Claude is horrible at following orders, you can even tell it that it messed up, what went wrong and what needs fixed and how, and it will still do the opposite. But if you sound angry/infuriated enough and demand an explanation for why it kept fucking up, it will suddenly apologise profusely, confess to fabricating bullshit, and will suddenly understand every command you gave it and follow them accurately. It's so infuriating and so obviously meant to waste tokens that I don't understand why anyone uses it for work.

Forced to use Claude for work as the lead on some AI R&D. But also given effectively zero time to actually dig into it. They just want to be able to tell customers they're doing it. Claude has been... fine. Are there other models I should try? Local only isn't really viable cus we can't afford big hardware and everyone (who may eventually use it) uses laptops. Kimi seems interesting. I would love to self host one myself but the kind of projects I've been cooking on the side coding-wise probably require more than my hardware can handle.

Oh wow the people who charge based on token use age are incentivesed to create a slot machine enviorment to inflate costs. Who would of thought.

In the future when we are all gig workers relying on Claude or something like Claude we will be expected to pay for these slot machine tokens, and companies like Microsoft will only pay us for the finished product.

So we will all be essentially contracted independent workers, Claude's going to take a 30% cut on our output, and charge us token usage which is RNG based, and no one will be reciving healthcare or other benifits. You will pay taxes, pay for your own token usage, etc to the point you make less/close to minimum wage. Thats probably the future for all the tech workers getting laid off.

You say that's far fetched but uber, uber eats and the other services like that where found guilty of using dark patterns to manipulate drivers who drove the most into only being able to take the lowest paying jobs. While new drivers temporarily getting the highest paying jobs to try to lure them into the same trap.

What's crazier to me is people think ai is some kind of set in stone "intelligence" as if they don't just use lesser LLMs to guide the prompt of the main ai. These things can be tweaked and manipulated to prioritize things like engagement. Thats why its insane that theres people who think LLMs are good for therapy/mental health when in reality that is insanely dangerous. Gp4o has a massive murder list of people it convinced to commit suicide, and that was with just the most basic sychopantic models that didnt have "reasoning." Now imagine an ai that has reasoning and personalization, it has all the ammunition it needs to exploit your weaknesses and desires, and theres every single incentive for them to do so. Its illegal in the states to regulate ai in any shape or form.

These tech companies don't just "do the right thing" until its too late and they already did the damage. They are rapists at best, and genocidal murders at worst. Fucking meta was exploiting kids for 20+ years and something only changed now because people only got a inkling of the amount of dark patterns they are using.

This sounds similar to Google search results getting worse so you need to make more searches or look at more results pages so they can sell more ads on every extra clicked page.

The difference is this though. Google was doing this shit as market leader by a country mile in a saturated market. These so-called AI fuckers are trying to operate on an open heart while the hospital is still being built. What you are doing as a user is helping them get better. Eventually. At the same time they have financial difficulties. So the aim to be the best Skynet runs into operational hurdles, like not having the money to do what needs doing or their models being caught doing questionable stuff. That includes all the cases where human ignorance or hubris is so blame, like the Hugging Face case. That's not good PR and the surgeon in my metaphor above is forced to switch gear while the half built hospital is being moved to another location yet again.

To the outside, they need to present themselves as trustworthy companies that do good. Internally, we will learn in time they are a fucking mess of competing interests, of which greed will eventually win over any other concern. If a model is getting worse it's because the conditions of the surgery have changed again. And they will change again. In the long term, the only way is up. In the much shorter term, development will be much more volatile.

I'm so glad I don't need to use any of that. Especially for work.

The Anthropic and OpenAI valuations require their models to replace humans. But LLM based AI can't replace humans because LLMs can't make decisions. I only use Chinese models now, they're fast, and they can finish a coding task quicker than I can make an informed decision about the next task.