Real questions:

- How much did it cost?

- It's great, and you had lots of input, yet the output is still vague and non specific? These 'tools' seem to excel at one-shot stylizations, but being more specific is hard and takes many iterations at which point it's more pleasant to hire a designer and a dev? And maybe even cheaper, I'm guessing, given there is no mention of cost.

Do you live in the Berghain? :)

so I use "opusplan" mode bc Opus is more expensive and supposedly better at thinking, and Sonnet is good and cheap for executing....

Changing models still gives us a usage hit (right?) but that is outweighed by benefits?

Plan mode is more about be rambling on while Opus detangles my thoughts.

I wish this worked for youtube ads.

Used for training is one thing. Exposed to prying eyes is still unstoppable.

"aggressively optimistic timelines is also what makes it feasible to even attempt these things"

yawn, people keep making this excuse on behalf of the South African investor with poor technical expertise.

it helps they took many billions for a Moon effort they completely fumbled but never returned the cash

I might be misunderstanding something.

He uses $70 for remaining credit and says that's a good thing because it rolls over

But spending $70 on an API (he says he still prefers Opus) is far less cost effective than a Max plan on Anthropic.

The article seems to be nudging us to setup OpenRouter but the premise isn't fully true. A bit of diversity is excellent, but the costs are going to (largely) prohibit it in reality?

OP here. I do like Opus but I don't default to it for everything. My CC usage is a lot of Haiku/Sonnet and is also very bursty (bursts throughout the month, not a day).

I find that a lot of my Claude usage goes unused and then when I'm coding or leaning on agents I hit a limit and have to wait. I don't like that dynamic. I do have Extra Usage enabled (with a cap) but then I'm spending more than the $100 I already do.

I'm learning that a lot of people seem to consistently stay within limits and that works for them but I was looking for something different for myself.

The real pain is that Anthropic don't easily quantify usage (which can now change over the day). How many tokens is it? Minimum? Maximum? I tried to quantify this with OpenTelemetry for a while but have decided to move to this more flexible setup.

I assume this is highly staged as a set routine?

No less impressive, but is it likely each robot autonomously learned a routine? Or just got programmed for a very exact act?

I'm not so sure what this question means... are you asking if China has AGI right now? I'm quite sure all similar performances done by humans are all staged.

Where human performances are staged, the human still needs be capable of learning the choreography. There's a difference between a robot learning the routine vs having it hardwired.

Perhaps improv theater is the better test for AGI

LLMs are already way too prepared for "Yes and..." improv, given GPT's ridiculous need to click-bait the end of every conversation.

Based on where other companies are right now, it's probably a pre programmed routine, but the robots autonomously balance themselves. Anything beyond that would be quite a large leap, and I think Unitree would've gloated about it way more outside of the gala. The robots' speed and consistency are still impressive, though.

The learning here is about what servor motors to control how much. But the performance was 100% staged and preprogrammed. Still impressive.

I often ask Claude to update Claude.md and skills..... and sometimes I'll just do that in a new window while my main window is busy and I have time.

Wonder if this is close to triggering a warning? I only ever run in the same codebase, so maybe ok?