micw

A strategy that could work very well is to let the AI

- explore the features and behaviour of the existing code

- first write lots of E2E tests for literally everything

- human-in-the-loop to review the feature list and test cases to see if some obvious or important stuff is missing

- then let implement it and verify by tests

I used this in (way) smaller scenarios and got really good examples with minimal human interaction.

Try GLM for such things. Better than DS Flash in my observations but less restricted than GPT

Thing is that you need incredible large masses in incredible heights.

Example: Lifting a 10-metric-ton stone by 36 meters stores just about 1 kilowatt-hour (kWh) of energy. This is roughly equivalent to the capacity of a small power station used for a camping trip.

> we'll all be around to see solar panels that cost 1/10th of what they do now

I doubt that. Currently the cells itself makes only <50% (maybe <30% meanwhile) of a PV module. So the limit to the bottom is by the glass and frame and cable, not by the cells.

But $/watts can significantly go down with more efficiency.

It makes no sense to make the devices so thin and then having a fat camera. This requires a cover to equalize the size with the camera for haptic reasons. Better use this space for other stuff like larger batteries and make the device the same thickness everywhere.

I use TensorX (EU hosted) and get 330TPS

It's LLMs, not LLM's ;-)

How do you access Claude from the multiplexer? AFAIK Anthrophic allows the use with their official tools only.

I have created wrappers for Codex (https://github.com/micw/codex-wrapper-advanced) and claude (https://github.com/micw/claude-wrapper-advanced) that uses their SDK (Codex) and the CLI (Claude) internally to align with the subscription ToS and still have a common API ;-) This way I can use both in any harness and can easily switch between both.

Also not available in their business offer. So I have now a $20 business plus a $200 "personal" plan in my account.

Yes, it should. Reduces a lot of noise.

The idea looks interesting. But when I think of "make incident recovery easy", the very last thing I want is a fork that differs from the standard that everyone else is running. I would have a better feeling if that would be an extension, not a fork.

I would even think that "smart" and "lazy" has a strong correlation ;-)

If you give a smart one a task, he will look for a way to do it with lowest effort. Probably most would do it but the smart ones will find way more the shortcuts, automations and exploits.

Edit: also smart people will more often ask "why" if a task makes no sense to them.

Tetris on postgres? Not doom? So not a candidate for everything!

Just kidding. Of course there's doom for postgres: https://github.com/cedardb/DOOMQL (pure SQL) and https://github.com/DreamNik/pg_doom (extension).

Oh and there's https://github.com/snaplet/postgres-wasm that allows to run everything else in postgres ^^

So is this a regression or a bug in systemd triggered by this change?

What does the (currently latest, https://github.com/systemd/systemd/issues/40262#issuecomment...) comment mean? Who are the "large folio people"?

systemd-journald also has rate limits that you can configure ;-)

I also saw this when I made a (personal use only) book translation (agent learns how book 1-3 of a row is translated, then translates book 4 the same way because it is not available in my language). Claude happily translated most of the book, except a few chapters which it denied. Tried several times always the same results (with no other context about the rest of the book).

> if you see an ad somewhere [...] close the tab immediately and never return to that site

I would even extend this: if you are bugged by an ad, avoid the product, the vendor and tell everyone why you do so. Also tell it to the vendor, why you avoid their product (aggressive, non blockable ads at facebook for example).

Why is this not being stopped by us jurisdiction? Or at least why does the rest of the world does not decouple from those markets?

I even upvoted, no idea why you're downvoted here. It _is_ a potential issue that you pointed out.

To be precise (Germany): 800W inverter, 2kWp solar panels (maximum 4 panels). This allows you to get the 800W during long periods of daylight (e.g. east/west configuration). You only need to add the installation to a public database and plug in the inverter. Before it was more complex (special plug needed, only 600W, need allowance from grid operator) and still we have ~1.5 GW installed plug-in solar plants. It's almost always a no-brainer, RoI is typically 3-4 years while the lifespan is 20 years.

Edit: My state had a subsidy of 300€ when I bought mine, so the RoI was instant ;-)

In Germany, where this kind of plug-in solar panels is very common (1.5 GW installed power) it is mandatory that the inverter powers itself down as soon as the grid is offline/disconnected. It has to be so fast that you can savely touch the plug after unplugging it. I guess it's standard everywhere.

No, they would collapse into a singularity with an infinite event horizon that eats up the universe at the speed of light

I wonder how Americans feel to see that their elected president takes tax money to forcefully destroy the planet and our kid's future :-(

Guess that's the exact point of the "intelligence" benchmarks

I recently did a few tests. And always the same prompt has been cached properly.

Paying $100/month and using Opus 4.8 all the time to get my work done. I had a brief look at fable (when it was available the 1st time) but it burned my allowance to fast. So I keep paying the $100 for Opus and enjoy friction-less uninterrupted work.

I'm usually at 60-70% usage at the end of a 5 hour window, that's the pace where I can still think about what to delegate and what to expect and verify the results. Could probably go faster but that would have a significant impact on output quality.

I would have expected at least a few sample real-world photos on that page.

More reason to frame it correctly, like "IBM debuts another 5 nanometer chip technology, selling it as sub nanometer chip technology"