The number of single family homes has grown about 70% since 1970.
 
I don't know what the average age of a sewer is before it needs it's first major repairs or upgrades, but 50 years doesn't seem improbable. Of course your running costs are going to be lower in the first 20-40 years. But the ground shifts, capacity needs upgrading, things fail etc. US population keeps growing, so more and more sewers will hit that magic 50 year mark (or whatever year you pick) where true Total Cost of Ownership becomes apparent. Europe is more expensive than the US, but their population began to stabilize in ~1985 and they've been paying the "sustainable" cost for many more years. The US will see the same result ~30 years after when the population begins to plateau.

The question is, what happens after the bubble pops? The toothpaste doesn't go back in the tube. The bubble popping is the short term, what is the world going to look like in 2035 when I still have a decade left before retirement, the tooling is stable, and Microsoft, Meta, Google etc have bought the scattered corpses of OpenAI, Anthropic etc.?

I think the bigger problem is that AI models are progressing faster than the companies who have sprung up to fix their problems, can fix those problems. By the time they have a real product, the base LLM model has fixed that problem internally. Or you can just vibe code your own bespoke solution in a lot of cases.
 
A company from 2023, that specializes in data normalization for LLMs is functionally obsolete in 2026. A voice cloning company from 2023 is absolutely obsolete, I can reliably clone a professional sounding voice on my phone today.
 
People click on these links thinking "ah-ha! finally proof that OpenAI is just wasting their time!" when in fact it's just companies that got out-competed by the models themselves.

A lot of AI startups exist(ed) to fix flaws in early models, or add functionality not yet present.
 
Post MCP, post skills, post TOOLS... a lot of this stuff is baked in now, or eclipses the products a lot of these companies were building. Cursor.ai, windsurf.ai have all pivoted to training their own models. Claude Code just demolishes what Cursor.ai was building. Even just regular VS Code + copilot running sonnet 4.5 is as good or better than Cursor was in January 2025.
 
It's no surprise those are money losers, if your product solves a problem that was later solved in the core LLM, you have zero sales leverage. The tooling is still maturing, anything that becomes a popular AI technology that isn't part of the core model.... will almost certainly be baked in to the core LLM at some point down the road. Investing in other than building a core model pre-2026 was probably not a smart decision unless you managed to get out near the peak. AI will continue to grow but I think it's going to be 2-3 more years before tooling matures enough that you can be confident that companies can build real products that are widely applicable to many customers that aren't office products already owned by microsoft (word, excel, outlook, etc).

The las vegas sphere was such an absolute boondoggle and continues to lose money it's suprising there's any management left to throw out new, worse ideas.

How are you going to find another blue chip tech job in nashville? If oracle is the only big tech employer in the area, when they tell you to jump you ask how high. The tech job market in the bay area is not amazing right now but if I got laid off tomorrow, I could walk across the street and be making 80% of what I am right now before my 2 weeks severance ran out.
 
If Oracle lays you off it's an emergency; you have to sell your house and move back to california or some other tech hub, likely 1000+ miles away.

Probably the exec whose passion project was the library, retired. Execs had wanted to kill the library for years but it was less of a hassle to just wait for him to retire, then dial it way back. So now that's happening. Microsoft has a lot of "untouchable" execs from the 90s era with weird pet projects.

For voice assistants it's helpful for it to be local. It turns out that 98% of commands fall into about 10-12 commands (Set a timer for 5 min, turn on/off the lights, what time is it, whats todays date, turn on/off tv, turn on/off the lights in another room). The device catalogs all these requests and then makes a list of the top ~30 requests and if the request matches something on the list with ~0.85 confidence it doesn't even go to the LLM it just runs the command. That's how you get the instant response for turning on the lights etc.
 
A fun trick is to do a complexity score of the request (for when it gets to the LLM) if the device is online and the complexity score is above a threshold (e.g. "explain the long term effects of brexit, and speculate what would happen if germany did the same thing"), it will dial out to a "real" LLM provider and then read back the response; if it's offline it will struggle bus to generate a local response
 
So yeah tl;dr a 7b model can totally work for standard tasks like setting timers, setting reminders, home automation etc. Even the 2b, 3b models coming out of google right now are pretty good; last year the 4b models could give you a 4 day travel itinerary for paris and most of it wasn't even hallucinated. The 7b models from last summer could explain how a microwave worked and how it excites water molecules etc. The models are getting alarmingly good for basic assistant stuff, and with the ability to do basic search are pretty extensible.

Apple announced they were cancelling production of their VR goggles, and as a pressure relief valve to save face, made vague hand-wavey gestures towards using all that R&D money for something in the future.... maybe.
 
If you're still standing up for Apple's VR ambitions, then you don't get it - VR has failed as a mainstream product. The same as 3D TV, 3D Blue Ray, etc. There will always be a core audience of VR users but they are vanishingly small, the number of VR headsets manufactured per year is in the 85,000-175,000 range. On a global scale it's effectively artisinal.
 
If Apple with all their technological, UI/UX might, their infinite advertising budget, could not ship 250,000 iGoggles, vr is dead. Sorry not sorry.

I wasn't aware you could click on the inside until I tried it just now. I never thought about it too hard but always thought of it as grabbing the outside edge and stretching the window like a sheet of rubber or something.
 
My guess is they looked at user data and the vast majority of users do this as well?

This is how the "walled garden" effect evaporates.
 
Amazon makes an alarming amount of money doing advertising, promoting products etc. Sometimes to find the product I want, I need to search and then click through the link, as it simply won't show up in search due to paid placement of other products. Especially name brand stuff from 3M (Scotch tape), Sharpie markers, Rubbermaid etc
 
Amazon is already banning agentic shopping tools from their platform, for good reason, it cuts out their advertising revenue, their upsell revenue, and optimizing of selling items that are closest to the customer to save on shipping.
 
If you had a tool that would auto-cross shop for the lowest price of the highest quality items, and you don't have to browse their website to buy things, it destroys a lot of what differentiates online retailers. A lot of the lock-in is the checkout process. It is two mouse clicks for me to re-order printer paper on amazon, even though I could get it cheaper from officedepot.com but I don't want to maintain yet another account. With agentic shopping, amazon just becomes a set of faceless warehouses with free delivery.

I'm excited for that future, but I'm not holding my breath.
 
That said the steam deck has been a smashing success; I don't have a need for a handheld device but that steam cube is certainly something I'm thinking real hard about, as a companion device to my normal desktop. 95% of my "gaming" these days would easily run at 4K @ 60fps on that device

Countries outside the US and a handful of other high $/capita areas can still get Disney+ ad free, but at half the price or less of what we pay here. Everyone else "pirates" it by sharing passwords. If you don't have someone's password, then you get the free, ad-supported version. There are a lot of kids whose parents make ~$250-$400/month (that's per month, not day/week, common wage for a school teacher in Colombia, population 55 million, more than the state of NY or CA, so not an isolated situation).
 
Those parents are lucky to have a cell phone with internet connection; their kids are growing up on the ad-supported version of Disney+. Based on my experience, kids will watch nearly anything, instead of play with their own toys. Also most of those kids are watching Disney+ on their parent's phone in the back seat of a car, not on a 70" OLED with 9.1 surround sound in a 25x30' air conditioned living room.
 
TL;DR the average viewer is not an american adult

2GB is a lot when you're building 3x TV per minute and shipping them out by the literal boat-load. There are 8 billion people on this planet and more than half of them will buy a new TV in the next 10 years

If JFK was your cousin, odds are you'd probably look good too