You just got used to slop and peeked behind the curtain when the wow factor wore off.

LLMs are next token predictors. Outputting tokens is what they do, and the natural steady-state for them is an infinite loop of endlessly generated tokens.

You need to train them on a special "stop token" to get them to act more human. (Whether explicitly in post-training or with system prompt hacks.)

This isn't a general solution to the problem and likely there will never be one.

LLMs don't think at all.

Forcing it to be concise doesn't work because it wasn't trained on token strings that short.

> Forcing it to be concise doesn't work because it wasn't trained on token strings that short.

This is a 2023-era comment and is incorrect.

Anything I can read that would settle the debate?

LLMs architectures have not changed at all since 2023.

> but mmuh latest SOTA from CloudCorp (c)!

You don't know how these things work and all you have to go on is marketing copy.

Yea you don't know anything about LLM architectures. They often change with each model release.

You also aren't aware that there's more to it than "LLM architecture". And you're rather confident despite your lack of knowledge.

You're like the old LLMs before ChatGPT was released that were kinda neat, but usually wrong and overconfident about it.

You no listen. Me give up. Go learn on fruit phone.

LLMs are literally next token prediction engines and nothing else.

Diffusion for text is not even an academic toy at this point and will likely never be a real thing.

They’re able to solve complex, unstructured problems independently. They can express themselves in every major human language fluently. Sure, they don’t actually have a brain like we do, but they emulate it pretty well. What’s your definition of thinking?

When OP wrote about LLMs "thinking" he implied that they have an internal conceptual self-reflecting state. Which they don't, they *are* merely next token predicting statistical machines.

This was true in 2023.

And it still is today.

1.5**12 is about 129.74, which is as close as you can reasonably get to a power of two.

So yes, the 12-tone scale is a universal thing - you want both octaves and fifths in your scale.

(12 is actually too much, so usually that's pared down to something like 4 or 5 or 7 tones, this is where you get cultural variation.)

> 1.5*12 is about 129.74,

Math checks out.

> So yes, the 12-tone scale is a universal thing -

I don't follow the logic here though. It's certainly true that a 12-tone / Chromatic scale is ubiquitous within the Western Music tradition .. but the universe is reportedly a little larger.

Even Western Music includes exceptions like the 9-note augmented scale, though the argument can be made that it's a 12-scale with 3 bits "missing" - not a case that can be made about a non-western 7 note percussive scale.

All scales in all cultures are based on octaves and fifths. (E.g., the ancient Chinese musical scale also has 12 tones.)

Also the so-called "Western music" standardized on 12 tones very late in the process, long after the Chinese figured it out.

> a 12-scale with 3 bits "missing"

That's all scales, even the "non-Western" ones. Microtonality is added to the standard 12 tone to add tone effects. (Synthesizers in pop music do the same trick.)

To confirm the claim that "all scales in all cultures are based on octaves and fifths" one might study the scales.zip scale files and find those that do not contain octaves and fifths, which should naturally be zero if the claim is true.

https://www.huygens-fokker.org/scala/

Note also that certain musical traditions were suppressed or eradicated due to their unfortunate habit of using dissonant notes such as minor seconds, as opposed to the consonant traids favored by a particular group recently in power around the world. Happy Easter!

The 12 tone scale is a natural property of the world that has been independently discovered by many different cultures across the world.

The microtonalities or strange scales (like gamelan) are evolutions of the natural octave/fifths grid.

Every synthesizer has a "detune" knob.

Thank you, I am somewhat aware of the knobs present on a synth, though fail to see the relevance given that various other instruments do not have dynamic retuning options. Which 12 tone scale did you have in mind (for there are many) and why do you think 12 (for there are many other numbers, some of which are used by various scale systems) is a natural property of the world? Perhaps with a more cogent argument you could make a better case for your opinion.

> microtonalities or strange scales (like gamelan) are evolutions of the natural octave/fifths grid.

What is your source for this opinion?

  Gamelan predates the Hindu-Buddhist culture that dominated Indonesia in its earliest records and thus represents an indigenous art form of Indonesia

~ https://en.wikipedia.org/wiki/Gamelan

How do you know anything that comes prior to recorded history?

129.74 is not really close to a power of two. 31-tet scales have a better approximation of a 5th (and an impressively better approximation of a minor 7th).

The obvious exception in the western system would be the blues scale, which arguably has 9 tones (7 equal tempered notes, plus a just tempered 3rd and 7th).

And Indian ragas break all of these rules. They have scales that don't have 8 notes, scales that don't use equal temperament, and even a few scales that don't repeat on octaves.

> 129.74 is not really close to a power of two.

About a 1 percent difference. It's close enough.

Equal temperament is a different issue. The 12 natural tones are necessarily approximations and can't be represented exactly due to the 1 percent difference.

> Music today is utter crap at all levels, this is a verifiable scientific fact.

No it's not, and it's not a verifiable fact. Unless you have a source? Rick Beato knocks the sami-ness of 'the charts', but there's more to music than that...take a look at who he interviews.

Bach is considered the greatest musical genius of all time, but he was part of an industry and composing was his day job. Each of those BWV's was written in a couple days. Bach's performers at the time didn't study for years for a single recital, they read the sheet music in an afternoon and then performed the BWV next day.

Beethoven improvised his pieces on the fly and performed them himself. This wasn't considered as something out of the ordinary at the time.

Can you imagine the average conservatory graduate imporovising anything today? Even a pentatonic blues riff?

Clearly we went off the rails bigtime somewhere along the way. The framework we're using to teach and compose music is actively hindering us.

Your ISP or telecom has to be compromised for TOFU to be relevant to anything. In practice that never happens.

Not just your ISP. If an attacker slipped a device onto your LAN and also you happened to be sshing to a new box for the first time then TOFU poses a problem. But that's an awfully limited attack surface. It's similar to the difference between leaking a fax while it's sent versus leaking years old emails that are just sitting there on an internet accessible server.

As for your ISP I think you should never rely on TOFU over the public internet. If you really don't want to do ssh certs it's easy enough to make the host key available securely via https.

You will have to manage your SSH CA certificates instead of your keys.

The workflows SSH CA's are extremely janky and insecure.

With some creative use of `AuthorizedKeysCommand` you can make SSH key rotation painless and secure.

With SSH certificates you have to go back to the "keys to the kingdom" antipattern and just hope for the best.

> With SSH certificates you have to go back to the "keys to the kingdom" antipattern and just hope for the best.

Whut? This is literally the opposite.

With CA certs you can create short-lived certificates, so you can easily grant access to a system for a short time.

And what about the CA?

It's no different compared to regular SSH private keys. You need to protect it from compromise.

However, it provides you an additional layer of protection, because it does not need to be on the critical path for every SSH connection. My CA is a Nitrokey HSM, for example. I issue myself temporary certs that are valid only for 6 hours for ephemeral private keys.

Yes it is different. SSH CA keys are harder to secure and attackers have a much bigger incentive to steal them.

You can also configure multiple CA for client auth, and on the client side multiple ca to verify host keys.

Exactly. We'd had discussions about building https://Userify.com (plug!) around SSH certificates, but elected to go with keys instead, because Userify delivers most of the good things around certificates without the jank and insecurity.

It's not that certificates themselves are insecure themselves, it's that the workflows (as the parent points out) are awful. We might still add some automation around that (and I think I saw some competitor tooling out there if you're committed to that path) but I personally feel like it's an answer to the wrong question.

> can find and report on some very sophisticated security issues sometimes

Fixed it for you.

> if the outcome is reliably and deterministically achieved

It's not. My favorite example: due to vibe coding overload literally nobody knows what configuration options OpenClaw now supports. (Not even other LLM's.)

Their "solution" is to build a chat bot LLM that will attempt to configure OpenClaw for you, and hope for the best, fingers crossed. Yes, really.

The openclaw situation is ridiculous. Configuring it is a nightmare, even with 3 different LLMs trying to help. Then I check their docs and it says three different things. Agents will take questions and turn them into a new config file, which consists of made up settings, causing the gateway to crash.

My setup is very simple too, just two agents, some MD files, and discord. Nothing else. These people using it for real work or managing their email and texts are in for a rough ride.

> microfeatures/microservices

Have you seen the code generated by AI? These things converge on the "1 million lines to make an API call" pattern. They're a lot of things, but certainly not "micro".