That tracks, because I'm currently restoring a 1948 Chrysler straight 8 engine (really that flathead design like what packard was famous for) on a 1948 rolls royce chassis.

I seriously doubt your claim of free speech with a chatbot will hold up in court.

They themselves proved you can't tell a chatbot to not be a therapist. You're not supposed to cut your hand off, there's no regulation telling people to not cut their hand off. You can't create an endless set of rules to tell you what you're not allowed to use free LLMs for. Anyone can train any number of LLM using free open source data sets with any restrictions (or no restrictions) it costs less than $1000 to train a basic 7b model with a credit card at any cloud provider. Next they're going to have signs on gas pumps instructing people not to douse themselves in gasoline and set themselves on fire.

I recall there was "premium" cable and that was HBO, Showtime, Cinemax and probably a handful of has-beens and regional ones that got bought out. I recall regular cable being ~$32/mo and with premium cable (HBO) it jumped up to ~$75 but they only had a few shows and movies would rotate, not too different from how it is now except it's streaming over the internet.

This was largely the purpose of gmail IIRC. Humans can't read your email, but there's almost no value in allowing them access, whereas letting computers build a customer profile to sell ads to you, is invaluable.

One of our engineers did this as a side project back in 2015 in an afternoon, setup a web scraper on aws and the next day we could visit all these things. I'm pretty sure the company did a new article on this... ten years ago.

My guess is phones and silicon will improve to make 4b on mobile by 2030. If not then those requests will get forwarded to xyz cloud service. I can see a world where 7-12b cloud models are ad supported free tier and you either pay or self host 70-600b yourself. I expect processing requirements to drop by half due to whatever breakthrough comes next and then there's a long tail of improvement after that. Token verification was a major improvement.

Yep I login to some of these sites and I see 3-4 tangentally-related things that are clearly algorithm rage-bait, designed to drive user interaction. I'm so tired of this. I login for 10-15 minutes a handful of times a week to check in on friends and family for updates, and then promptly uninstall the app.

Probably in 3-10 years there will be a handful of open, legally copyright free training sets anyone can use to train their own ~600b class model with whatever architecture is current state of the art. Researchers are already putting together 7b training sets like this. And LLMs can use tools like search now so they won't always need the most up to date news or info - they can use tools for that instead. Most of finance is analysis of documents, which LLMs have been excellent at for a while now.

Google announced roughly the same thing, on device models for phones a couple weeks ago at their developer conference. The 1b model is fine for basic tasks like turning on lights, checking email, social media notifications etc and runs ok on midrange phone hardware. The 4b model technically runs but it's borderline unusable speed but it can answer questions like "how does a microwave work?" with moderate accuracy at a semi-scientific level which is impressive. I suspect most devices will be able to run a 1b and by the end of the decade most everything will run a 4b model at least at talking speed. There's a concept that all AI processing will be done in the datacenter, I suspect 80%+ of consumer LLM will happen on the device, and more complex tasks will get routed to the cloud. For a lot of end users (high school students, etc) 98%+ of requests will be on-device.

LLMs are really good at stuff, better and faster than humans, as long as the complexity isn't much more than ~200 LOC (lines of code). 250-300 LOC and things start falling apart quickly. Sometimes (1/50) you'll get ready and it'll pop out 400 LOC without major errors but that seems to be the absolute limit of the current statistical model family everyone is using.
 
LLMs are really good at analyzing and summarizing text though, it has no problem analyzing 20-30 page PDFs of economic or financial data.
 
But yeah there was this idea that if you just kept training on bigger datasets for longer eventually you'd just arrive at AGI and it's pretty obvious via many many research papers that the error limit right now is ~1% and getting below that is really really dang hard. We're going to need a new breakthrough to get the ball further down the field.

Microsoft CFO was publicly complaining during last month's earnings call that the current Microsoft AI chief wasn't cutting it, so they're putting it under this guy instead. Who knows how long it will last; they'll keep reassigning it until it becomes a profitable, or path-to-profitability. Right now Microsoft is just barely ahead of Apple in public AI success, despite being an early partner with OpenAI.

The big issue is they wiped out their ci/cd and more importantly their entire AWS platform the system runs on. Sure you can probably re-deploy the code, but if it's not infrastructure as code then a lot of that stuff is bespoke hand-wired stuff and there's probably no backups for the config files etc. Maaaaybe you can deploy everything on a giant EC2 instance but you still lost all your database backups which means all your customers now have to create new logins, you need to create a new account with your credit card processor (if you can even get them on the phone) test everything.... I sorta specialize in business continuity having something like terraform or other infra as code helps a lot, and we do monthly or at least quarterly wargames where we restore develop from scratch to ensure there's been no drift in the interim. Quarterly end to end test of your backups isn't perfect but it's unlikely you'd be down for more than a couple of days for all but the largest enterprise setups. With something like terraform and kubernetes you ought to go from cold dark black to up and running inside of 12-18 hours, depending on where your offsite backups live.

Evaluation is a lot less computationally complex than generation; I would expect anything that gets scraped by a larger AI company goes through at least a two pass system 1) is this even relevant/compiling code? 2) does this meet a minimum standard? and probably for largest companies building a library of quality content 3) is this high enough quality to include in the library for future training? Steps 1/2 can probably be done with something like a 30b or 70b model fairly quickly and step 3 is probably evaluated by a 300-600b model. Evaluation is something like 10x faster than token generation so it's pretty straightforward.

The movie studios (particularly 900lb gorilla behemoths like Disney, and whatever the name of the company that owns HBO is this week) already own the copyright on enough content they don't need to train outside of their own copyright library. LLM summary of script -> Script of trailer -> shot by shot description of trailer -> 2-5 second "video clips" -> human or machine stitches them together, throws on a backing track is definitely something they can do/train on using their existing library. Since that kind of tooling isn't for export to customer facing they can train their model on whatever they want (even content outside of their own library) so long as they don't sell it.