I trust Weather Underground instead- it's private citizens, not government propaganda, and it's far more accurate.

Post-COVID, I don't trust governments or corporations to do science. I saw too much statistical abuse, p-hacking, politics, and outright lying about the scientific method to trust federal funding OR corporate funding of science.

Science is best done by private citizens funding their own experiments with outside jobs, not academic peer-review cancel culture bubbles.

This sort of system is only useful because LLMs are limited. If they can be told to farm certain categories of step to middleware, then when they encounter such a step, they should farm out the request. I've found, with trying engineering problems, that LLMs consume a lot of steps finding out what to collect, with a risk of hallucination. That's exactly the sort of thing that can be farmed out.

According to both Claude and ChatGPT, that sort of process is the focus of a lot of research, right now, although apparently it's not actually been tried with reasoners.

Yeah, that would count.

Specifically, with something like Protege, I can define how things relate, so I could set up an ontology of cameras, with digital and film as subtypes, where lenses are a component of cameras, films are a component of film cameras, and pixel count is a property of digital cameras.

The reasoner could then tell you about how the bits relate, but a SPARQL search could also search for all records in a database pertaining specifically to any of these parameters.

At least some search engines let you do this kind of search. On those, you can specifically say "I want to find all pages referencing a book where this is specifically identified as the title, it isn't referencing something else, and that specifically references an email address".

So, in principle, nothing would stop Claude or ChatGPT saying "I need to find out about relationships involving pizza and cheese", and the reasoner could tell it.

That means if you're designing a new project, you don't use any tokens for stuff that's project-related but not important right now, and you use one step no matter how indirect the relationship.

This would seem to greatly reduce hallucination risks and keeps stuff focussed.

What you're suggesting can bolt directly onto this, so Protege acts as a relationship manager and your add-on improves memory.

This triplet would seem to turn AI from a fun toy into a very powerful system.

Why are they monitoring syscalls?

The correct solution is surely to use the Linux Kernel Security Module mechanism, as you can then monitor system functions, regardless of how they are accessed. All system functions, not just the ones that have provision for tracepoints.

For something like security software, you want the greatest flexibility for the least effort, and Linux allows you to do just that.

Because it's fine-grained, security companies can then pick and choose what to regard or disregard, giving them plenty of scope for varying the level of detail. And because the LSM allows services yo be denied, there's a easy way for the software to stop certain attacks.

But I guess that the obvious and most functional approach would mean that the vendors would have to write a worthwhile product.

Highly paid software developers are capable of formal methods and other advanced techniques, which would require a far far caster LLM to perform. So the LLM woukd still be more expensive.

Since R1 has good reasoning, but no real breadth, and is open source, the logical thing would be to modify R1 to pre-digest inputs and create an optimised input to 4.1. The logic there would be that people generally won't provide prompts ideally suited to how LLMs work, so LLM processing will always be worse than it could be.

R1 should, however, be ample for preprocessing inputs to make them more LLM-friendly.

The bottom line is that AI software is, at present, nowhere near where it needs to be to be useful. The approach used has serious flaws and more power won't help that.

Artificial silicone life certainly exists in America. But I don't think it's what you're referring to.

Microsoft today isn't the Microsoft of the 80s and 90s?

I think this rather indicates there may be some lies involved in that claim. This is very much the tactic of Old Microsoft.

Gemini struggles. I've repeated the runs on Gemini and can now report it says files were truncated two times in three. After reading in the files, it struggles with prompts - sometimes erroring, sometimes repeating a previous response, sometimes giving part of an answer.

When it gives some sort of answer, further probing shows it failed to consider most of the information available. In fairness, it says it is under development. Handling complex data sources clearly needs work.

But precisely because that is its answer, I must conclude it is not ready yet.

I would be absolutely happy if any of these AI companies were to use the files as a way to stress-test, but I'm sure they've better stress tests already.

Still, it would be very unfair of me to argue a need for improvement if I were to insist on not providing either evidence or a means of testing.

I would estimate 190,000 words in total, and reading the files one at a time and asking for a count of concepts, the total seems to be around 3,000.

I'm using the web interface, using the file attachment button then selecting documents. The prompt used explicitly instructs using the largest input window available and to report any file that failed to be handled correctly. It gives the OK for most of the files, then reports several got truncated.

Yeah, if it was something simple like input exhaustion, they'd not be making the claim. That would be the first thing anyone tested. So I'm reasoning that the problem has to be deeper, that the overflow in Gemini's case is in the complexity and trying to handle the relationships.

Testing it further, after just a short list of questions, the answers cease to have anything to do with the prompt. This lends credence to the notion that it's the relating of ideas that kills Gemini, not the file size.