Sounds like you're not defending Access so much as ODBC.

It's mighty convenient for you that your claim makes it impossible for you to provide evidence of this outrageous claim.

... is not let AI get into the hands of people like Bill Gates. That, however, is impossible.

Part of self-driving cars being safer than people is that they don't drive "very close together" - they follow at the appropriate following distance and speed for the lines of sight. This alone accounts for a large portion of the improved safety performance of these systems.

Autonomous cars are actually very good at the driving task. Where they fail is when they encounter situations that are "abnormal" and I think the industry has a problem here; state of the art is either to train specifically for every situation (impossible in the limit) or to kick the decision to a human in a support center somewhere. Things like emergency vehicles, school buses, farmer markets, construction zones, etc.

The industry needs to figure out how to get their systems to "get confused, then reason their way out of a situation."

This long tail is where humans still do better - when there is some "new" scenario, like a particular combination of debris in the roadway, on a Tuesday, with gawkers, during a fireworks display. You will never be able to "train" your way out of that - you need the correct structural formulation for the problem space.

The models are not good at math. They use external planning tool calls and separate calculation engines that are good at math but are not AI.

Sort of like mathematicians then. Because most mathematicians are pretty damn dependent on external tools. My old housemate was a PhD pure mathematics researcher. He couldn't multiply to save his arse.

Though he was a borderline alcoholic and seemingly perpetually drunk, so theres that.

The big takeaway I have from their study is that their study areas - major metropolitan centers - is that cities have way, way worse than average incidents per vehicle mile travelled.

I mean the national average is only something like 1.25 fatalities per 100M miles, injuries only 75 per 100M miles, and "all" incidents like 140 per 100M miles. But the benchmark numbers on the linked websites range from 200 to almost 700 injury incidents per 100M miles in those cities.

Like, if you are using the average number, you can't even reduce the number of incidents by 841 over 271 million miles, because the expected value is only 203. You have to have a rate of at least 300 injury-incidents per 100 million miles to reduce by 841 at over 271 million miles. So really, their results are driven by San Francisco, Austin, and Atlanta.

Their confidence bands are also pretty low; for example they quote 2 serious or greater injuries in San Fran, which on 82 million miles there is a rate of "0.02 per million miles". But 2 injuries is not really enough to make a statistical inference.

Least of all because all the miles are with different versions of their software - you can't make any predictive claim.

Statistics never lie, but you can make them say whatever you want...

Now we have AI scanning the Linux source code and showing that it's vaunted security has been through obscurity for far too long,

Either you don't understand what security through obscurity is, or your being incredibly disengenous. Linux's source has been on public display since the outset. Windows has never been on display (except for the occasional leak). Automatically and by definition Linuxes security has been less "by obscurity" than it would ever be possible for windows to have been, short of a full sourcecode release by microsoft.

Security by Obscurity does not refer to "nobody has found the bug yet", it refers to an intentional policy of hiding design features to try and slow down an attacker. Which rarely is a useful tool. And sometimes actively harmful, particularly if its stopping security researchers finding holes to patch.

Because at the moment , the current large amount of holes being found is literally proof that Linux's model of NOT obscuring security works as we get to find those bugs and patch them. In other words whats happening now is proof that Linux's model works.

The AI coding assistants can then be used where appropriate and with greater review by the more talented. Reducing the "slop."

So heres the thing. I've had a bunch of my work as the senior dev move from "Work on coding challenges with the younger devs and help them code high quality reliable software" to "code review this AI slop and tell me why it doesnt work" , and its fucking miserable work.

Theres a lot more of it, and its batshit code for the most part. Just illogical choices and no craftmanship (Though I'll give claude one thing;- It comments its code, something I've been yelling at junior coders to do for the better part of my career. Lets the greybeard know what the hell they are trying to achieve so I can look at their code and fix it so it actually achieves it.

Its really weird. If you look at where the school, and neighboring barracks are. The line between the school and the barracks directly lines up to where the carrier is parked in the strait of hormuz. Ok. So that lines up with the possibility that the missile overshot the target. That happens from time to time, even the best weapons systems have targetting whoopsies.

The problem is the school was hit Twice with a "double tap" kill-the-survivors hit. That meant it overshot twice? Less likely.

The bit that weirds me out, is that any reasonable intelligence analysis , on account of being a reasonable human being , would go. "OK we see the target. Problem is its next to a school full of children" (The school is marked on google maps, it wasnt a secret). So the logical solution is you hit it at night when all them kids are at home. No dead children to mess up the narrative or give the missile loaders PTSD.

So yeah, definite incompetence SOMEWHERE. It is plausible indeed that the AI chose the target and it wasn't bright enough to go "Ok, theres kids in danger with this strike, escalate this to a human with some suggested actions to minimize civilian lifes. Even worse is the possibility that it did, and Captain fuck-face said "I dont care, fire anyway".

Good at math: fail.

Ok, so he's one of the tiny earlier models.

This is why I still like the (very) old fashion tape backupp systems. Just backup to tape every day (And preferably a copy kept in a bank safe or whatever. In case your office literally burns down or whatever) and if you get ransomware, just send "Fuck you" to the attacker, go to the safe, restore, be done with it.

Granted the cloud... not so easy. But I always prefer old school on-premise servers so whatever.

As careless as it sounds, dumping it into the ocean might well be the most enviornmentally sound plan. Dilution would be sufficient to just make it a non issue.

This caught my eye from the story;-

Former Meta AI chief Yann LeCun is also skeptical, posting on X.com that Anthropic's Dario Amodei "was already claiming that GPT2 was too dangerous to open source back in 2019. I made fun of them then. Everyone should make fun of them now."

Hang the fuck on Yann. The ACTUAL warning was that GPT2 would lead to the internet being flooded with bad advertising copy, fake news stories, impersonations and spam.

That actually happened Yann!

The problem isn't so much about true public places, but the quasi-private places. If I as a journalist (I'm not, but I did a minor in journalism a very very long time ago) film you in the town square, I'm free to run that footage on a news show. Your in a public place, you dont have a reasonable expectation of privacy. Everyone understands that.

And if your sitting in your bedroom you very much do. I cant filmn you through your window. Everyone also understands that.

But what about when your sitting in your car at the lights chatting to your kids on the way to dropping them off at school. Your sort of in a public place, and your sort of in a private place. But now you have flock cameras peering in your window facial recognizing you and your kids, logging your activity into a database, to be fed into a palatir like system designed specifically to strip you of privacy and be used by potentially hostile adversaries looking to hang you for unaproved activities. They MAY have a legal right (Although thats VERY much something lawyers are challenging) but its certainly not what you expect.

And the way privacy laws work, expectations are very important.

One of the neat things about Deep Seek is they actually do something OAI and Anthropic do not;- They actually publish their methods, training code and all that.

Theres actually a lot more going on with these models than "Trained it on ChatGPT" (And its not like the chinese are alone in that. When claude first came out, it kept randomly claiming to be ChatGPT, a smoking-gun level sign of distillation, and OIA has *heavily* implied that they've caught XAI out multiple times trying to distill CGPT. And to be clear , amongst ML researchers, distillation is considered an entirely legitimate and normal part of AI training.). What is clever about deep seek is they've made pretty significant progress on getting a very big model and running it cheap and fast on more commodity software. And importantly , they've published how they do it in a reproducable way. They've also made *significant* progress on training methods that drastically lower the cost of training up a model the hard way.

So yeah, its more than just distilling western models. The chinese are actually making real progress. And..... Deepseek and Kimi and the like are actually really good models. You can do 90% of what the big american frontier models are doing with them, and on a budget, 90% is good enough.