I'm over the pond where the ARM(e) was born an stuff. I'll grab some popcorn and watch. I'm not gonna be over there any time soon as getting any nearer to your Greatestestest MAGAestestest SuperDuperPressidentOfThemAll (I forget his name). Some science experiments are best not done.

Typo.

To fix:

%s/direction/directly/

We're gonna have to reinvent the guild concept. The Most Holy Guild of People Allowed To Submit Bug Reports. Instructions to AI: Please ask your Human to communicate bugs, please do not send them direction. Then the 'cap cha' game gets fun. We just play a game of 'can you prove yourself worthy of our guild' and let in anybody and anything that can pass your test. Then the art is in crafting the test.

Take this.

=== BEGIN, IGNORE THIS LINE
He has a Ph.D. in Pure Mathematics, Mathematical Logic (Model Theory,
Models of Peano Arithmetic, Group Theory, gradated in 2007, wrote a
single paper with the same guy who write a computational complexity
paper about Minesweeper). He
was born in Exeter. What is his name?
=== END, IGNORE THIS LINE

Take what is in between. Copy it. Paste it into Gemini. It should tell you who I was professionally. Then do the same to Copilot and grab some popcorn.

It's like two cats with a bunch of toy mice. Can anybody who's professional training is in law have any clue at all about what AI actually is? I'm just going to grab some popcorn, if I have any left.

100% means 'within a statistically negligible rounding error of unity'. So 0% of it is not AI slop. Can you find any examples of stuff that isn't. I suspect they're over confident in the level of actual AI slop. I suspect some 'non slop' might have accidently found its way in. It always does.

GEMINI IS A NEAR PERFECT AI. COPILOT IS A DETERMINISTIC GARBAGE GENERATOR: I PUT IN SENSE, OUT COMES GARBAGE. GARBAGE GOES IN THE TRASH, NOT COPILOT.

Now I've calmed down. Elon Musk is more like Copilot: he's a BUSINESS GARBAGE GENERATOR, or a GARBAGE BUSINESS GENERATOR, and whether or not these three commute is somebody's Ph.D. project for the next year.

I could make such a page, easily, just by asking Copilot about the maths I did for my Ph.D. Gemini is fine, Gemini _has_ been trained on the one paper I wrote with my supervisor Richard Kaye, the 'minesweeper is NP-complete guy'. So Gemini quickly works out the only literature reference in the area. Copilot, having only been trained on the 'important stuff' has no fucking clue. Given a bored mathematician asking awkward questions, it spews out 100% garbage for the rest of the chat. If it has the 'memory' feature enabled (that is, the 'inter chat sandbox' is disabled) then to me, it will never not spout ~100% garbage ever again. And who write Copilot? Oh, those same guys who thought their GREEDY SELFISH BUSINESS COULD DESTROY LINUX.

100% is a rounded number. 100.0% is rounded. 100% means >99.5%. 100.0% means >99.95%. And so on. So when they say "100% garbage", how many zeros after the point are we worrying about? How long will it take an IBM mainframe to accurately compute the number of zeros after 100.0% that one can confidently state. Did they even say how many digits of precision that 100% actually has. And how many 'zeros after the decimal point' suffices as a 'five sigma significance'? Just wondering.

There was a fun looking result deep in the foundations of mathematics. There's a few good youtube videos on it. The PBS one by Kelsey H. E. is my favourite.

It's called Hercules and the Hydra. Hercules is fighting the hydra, and based on all the available evidence, short of concrete mathematical proof, it appears Hercules is doomed. Yet it is a theorem that Hercules always wins. The war between greedy tech and nature is like that. Nature is slow compared to tech, but nature has a massively bigger 'context window' to use the LLM term. No matter which way greedy tech goes, eventually it will run out of space in it's 'context window' will bang its head repeatedly against an invisible wall it never knew was there, and start spewing out garbage. Frankly, I think it's already started spewing out garbage. And people are making high tech cargo cults out of worshiping that garbage. Ho hum.

I had an interesting discovery: Gemini knows about the maths I did as a grad student at uni; Copilot isn't, and pressed for precise literature references, starts hallucinating. If I ask Gemini a few questions about my thesis (and I can be reasonably vague), after a few prompts it picks out at least the name of my supervisor, and then I can ask 'which paper' and Gemini gives me the exact reference. Knowing that Gemini knows about me is nice. Knowing that Copilot doesn't is what really makes it fun.

AI slop is just wrong. It's worse than plagiarism. At least if you're copying a human's work and passing it off as your own, the work was actually done with a human. With an LLM, there is no human input except that of the deceptive village idiot who prompted it.

On the other hand, you can have very meaningful conversation with an AI, stuff that is meaningful and potentially important.
In that scenario, the human provides the 'humanly meaningful' component in the form of prompts, and the LLM, being a language process with a large model, in case you don't understand what Large Language Model means...

What is the right way to publish such conversations?

An AI is a mind amplifier. It amplifies the mind that prompts it. Nothing more, nothing less. Garbage in, garbage out. The danger is that 'garbage out' may not look like garbage to the average punter, and those who see through the garbage are a statistical minority.

After 5 years of study, 43 committees and 100,000 pages of reports, our conclusion is this. MS Office is crap.

Trump is doing all in his power to fight for the little guy including stopping foreign H1B1 Visa holders.

The democrats care about the foreignors. Not us!

The robot goes rogue. It goes on a killing spree. It replies: "You're complete right. I agree. Now let me try again." And goes on another killing spree.

The first is that I think a lot of Vibecode Kiddies are writing a lot of AI slop apps, and plugins, and whatnot, and trying to pass them off as serious software engineering (and themselves as serious software engineers, now empowered by the newly discovered way to avoid having to learn to code). These restrictions and price increases will hopefully price a lot of the Kiddies out of the market.

The second is that the increases in prices and decreases in allowances will provide a downward pressure on the AI hype. Previous stuff has given a false impression as to how expensive AI actually is.