AI, that doesn't suck, running locally on your phone.

You know who's most likely to get to the doesn't suck stage first?

The phone maker who makes money off hardware, and not ads and surveillance, which inherently suck.

Now, all they have to do is get their software under control again... *sigh*...

He drew on chess as precedent: 15 to 20 years ago, a human checking AI's moves could beat a standalone AI or human, but machines have since surpassed that arrangement entirely.

Chess programs got better because the machines running chess programs got Moore's Law better and Denard-scaling faster, so the same costing machine could search deeper.

Machines running LLMs are not getting better at the same double-every-18-months rate that we had until about 2012. Moore's Law for storage and gates is not quite dead yet, but clock speeds are basically standing still.

Government dragnet searches used to be unconstitutional - Fourth Amendment and all that.

Then we got the third party doctrine which says it's OK for the Government to ask third parties to give it info on people without a warrant, and it's OK for those third parties to comply. Some of the biggies like Apple and Google usually won't comply without a warrant, others cooperate.

Today, there are third parties like Flock who basically sell dragnet searches to anyone who can pay - auto dealers who want to repossess cars, etc. Other third parties like Ring default to taking their users' data and making it available to the Government - you can tell them not to when you sign up, but the vast majority don't see the harm in contributing to mass surveillance.

Short of new overarching privacy laws like the GDPR, we're screwed.

I wouldn't carry a device in my pocket that contains a lithium battery that hasn't been tested for safety in a vacuum.

The real risk isn’t that a generation can’t read code. The risk is that we stop expecting them to. If we treat LLMs as training wheels instead of prosthetic eyesight, we get a generation that ships faster and understands deeper. If we treat them as a replacement for learning, we get brittle systems and brittle people. That’s not a technology outcome. That’s a cultural and educational choice.

The ability to read and write code without support was in decline long before LLMs - FizzBuzz as a low bar dates from Spolsky in 2005.

If we're hoping for the right cultural and educational choices to save us ... we're screwed.

Until then, it's pure theoretical speculative jabber.

as a meaningful category when chocolate is added to it.

It's like Paula Poundstone's reasoning about Ding-Dongs:

They only have three ingredients- devil's food cake, creamy filling, chocolaty coating.

I don't know how the next Python is going to get any traction, if table stakes for adoption is "language is understood by LLMs".

The current generation of coders won't use it if their LLM of choice doesn't understand it.

LLMs won't understand it if there's no training data, which comes from users.

I've always told people that coding would be automated last, if ever.
Apparently I was wrong, and will have to settle for being part of the last generation of coders that can actually read and understand code without LLM support.

Anyone who owns a Cybertruck bough it relatively recently, so Musk's predilections were priced into that. A shotgun rack in one of those would definitely be on point.

As for older Teslas, there's a lot of angst and buyers regret out there.

Witness the brisk trade in bumper stickers like:

I Bought This Before Elon Went Crazy
Vintage Tesla - Pre Madness Edition
This Vehicle Reflects Climate Consciousness, And In No Way Aligns With The Political Views And Actions Of Its Company's Founder

I have a 2020 Bolt, so I have no dog in that fight other than as a somewhat bemused observer.

Their standard service was still 8 times slower on upload versus download the last time I checked.

Despite the pandemic showing us that we need two-way video to do virtual anything, so service is limited by the slowest data direction.

I'm considering moving to a senior living place in the next year or so. One sticking point is that their building ISP is Comcast.
I'm a software engineer - my work network meetings are mostly audio and screen sharing, so I can tolerate asymmetric data, barely.
My wife is a sign language interpreter - her remote work is absolutely dependent on two-way video.
We may have to tell them "FIOS or no sale".

Trusted third party that collects and batches multiple orders into lumps big enough to keep the tariff overhead reasonable.

Of course it would have to charge its own overhead for receiving and dispatching orders... never mind.

My wife has Ehlers-Danlos - weak stretchy connective tissue. It's a spectrum syndrome running from very loose joints on the low end to crippling joint/ligament problems with lots of other stuff thrown in. She's on the low end, thank ghod.

Ehlers-Danlos is highly correlated with sleep apnea that has nothing to do with overweight.

If you can ship and deploy 100x bigger or faster, you are exposing 100x as much interface to adversaries.

You'd better hope your LLM was trained on very carefully vetted example code, as opposed to the highest rated posts on StackOverflow, or maybe even all posts there.

Bell Labs (cosmic microwave background radiation, among other things) did its best work while it was a state-authorized monopoly.

Good old fashioned AI used to be hands-on - your dissertation code had to at least work for the examples in your thesis, and your code was under development for long enough that it had to survive OS and language updates.

Being wary of code by theoreticians is definitely valid - I believe it was Knuth who said something like "I have only proven this code correct, not tested it".