> I think it's clear that inference is a viable business model.
Only if you also have the model thats better than anyone else's.
As soon as models are free, or there are no newer models (assuming thats going to happen, and thats not a given) then the only thing you can compete on is price.
This means that the only thing you have to differentiate is either price, speed or ease of use. (or regulatory capture...)
We are at pets.com level of spend currently. Unless model development becomes cheaper, then we are going to run out of novel debt but not really debt mechanisms.
But I also think there are multiple ways to differentiate. There is at the very least: "intelligence", price, latency, throughput, reliability. It's not clear to me yet what this looks like, but maybe there is also a services and integration level of differentiation. And then there is the universal stuff: sales, marketing, branding. And then on the other side of the ledger there is operational efficiency, management capability, cost of capital, that kind of stuff.
I mean, there is no kind of "model quality" difference between AWS and GCP or between Delta and Southwest or between Wal-Mart and Costco, etc. but all of these businesses remain viable in very competitive markets.
I totally agree that the level of investment / capex is not sustainable though! But I think what's going to happen is that it is not going to be sustained, while AI continues past that point as a viable business (but maybe with different specific companies leading that industry).
I mean, the obvious analogy is to other llm hype cycles. When chatgpt came out, everyone wanted to figure out how to use it for everything. Turned out it really was good at a lot of things, while still being overhyped. Same thing when chain of thought models hit the scene. Same thing with coding harnesses. Same thing now.
My base case is that this will probably be pretty useful, and also not as useful as the current hype suggests.
there's some nugget of usefulness to it as an idea, but overall jev itself for some reason smells like a scam. scam in a sense of an engineered marketing push towards some bs that will turn into a saas. technical merit is next to none.
For english people, as in actual english not from the UK, its a historic, unless its for emphasis.
There is/are some religious overtones, that are infused into this. Catholics used to (or still do) say "hay-tch" for h, where as non-Catholic would say "ay-tch". However that distinction is not so acute in england(unless you're in/near liverpool), but it might be in scotland and its certainly there in NI
It's interesting in NI because haitch-vs-aitch is so much more obvious than any of the other common shibboleths that it's become almost a comically exaggerated thing - something that people feel free to comment on or even joke about in a way that they wouldn't about anything else.
(Derry vs Londonderry might beat it for sheer obviousness, except for the fact that no-one, even the most hardened Unionist, actually calls it Londonderry in everyday speech unless they're making a point... or work for the BBC)
But, in reality, I think it originally comes from Hiberno-English vs Mid Ulster English, the two main dialects of English spoken in Ireland, rather than anything to do with religion. The association with religion came with the post-famine wave of emigration to Britain, and it was only after that that it began to filter back to Ireland.
(That's the story I was told at school, anyway...)
Aha, this is interesting, because I didn't know the origin of it. The only reason I'm really aware of it is because some of the children I look after go to a catholic school, and say haitch really obviously.
> even the most hardened Unionist, actually calls it Londonderry in everyday speech unless they're making
thick NI accent the SIX COUNTIES! (at this point I would then quickly withdraw to a safe distance and hope I haven't set off a riot.)
well not specificlly but its bound to it. Having durable history stored in an SQL database (with secrets filtering and option syncing) you can search global history, or local. (as in if you have two terminals open, you can just search the commands you've typed in one terminal for that session, that session plus history or that session plus what you've just typed in any other session. )
> start talking about where administrative access and encryption keys lived.
As soon as you start specify technologies, rather than "sovereignty" you end up needing to created specific legal tests to stop people getting around it.
"Data must be stored domestically" is a short hand for being held in the same legal jurisdiction. This means for somewhere like the UK, you get all that battle tested data protections law for free. (new laws require case history to be reliable. Ie, prosecuting under a new law is hard, because if its on the edge of being legal, it can create a precedent that undermines the entire law)
In civil code places, its different, but I don't know enough to offer even a half arsed opinion.
The reason why jurisdiction is important is because if you are storing data outside of your legal protection, when something goes wrong there is little you can do to discourage fuckery.
This is the problem with blinkered engineering thinking. Yes geographically distributed data storage is good. But as you also know, storing it in place with lots of other data, means that its a target. The more places its stored, the more physical security you need. This means that there is higher chance of people being bribed.
Its not a binary, its a multi-dimension graph, with no one answer. Every dimension has a tradeoff.
UAE's tradeoff was: not even trump would ignore all the wargaming that clearly shows kicking iran in the nuts would have inflation rising consequences
1930s modernism was about new materials to allow luxury for the modern man (electricity! large windows! electric oven! washing machines! fridges! indoor toilets!)
1960/70s modernism was about creating new communities, adapting land to fit the modern car based world. No more pokey little shops with only a few items, requiring a trudge along a wobbly high street. NO! drive your car, go to a futuristic supermarket only once a week.
I don't think we see a modernism now, not as a movement. I will now divert to UK building, because in the US its different. Part of the reason why thats the case in the UK is that councils used to build houses, facilities and infrastructure with their own money. This meant they had their own architects. This lead to town plans that were coherent, if not always successful. See Birmingham, milton keynes, roehamton.
Housing wise, council houses were built, or should have been built to a standard. They were, especially in cities much bigger, brighter and luxurious compared to what preceded them. Were they pretty? no, modern is never pretty, because its not what you grew up with. When I was a kid everyone _hated_ victorian terrace houses, because thats what poor people have. They were more pokey compared to executive estate housing, and in some places didn't have a drive for your four door saloon.
Now, those estates are meh and terrace houses are very expensive, despite being cold, draughty, dingey little shitboxes with odd toilet arrangements. (unless you've slapped in a loft conversion and extended over the kitchen/outdoor toilet)
> Dieter Rams as some aesthetic pinnacle rather than soulless mass production.
The problem is, Rams was about bringing the complex new object into the house in a was that was reliable and simple. If we were all about Rams' vision, then we'd have got rid of radiators, had decent windows, insulated the place, had lots of cupboards and cleanly separated living spaces. Not an everything room and a bedroom with inbuilt mould farm (shower + shitter)
Yes the modern apartment is inspired by loft living and dieter rams, but that requires 100sqm/1000sqft and high ceilings. none of which most places have. So you get a kitchen diner because it allows developers to shave off one room from the floorplan and increase density by 1/3rd without extra mateiral costs.
But thats not modernism, thats poor fucking standards, and the lack of competition in the uk in the form new council houses. (a 3 bedroom council house will be at least 90sqm with seperate kitchen, living room)
porting an OS to a known platform which has been done a tonne of times before is very different to trying to create a power efficient, secure and fast OS on hardware that is not yet locked down, let alone released.
Moreover the interaction between firmware and the OS is a massive headache to overcome, especially if you are dealing with battery powered devices.
OpenAI want to make smart glasses, but can't do that yet because its not possible.
So they are almost certianly making an alexa type thing with cameras. So they can build a device capable of recording the entire context of a household.
Because the central limitation is still mass; a friction-fit frame light enough to sit on your face all day without fatigue is the market expectation for eyewear, and devices generally require parts which weigh more than glasses.
They already exist. Glasses used to have heavy (at high prescription numbers) glass elements instead of plastic. HUD glasses can already fall in that weight envelope, with better weight distribution.
World locked rendering isn't a requirement for smartglasses. Some existing ones in glasses for factor do have more limited visual positioning systems. In more obtrusive formfactors they have positioning suitable for world locked.
Of world locked rendering is just a huge waste of power at this point.
for assistant glasses to be useful, you need to know who, what, when, where for each operation that the user cares about. In the experiements that I ran, that took about 35 watts continuous on the headset (using a jetson) and at least 2 big GPU machines in the backend. This was in 2024, so perhaps its not quite as heavy now. but its nowhere near 130mw
> Agents need packages like the rest of us. Ruby gems, npm packages, Maven, pip, docker images..
Yes, yes they do, but read through artifact proxies are dodgy as fuck, which is why and facebook (and I assume a fuckload others) don't have them.
Also semi-airgapped labs are a lot less expensive than you think at that scale. Once you have to do multi-region VLANs with machine certs before you get access to juicy VLANs, the difference between "no internet for you" and "mostly airgapped" falls to almost zero.
Also I would want an artifact mirror because a) that give a good signal about how the model reacts, and what training material its latched onto, b) it hides what the models are doing from the outside.
Only if you also have the model thats better than anyone else's.
As soon as models are free, or there are no newer models (assuming thats going to happen, and thats not a given) then the only thing you can compete on is price.
This means that the only thing you have to differentiate is either price, speed or ease of use. (or regulatory capture...)
We are at pets.com level of spend currently. Unless model development becomes cheaper, then we are going to run out of novel debt but not really debt mechanisms.
reply