Cracker

Charlie Kirk was assassinated for exercising free speech. The left celebrated it. There will be no unity.

[Cracker @AFlaCracker]

@JohnKasich No, we don't need to deliberately and preferentially import people from the worst places on the planet.

That's an unbelievably stupid thing to do.

[Cracker @AFlaCracker]

Current models aren't capable of doing what I would want, and even Fable 5 for the little time I used it wasn't looking too promising.

I guess that's job security for me... but it sure would be great if I could enjoy some of the same benefits that others see with AI.

That's the difference between working on the millionth version of something versus cutting-edge, unpublished, and never been attempted.

[Cracker @AFlaCracker]

@UziCryptoo @uoislame All of humanity has survived on "potatoes and beans", or their regional equivalent, throughout most of human history only up until the last ~50 years.

[Cracker @AFlaCracker]

@Chris0x88 @jun_song @leoqin Yes, context matters.

Context also matters in Chinese, perhaps even more so.

[Cracker @AFlaCracker]

@nomad_stable @jun_song "Mythbuster: Chinese Language Is Not More Efficient Than English in Vibe Coding: A Preliminary Study on Token Cost and Problem-Solving Rate"

https://t.co/DuES5f7PTQ

[Cracker @AFlaCracker]

@Chris0x88 @jun_song @leoqin LLMs don't process characters. They process tokens.

And each *token* can represent just as much "meaning" regardless of the tokenized language.

[Cracker @AFlaCracker]

@jun_song This is nonsense.

LLMs don't process individual characters. They process tokens.

And the efficiency with which "meaning" is packed into tokens is a property of the tokenizer and the language it is optimized for.

[Cracker @AFlaCracker]

@DanielSmidstrup Invent, implement, and then optimize novel algorithms not already found in thousands of separate GitHub projects.

[Cracker @AFlaCracker]

He isn't making a very good point.

Uber is a taxi service. Taxi services have been around for a long time.

The difference is in how you call a cab. It used to be with a phone call, or by flagging one down on the street, and now it's with an app.

We'd still have taxi services and taxi drivers without an Uber app.

[Cracker @AFlaCracker]

@hanakoxbt "Our agents fix themselves 100 times" ... and we get to pay for each of those 100 mistakes and 100 fixes.

... great.

[Cracker @AFlaCracker]

@sairahul1 "No memory file, so every loop starts from zero.
No sub-agent split, so one agent tries to do everything.
No stop condition, so loops run forever and bill you in your sleep."

Yes, this sounds exactly like how Anthropic would want paying customers to use it. 🙄

[Cracker @AFlaCracker]

@ihtesham2005 Taxis existed a generation ago.

[Cracker @AFlaCracker]

Dario himself made the point that capabilities of this magnitude always before been developed and initially controlled primarily under government auspices, and he doesn't think AI should be the first major example to fully buck that historical pattern without strong public oversight.

And he's right.

This goes far beyond your "right" to one-shot vibe-code the millionth Minecraft clone.

[Cracker @AFlaCracker]

@zostaff A compute bottleneck is a model bottleneck. A more efficient model would require less compute.

[Cracker @AFlaCracker]

It's more than "point it at your codebase and it will find vulnerabilities" with API access to the underlying LLM and your choice (or your own implementation) of agentic scaffolding.

Give it real tools and a permissive sandbox (or lack thereof) and "point it at your codebase" turns into "scan this subnet and exploit vulnerable hosts" in real-time.

[Cracker @AFlaCracker]

@alex_verem Well, if you can't make it secure then you don't get to distribute it freely to the world.

Because where "an unjailbreakable LLM" is an impossibility, export controls aren't.

[Cracker @AFlaCracker]

@gailcweiner Again, that's a them problem. What can be accomplished is export controls.

You either secure your cyberweapon or you can't publicly release your cyberweapon.

The "release it anyways" option is not on the table.

[Cracker @AFlaCracker]

The "g0VerNmeNt d0esN’t uNderStAnD AI" crowd is ridiculous.

AI bros: "This is so awesome and dangerous. It's basically a cyberweapon."

USG: "Okay... you can't export a cyberweapon. You're going to have to secure it."

AI bros: "You just don't understand AI! It's impossible to secure it."

USG: "That sounds like a you-problem. You still can't export a cyberweapon."

[Cracker @AFlaCracker]

@elliotarledge The people saying it "can't be done" evidently don't understand AI either.