conflict of interest.

If they’re bigger folks or starving and someone doesn’t want bone-in chicken, I could see it. 3 large 4 piece combos is $55.50 in Miami, and I think there are other things in that range (eg a 5 pc tender meal if someone hates bone-in chicken so they can’t get a family meal).

The family meals are substantially cheaper than individual meals, if you can get everyone to agree on bone-in chicken and the same 2 sides.

A 20% tip would push that up to something like $66.

I do think it’s a step function improvement, as someone that enjoys those crappy but infinitely-useful tools.

Those tools tend to suck in the sense that they were written for a single person, so the tool assumes that you wrote it and have the context of the author. Help text output often sucks, the UI often sucks or doesn’t exist, it might require 15 packages to be installed that no one has documented because it was never meant to be distributed.

I like it as an MVP thing to trial whether other people would use it and I need to write a “real” version of it, or if I’m the only one that will use it and I can just vibe-code the hell out of it permanently.

You used to have to decide that pretty early on, which meant a lot of talking to other people but also trying to prevent it spinning into a whole project with a PM that will never actually be done.

Dunno, ymmv, I like it a lot in the domain of “stuff that doesn’t have an oncall rotation”. I do find it super useful in that space. Still useful but much less so if the code is something I could be paged for.

I was thinking of studies like https://www.science.org/doi/10.1126/science.adh2586 I may have misremembered the domain, or I may have neglected to remember the publishers. This time around I explicitly excluded any AI-tied sources (I don’t trust Anthropic to have an accurate study of productivity boosts).

While I don't think OpenAI and Anthropic will capture all of the potential upside, I do suspect they will do much better than other players despite the competition

We may have a fundamental disagreement here because I think you see reasons for businesses to prefer LLMs beyond price, but my point was that I think a substantial portion of that top line number will be eaten up by needing to be cheaper than humans.

Like 30% cheaper than humans for the same task isn’t an unreasonable demand for onboarding, and that would lower the addressable market to .75-1.25 trillion. That’s also revenue, not profit. Their best case is probably 30-50% margins (even that might be high, I suspect it will become low margin as models are commoditized), so $600 billion-ish in profit spread across all the players and that feels pretty generous to me.

I suspect a lot of general knowledge work tasks could be done with fairly dumb local models. The small Qwens could probably respond to my Slack and emails for me.

Typically yes, but there are reasons companies may be willing to pay the same amount or even more, such as "AI doesn't need sleep, holidays, insurance, or benefits" and "AI is easier to procure and replace than humans."

It’s also an enormous supply-side risk. Your AI provider could charge basically anything they want and you have to pay it, for the same reason people still pay bonkers Oracle licenses: too hard to migrate away.

Unless companies maintain their ability to swap providers, but then I can’t see why prices for inference don’t crash down to the cost of GPU time and electricity. If Claude doesn’t have some secret sauce that keeps me locked in, I don’t see why I wouldn’t use Deepseek or Qwen or one of the other dirt-cheap inference providers.

I just haven’t seen anything that seems to keep people locked in so far. Even the non-tech people I know that like AI are rotating providers to find their ideal price:performance ratio, many are finding success outside of Anthropic/OAI. I do know a couple holdouts that refuse to use anything Chinese. They mostly stick with OAI or Anthropic, though they do moan about paying more.

They can be impeached, though the efficacy of that is questionable these days.

I'm thinking the future of this tech will likely be better tooling with better IDE integrations rather than "Claude plz make me a SaaS kthx"

I think this sort of thinking is a trap, because it presumes that all software has the same constraints.

There's a spectrum of requirements between "chuck this over the wall at Claude, it only has to work once" and "this is a literal rocket ship, formally verify the whole thing".

I've made some things with Claude I don't understand and don't control. It's fine, they're still useful to me. Things for the house that I wasn't going to build manually, some dashboarding stuff and scripts for work, stuff that can crash and burn and I'll be fine.

They won't justify trillions in investment, but they are useful.

Equally, I do agree with you on some things. Sometimes I hand-hold the LLM or forgo it entirely because I want to be 100% sure I know how something works, and can justify a decision if it causes a production outage.

I think the future is probably multiple different tools with different goals. Better IDE integration for some uses, an entirely separate "LLM herd controller" kind of thing for when you're okay with vibe-coding, and the most interesting is something in the middle where you're more in the loop than pure vibe-coding, but don't see the full context like in an IDE. Something where it surfaces changes to key components, but hides things like test changes.

Note, at 5% productivity boost, humans are not just in the loop, they are the loop. AGI or large-scale replacement of humans is not even needed, but the financial opportunity is already immense, and it scales with how much human productivity can be improved (i.e. how much work can be offloaded to LLMs.)

The studies I've seen recently (at least in the software space) put it at something like a 10% increase in coding speed, which for me would probably translate to something like a 3% increase in productivity. I spend a lot more time on things like getting agreement between teams, documenting approaches to things that don't exist on the wiki, etc, that LLMs are significantly less effective at. Or just can't do; no one will be happy if I send an LLM instead of me to meetings.

I suspect a lot of roles are like that. They give a 10-30% boost to the core role function, but that core role is still only 30-50% of what you do.

that is ~1.5 - 2.5T in value annually

That seems really large, but it's ~2-3x Walmart's yearly revenue, and OpenAI and Anthropic both have estimated valuations that compare to Walmart's market cap. And this is before we consider that they need to do it for cheaper or why would anyone bother. Realistically, potential revenue is probably half that at best.

It's also before cutthroat pricing really kicks in. People are willing to pay for Claude right now; I still suspect that as time goes on people will start looking towards Deepseek/GLM/etc models that provide 95% of the performance at 10% of the price. That'll cut the market even further.

The question is how much demand for knowledge work swells as prices fall, and whether that's a soft landing or a crash.

There’s no way to justify their valuations if they get downgraded to a pair programming tool. They need fully agentic stuff to work and replace human engineers to even come close.

Offhand, I’m not even certain whether a model like that could justify the constant retraining we’re doing on the agentic models.

It doesn’t make a lot of sense to spend millions or billions on training to reduce hallucinations by 0.3% if your model assumes a human is in the loop to course-correct them.

normal people who think expediting the interdiction of stolen vehicles (which are vectors of violent crime) is a perfectly reasonable thing for a city to invest in.

The effectiveness is dramatically oversold. This has literally happened to me. Called 911 midway through the theft, cops pulled it up on Flock, the best they could tell me was "they went towards the highway" (no shit, Sherlock).

I saw the parts on marketplace a week later (definitely mine, unique staining), 2 hours after they were listed and called it in. The PD where the parts are wouldn't do anything unless my PD called, my PD said the officer in charge was on vacation so they can't do anything until he's back.

Take a wild guess who hasn't heard anything back. This should be their ideal situation; they got called mid-robbery, and someone else located the parts afterwards. If they can't solve that with Flock, lack of information is not the holdup here.

It would be so easy, to generate subconcious addds with an LLM. Like weave a slogan reminder into the conversation...

Probably not without changing the underlying response and tanking the benchmark ratings. Like I ask it to generate code, it tries to add a slogan reminder, now the comments on my code have Tide's slogan in them and the variables are all named "bleach".

Probably less humorous than that, but the underlying point stands. People will be extremely upset about their prompts being tinkered with, and any perceived degradation in quality will be quickly attributed to them. The hidden nature of them will probably cause speculation that they do the same on the higher tier plans.

That's not a lawsuit by a gun manufacturer, those are regular private citizens and the "California Rifle and Pistol Association". So far as I can tell, that's not an industry lobby group, it's a private interest group although I'm sure the manufacturers like them.

That changes the above quote. It's not a gun manufacturer claiming they can't find anyone to microstamp their guns (though not dispositive of the same), it's a citizen saying that commercially no one sells microstamped guns.

If you expand out your quote, it neatly explains what I would wager is the reason why no one sells microstamped guns:

"AB 2847 also imposes an additional amendment to the UHA: for every semiautomatic handgun that satisfies the new one location microstamping requirement (in addition to havingCLI and MDM) and is therefore added to the Roster, the State must remove three (3) grandfathered semiautomatic handguns from the Roster, in reverse order of addition. However, this has not yet occurred because microstamping technology does not actually exist in any commercially available application on a handgun."

In other words, if a manufacturer decides to make a new microstamped model then 3 models that are likely still in production will become illegal. It's a losing proposition for the manufacturers. There's no money to be made by creating a new production line to build microstamped pistols if it's going to shut down an existing line; it's a bunch of money to re-tool factories to sell to the same people the shut down line was already selling to.

Without getting too into the weeds of politics, it's exacerbated by it being California. Whether you agree or disagree with the goals of the policies, they do change the requirements for guns very frequently relative to other places. I can understand manufacturers being reticent to set up new production lines to comply with policies that will probably change again in the next 5 or 10 years.

My reading of the lawsuit is that they think this was the intended end state of the bill. Require new guns to implement a feature, ensure that implementing that feature will generate financial losses, and now no new guns get made but you can point at gun manufacturers and say you didn't ban them the manufacturers just won't comply. I'm not totally sold on that, but it is a remarkably poorly written bill when the incentives the bill sets up run exactly counter to the intended goals.

* there is LangChain and LangGraph - used a lot, but framework bloat is hated as well

I've used them a fair bit, and I'm not a huge fan. Only self-hosted, I can't comment on their cloud-SaaS agent runner thing. The observability looks neat, though.

LangChain is nice enough, I appreciate having a unified API across providers. LangGraph is... just not all that much? As in the DAG is too much for a simple agent, but when I start thinking about a large agent and dealing with that flow in their their DAG DSL my head starts to hurt. "Go To Definition" isn't going to help navigate that very well, the state is going to be a lot of Optional's with not a lot of info on when they have a real value at which stages in the DAG.

I had substantial issues with branches in my DAG because the state has to include all possible fields for every step. It gets hard to mentally track all the combinations of fields that will be present or missing depending on the path taken upstream of this node. Do I have RAG results? Not sure, it depends on whether the query includes X, but then later it also depends on whether a tool returns a particular result, in which case it can either be missing, have a single value or have 2 values. Yada yada, in a sufficiently large DAG it gets hard to track those. Things are much cleaner in function world where you can declare "this functions require X and Y, and can optionally provide Z".

I mostly go directly to the API these days, but I'm fairly settled on Ollama. I might use LangChain if I think I'll want another backend, but I also might use OpenRouter. I haven't yet, but it seems cool.

One is a specific pattern, with specific distances, colors, etc.

The other is a human-invented category of objects that lacks consistency in size, shape and design.

“Don’t print this specific pattern” is a much more closed loop than “figure out if this could be part of a firearm, bearing in mind that guns share a ton of parts with perfectly legal things”. Especially since the things that are most indicative of being a firearm are usually bought instead of printed.

Eg a “3D printed AR-15” usually means a 3D printed lower and the rest of it is bought off the shelf of a gun store (the rest isn’t regulated, you don’t need a background check to buy a barrel or upper or etc). So it’s really just a grip, the slot for the magazine, and some holes for pins in the right spot.

That’s super hard to identify. Grips come in a variety of shapes and sizes and have a bunch of legal uses (nerf guns, slingshots, Wii remote holders) and are hard to detect. The mag well is just a rectangular hole? Probably some slots for springs, but you have to assemble it afterwards so the printer doesn’t see that.

The pin holes are probably the best indicator, although they’d be trivial to drill or melt afterwards, or print them as a separate piece and screw it on top of the lower.

It’s just an almost impossible problem. I don’t even think humans could do it reliably with the info the machine gets. Whether it’s an AR lower or a prop/nerf gun/airsoft/cosplay/accessibility attachment (pistol grips have some nice ergonomics) gets into the nitty gritty of the interior and whether they support real linkages or not. I just don’t think we can do it with any degree of accuracy.

That’s without getting into the more avant-garde 3D printed guns that are made either out of hardware store parts rather than gun store parts (common in Europe, barrels and what not are regulated there), or are made out of an amalgamation of parts from other guns (trigger from a Glock, barrel from an AR-15, mags from something else, etc).

Generalizing an algorithm that can detect “things that aren’t a gun I recognize but could be used as a gun” is going to be fraught with false positives. You don’t 3D print the linkages, so it’s just a frame, and there an awful lot of things that could conceivably be part of a gun. Anything vaguely semi-circular could be a trigger, anything with a tube could be a barrel, anything hollow and rectangular could be a magazine.

I don’t see a way that either a) it only blocks current designs and the community quickly adapts around it and continues, or b) it tries to block anything gun-like and false positives spiral out of control.

It feels sort of moot anyways. Phillip Luty wrote a whole book about assembling a fully-automatic submachine gun out of parts from the hardware store all the way back in the 90s. It does require familiarity with hand tools and access to more tools than most people have at home (drill press comes to mind), but you don’t have to be an expert machinist to do it even without 3D printers.

This is all besides the fact that it’s so easy to obtain a real gun in the US that 3D printing guns is a niche hobby. I’ve looked at them and watched some of their competitions, they’re still pretty crap. The fully 3D printed ones still jam constantly and I believe still use off the shelf magazines because those are surprisingly hard to make well. The partially 3D printed ones (eg lower only) tend to be substantially less durable than market alternatives. Unsurprisingly, a plastic 3D printed lower is more likely to shatter while firing or dropping it than a milled steel/aluminum one.

They don’t claim it doesn’t exist, they claim it bears a cost significant to the overall cost of the firearm and is unlikely to meaningfully accomplish its stated goals. Ie prices will go up by 10% or whatever for basically nothing.

It’s a useless place to put an identifier. It can be filed off easily, and checking whether it’s present involves a microscope and/or disassembling the firearm (depending on the gun and firing pin setup).

You can even just straight up replace it with hand tools at the range. More insidiously, if this passes, you’ll probably be able to buy firing pins that are stamped with someone else’s identifier. I have little doubt a market will spring up for falsely marked firing pins.

Or go super low tech and just throw a brass catcher on the receiver to catch the shells. In a pinch, duct tape and a shopping bag will prevent leaving brass behind.

This is just a money grab by the folks that invented the tech. We don’t even do a good job of keeping stuff like select fire switches for Glocks off Facebook, this is unlikely to do anything but raise the base price of a gun.

New stuff does, but most of my peripherals aren't new. A couple are a decade old and probably pre-date USB-C being standardized.

I am semi-frequently annoyed that my laptop has 0 USB-A ports. At least give me one.

Net neutrality is about a natural monopoly - there can only be so many cell towers, satellites and fibre optic cables.

This is a misinterpretation, we could support an absolute ton more physical infrastructure than we have in the wired space (cell towers and probably satellites are limited by spectrum, but still not the physical footprint of the devices).

Fiber cables are tiny relative to their bandwidth. Ignoring cost, if we made water mains sized fiber runs under the sidewalks we could probably get hundreds of 10Gbps fiber runs to every house. And I think there’s still a ton of space to fill with cabling if we wanted to for whatever reason.

The two most significant factors at the physical level are

1: it’s a natural monopoly not because of space, but because building that infrastructure is so expensive it’s unlikely any competitors could emerge. Think about where you are and where the closest peering point is. That run alone is probably millions of dollars and a decade of lawsuits to get easements on the intervening properties to even be able to run it.

2: it’s incredibly wasteful to run parallel lines when each house will only realistically have one set of them active at a time. Few people pay for more than one ISP, it’s basically setting resources on fire.

AI companies are frankly far more limited. GPUs are scarce, I wouldn’t be surprised if we’re already building faster than GPUs get produced. Power is scarce, so far there’s been a lot of hand waving about how we’re going to double our power production. Land is fairly scarce when you scope it down to “land that has enough power access, and usable roads for trucking in materials, and access to water for cooling, and is far enough away that the noise won’t make people riot”.

There’s few implementations of the engine, but many implementations of rules for that engine.

I think OPs point is that many of those rule sets probably don’t do what the author intended.

I would second that, because CORS questions are common and “turn on the allow all pattern” is almost always in the top 3 suggestions.

Semi-related tangent, it annoys the hell out of me that create-react-app (and the newer incarnation) don’t come with an “allow all” CORS rule. Don’t force me to figure out which arcane setting configures CORS headers, I’m the one writing the code, I’m okay with wherever the HTTP requests are going, I’ll set up real CORS headers on nginx for prod.