You should Google how often the world mocks the UK for their "heat warnings" or whatever you call them over there, rather than listing all time heat records. Great job pointing out the us and uk use different measurement systems, literally nobody knows that, so it was very useful to learn today, thank you for that deep insight. You will surely be the person to turn around the uk's declining fortunes. Stiff upper lip, I'm rooting for you.

Whats "extreme heat waves"? In england a "heat wave" can be 75F, in Phoenix it's 120F. You could have three global warmings and heat waves in england would still struggle to crack 88F. In southern california 88F is a tuesday.

You should check out the latest state of the art stuff coming out of Google and OpenAI; they're releasing stuff that runs just fine on an 8gb consumer gpu and performs as good as summer 2024 models, with offload to system ram through new architectural techniques. A "20b" model will run performant on a 8gb card these days, provided you have 16gb system memory. You don't even need a good or modern gpu just wicked fast VRAM. Google has released some gemma and Gemini models in the 270m to 4b (that 'm' isn't a typo) that are functional for running lights, thermostat, kitchen timers etc that can and do run on raspberry pi hardware now.

And yet people don't feel the need to take their TV remote to a shop in the mall to pay to swap out the two AAA batteries in the back because you can open it with your thumb and swap them out in seconds.

The devices with screens (that can show you ads) tend to redirect you to an external website, whereas the hockey puck screenless devices tend to answer the question directly. We don't use the screen devices for much beyond turning lights on and off anymore as their answers involve looking at the screen. I'm looking forward to switching to a separate private hosted llm solution as soon as good hardware becomes available that's not a raspberry pi in a 3d printed case

yes and this is already done it's called speculative decoding. it's mostly applied to allow smaller models to write the majority of tokens (words like, the, and, if etc) and then those are fed to a larger model to check. checking tokens is way, way faster than generating them.

A lot of these smaller 3-12 person companies will develop some proprietary tech on top of (probably agnostic) state of the art models and rather than sell the product they just maintain it for their existing customers as a professional services company. Something that solves a difficult to solve problem, but has to be uniquely wired into each client's system differently, and then add/tweak features per customer.

> It also shows me that locally running LLM are not nearly as bad as some portray these to be.
 
Well, also, things have improved a lot since last summer, and there's been a lot of work lately on using archetctutal strategies in flagship models, in small and micro models. A 270m model today is about as good as a 4b model was two years ago. Back then a 4b model couldn't write a haiku or sonnet, now a 270m will at least make an attempt on par with a 6th grader even if it isn't exactly perfect. 270m sure seems like the lower limit to get coherent responses from and do basic analysis. But it's enough to do tasks like check the time weather and date, set timers and alarms, and know when it's out of its depth and hand off to a larger model. 270m is probably just small enough to run on a raspberry pi with half a gig of RAM

You can self host a pretty competent llm on an 8gb gpu and a real competent llm on a 16gb these days. A 30b model might only hold a couple layers in gpu memory but still be fast enough for daily use. Pretty soon if you have 64gb system and 16-32gb gpu you'll be able to run in a year or two summer-2025 state of the art models locally. You can already run Winter 2024 state of the art models on consumer hardware. The only thing self hosted doesn't have access to yet is search tools

If a competitor offers an ad free solution, at a reasonable price, I'll switch immediately. There's no moat on this sort of product anymore. Anyone can host deepseek R1 or Qwen which are arguably as good as any of the GPT4-era models. And they'll get better.
 
I'm done seeing ads.

Videos need to be at least N length to qualify for different tiers of ads and sponsorship. Yamering on for a minute about what the video is about, for whatever reason, helps keep people through that critical first 30 second period, which boosts your viewership and tells the algorithm your content is worth watching. If the video has over 10k views you can pretty much always skip over the first 30-65 seconds

Yeah there are valid uses for hydrogen. Powering Karen's mini van is not one of them.

Absolutely nobody actually focused on sustainable energy gives hydrogen even a second of thought. Big oil has been pushing hydrogen forever as you can make it with fossil fuels, and requires siloing and pipelines and refineries - all stuff oil companies have in spades.
 
Other than Japan (which has no significant fossil fuel resources domestically) nobody is actually spending a dime on this stuff. It's a total red herring. Do yourself a favor and just forget about hydrogen. It barely made sense in the 90s before lithium batteries were affordable, and today it's just a hollow black hole where funding goes to die.

Usually software support ends before the device becomes unusable. I still have my mother in law's 2nd gen iPad floating around on a bookshelf somewhere but it's basically useless as no apps support the newest OS you can load on it. I've never ever paid for phone or tablet repair. If you regularly need device repair you probably should be buying cheaper plastic devices, not glued together pieces of metal and glass.

I think the public datasets are well established now. There's no need to manually scrape websites in 2025, you just download the stackexchange.all.2025.torrent and then train on that. Lmao nobody is manually scraping stack exchange in 2024 let alone 2025