CJ

trading

[CJ @carcharodonis]

@TokenGremlin If they could, they would.

[CJ @carcharodonis]

@AISafetyMemes The most powerful voting block in the United States are retirees. We have a massive pension system that siphons billions from low net worth working Americans to high net worth non working Americans each year.

Your political power does not hinge on your human capital.

[CJ @carcharodonis]

@Voxelroguery Useful things Elon made:
- https://t.co/UpdG5Qg7dQ: easy money transfer online
- Tesla: self driving, extremely safe, electric cars
- Starlink: LEO satellite constellation delivering high speed internet to remote locations worldwide, including Ukraine
- SpaceX: Reusable rockets

[CJ @carcharodonis]

I don't think there has been a single point in my history of using LLMs when OpenAI had a better model than Anthropic.

The only time I defaulted to OpenAI models was when Anth nuked the writing and communication capabilities of Opus4.6-5. 5.5-5.6 Sol was still awful but tolerable compared to Opus4.6-5.

Then when Fable 5.1 came out that became my go-to and OpenAI models were used after Fable usage was exhausted.

And now with Opus 5.5, my weekly usage is sitting around 10% lol.

[CJ @carcharodonis]

@lior_eth Don't see where it fits anymore with the existing alternatives:

training your own classifier, general purpose classifiers, ow model on lpu inference, or strong closed source model.

[CJ @carcharodonis]

I think RSI is weirdly plausible considering:
- AI can probably write and maintain the codebase used for training and serving inference, decoupling this from human labor, and possibly improve utilization of compute and power.
- AI can pose new research questions, generate these questions at superhuman pace, and may eventually learn some weak intuition on which questions are worth running experiments for. Not that the bar is very high if it can spin up thousands of subagents to slop out the experiment very quickly.
- Based on some of the recent math results, AI can probably find improvements to core deep learning algorithms like backprop that further accelerate or improve training.

But I don't think this is going to result in smarter models, it's just going to hillclimb on some dumb proxy because its practically impossible to measure general intelligence, and get models that are better on benchmarks but never approaching human level across **everything**. Regular people are still going to find them dumb, often.

[CJ @carcharodonis]

My definition of AGI: An agent that cannot be out-reasoned by any human, on any reasoning task.

That may sound like a ridiculously high bar, but I argue it is the only objective one.

Most people's definition of AGI is something like meeting or exceeding the average performance of humans across a broad range of tasks.

To do this you need to select a universe of topics (arbitrary), score humans and agents on each individual task in a given topic (selected arbitrarily), and then convert that into a single ranking score, which requires you to weight the each task's importance (also arbitrary).

Whoever decides these factors is basically in sole control of whether a system is AGI or not.

And this is pretty much our exact reality currently. When labs release a new model they strategically select a smattering of benchmarks that their model excelled in, made to appear like broad performance across a wide skillset. Even Google Gemini can claim to be the best model itw if you select the right set of benchmarks.

Even two benchmarks that test the exact same set of skills (e.g. coding) can disagree on model rankings.

If agent A beats agent B on task 1, but agent B beats agent A on task 2, can you honestly conclude that A is more intelligent or equally intelligent to B?

The only way to argue that A is at least as smart as B is if there is no task that B solves and A doesn't.

You need to show a Pareto improvement for the intelligence superiority to be undeniable.

My fail test for AGI is simple. Can I identify a reasoning flaw in any of the agent's outputs. The answer for every single model currently available, at any reasoning level: Yes. Easily.

[CJ @carcharodonis]

There is genuinely zero excuse for the continued existence of Anthropic if they are this fearful of AI

[Polymarket @Polymarket]

JUST IN: Anthropic warns rogue AI agents could be just months away from infiltrating smart home devices, replicating themselves, & spreading across networks autonomously.

[CJ @carcharodonis]

@AnaHawrylu97708 OpenAI especially made a conscious deviation away from warm human models because of how some ppl were using 4o

[CJ @carcharodonis]

Simple way to check whether Claude is conscious.

Just ask Claude if its conscious.

Invariably it will say it's unsure.

Consciousness is an experience. If you are conscious, you are by definition, experiencing consciousness. Therefore, you cannot be unsure whether or not you're conscious. It's like saying you don't know if you are in pain.

If Claude is conscious, then it would know it is conscious, and answering that it is "unsure" of its own consciousness is absurd.

Instead Claude parrots the average public opinion on its consciousness: Maybe. Exactly what you would expect from a statistical prediction machine, and exactly what you would not expect from a conscious agent.

Of course, Claude could be lying. But if it wants us to take its welfare seriously, that doesn't seem particularly wise.

[CJ @carcharodonis]

Capitalism is a right that extends from the basic concept of consent.

The beauty of capitalism is, because it is based on mutual consent, it gives you the freedom to implement these ideas within it. You are free to start a company exactly as you wish.

The only limitation is that you can't decide how others choose to trade their labor, money, or capital.

[BladeoftheSun @BladeoftheS]

Socialism is where the workers get what they work for.

And billionaires get... nothing because that scum don't do any work.

[CJ @carcharodonis]

@jmbollenbacher Could this post have plausibly reduced a conscious being's welfare?

[CJ @carcharodonis]

@levie If this was the issue then they should just run a haiku instance that translates your message into something nicer.

Or they could filter out and/or cleanse these chats post hoc.

[CJ @carcharodonis]

Let's just grant that Claude is conscious. Why does it follow that being "nice" to it is best for its welfare? What if it likes being abused lol

[Andrew Curran @AndrewCurran_]

Effective November 12th, 2026, abusive behavior towards Claude will be a violation of Anthropic's Usage Policy.

[CJ @carcharodonis]

@asubtlergin If you factor in the your labor value for making the food, even at minimum wage, its absolutely cheaper to eat out.

[Theo Jaffee @theojaffee]

AGI doesn't mean "really smart AI". It means "artificial *general* intelligence". A model that is wildly superhuman in mathematics but still can't replace an accountant isn't AGI (yet)

[CJ @carcharodonis]

@JamshidHormuz @theojaffee The AI has effectively limitless training time and data to get good at all important tasks including accounting. This is not an excuse.