I’m not really interested in analysis of the weaknesses of such models because in my experience many weaknesses disappear entirely as models get stronger and reasoning effort is turned up. Especially if you tell them what you want them to do.

Also, it’s not surprising to learn that when more acceptance criteria are added the failure rate increases.

Things now seem to be stable enough that you just need the Cabal solver - but I suspect the broad nightly builds and visible failures/blockages probably help keep the ecosystem solvable.

The AI results were bad beyond all human understanding - the sort of product that I thought only a walking corpse like Microsoft could release so broadly. But, that is essentially what Google Search is now. Clearly no one in a position of responsibility knows or cares about product design or performance and it only continues to exist through sheer inertia.

Switching to Kagi is how it felt when I first used broadband internet in college , or when I switched from Alta Vista to Google: the internet works like it should.

You need to understand what a nuclear chain reaction is. https://en.wikipedia.org/wiki/Nuclear_chain_reaction

> We've had orders of magnitude more energetic events in earth's history

It isn't about energy. There was never an unbounded nuclear chain reaction of anywhere near this magnitude on the planet before Trinity. A large asteroid impact doesn't cause a nuclear chain reaction at all. The moon impact melted the entire crust but didn't cause a nuclear chain reaction.

In fact, the only chain reaction that happened at all before Fermi's experiment in 1942 - that we know of - was in Oklo (now Gabon) about 1.8 billion years ago. We didn't learn about that until 1972, and anyway that was more like a controlled reactor pile and it only happened because there was so much more Uranium 235 so early in the Earth's history.

The event at Trinity was completely different because so many neutrons were released at exactly the same instant. They had good reasons to be very confident in their models and calculations, but they were not 100% sure, and as TFA points out, the blast was several times more powerful than most models predicted.

Stockfish's neural net evaluation model was trained on millions of its positions with its own original algorithmic evaluation function (entirely developed by humans) and search tree. The result was a much smaller model than Leela's that requires little computation (not even a GPU), paired with its already extremely efficient search/pruning algorithms that made it stronger than Leela in competitive play. Leela's evaluation function is much stronger (at one ply it has an ELO of around 2300, Stockfish is probably closer to 1800), but it requires vastly more resources and those are always bounded in a match.

Humans haven't learned as much new information about chess from Stockfish as we have from Leela.

That is not what the mathematicians are saying. I don't have the knowledge to evaluate this myself, but a number of mathematicians - for example, in the SP - are saying it goes further than that - they really do introduce novel ideas. Of course everything is based on and inspired by some previous work, but that is true of all human mathematics as well.

LLMs that have been trained through reinforcement learning on mathematics are NOT simply token predictors. Only base models can be accurately described that way. They have learned how to do mathematics. They have learned to do coding. Its really amazing we're three years into instruct models and such a large part of Hacker News still does not understand the most basic facts about this field.

Their point is more that SwiftUI has generally poor performance. Lots of native Windows frameworks have poor performance as well.

Native UI development is a minefield. If you want to build an app today that will still run in 20 years without a complete rewrite in the UI layer you should probably use wxWidgets if you are committed to native - even if only targeting one OS. But that model is really only appropriate for building traditional desktop apps. I don’t think the market would accept a Slack or Notion built that way today.

Agreed. In Chromium all the content from HTML is rendered inside a single object from the point of view of the host UI; much like a game engine’s UI rendering. Chromium draws everything itself. Host events like mouse and keyboard events are sent to that top level object (although there are some shenanigans involved to make it look more native to accessibility tools).

Also, Python does not build or run large language models. It orchestrates C code that does that, and it was probably good enough to do that in 1998.

The reason Oracle can continue failing at those massive projects is simple: everyone fails at them routinely and often it’s the customers fault.