> and google just started letting all their engineers use claude

That's misleading.

1. Having different models available is useful for A/B testing and helping improve Gemini itself.

2. They have an enterprise offering for Antigravity (their agentic coding platform), and they need to test that it works well with non-Gemini models too.

> late last year/early this year

That's an eternity when it comes to coding models.

In my personal experience, we've had almost a step change every ~3 months this year, at least for bigger one-shot tasks. For example looking at Gemini Flash 3.0 vs 3.5 vs 3.8, it went 5% -> 30% -> 75% on DeepSWE, all since the start of the year.

In one sense you're right.

In another one, Opus 4.6 level already solved 90% of my work-day tasks, so while better models have been instrumental into handling a higher % that does not mean that defaulting on cheaper models can't be good.

I run DS 4.1 flash daily, and then cross check with gpt-6-astra and I've nuked 90% of my AI monthly bill while having higher limits and better performance/intelligence than I did just at the beginning of this summer.

tbf, I the happiest I've been working with claude is late last year/early this year (before March)...

That's because Opus 4.6 was the last good assistant model.

Everything after it might be more "intelligent" but is super tuned around end-to-end task (and related benchmarks), not to act as an assistant.

Now it's *you* being the assistant, reviewer, etc.

Is that right? Why the shit would that happen?

Their ambition isn't your work being amplified by their model, they want you running fifty autonomous long-running agents.

Because they’ve essentially exhausted pre training scaling and are looking to post training to expand capabilities, which is really just optimization via reinforcement learning against specific tasks aka bench maxing.

Right, in terms of hardware and price, I've always been tempted by the Meta Quest, but the ecosystem and the fact that it's locked down has always been the major blocker for me. It does suck that this more than twice as much (although some of it is obviously the current surge pricing), but for hardware that I own, it's well worth it.

Curious if there's any info about how much Foveated rendering helps performance roughly. Is it like 50% more frames?

It'll be highly dependent on the game. Unlike foveated streaming the implementation of foveated rendering could mean many things. Actually rendering at a lower resolution in areas you aren't actively looking is probably the most complicated approach because it seems like you'd have to render two times (low res + high res) to get the desired effect.

> Actually rendering at a lower resolution in areas you aren't actively looking is probably the most complicated approach because it seems like you'd have to render two times (low res + high res) to get the desired effect

In the newer implementations, you can split the display into tiles, and render each tile at a different resolution, so it's not that bad. A bunch of the intermediate rendering targets are already rendered at reduced resolution, so you won't see a full 4x speed up in the low-res regions, but you should be able to achieve a ~30% overall speed up with modern foveated rendering.

That's really cool to know, thanks for sharing!

The framebuffer that you are filling out is still at native resolution, but while computing that frame buffer you can do many of the steps at reduced resolution. For example, instead of evaluating the pixel shader code for every pixel, it could evaluate once for a block of four pixels in the reduced resolution region and use the result for all four.

From my experience, it's quite hit and miss for every app.

If you use Google 99% of the time and only check other apps when Google is wrong, then that's a biased experiments.

I disagree.

Similar to you, 95% of the time I do the simple use case that I've memorized. The problem is that even after a decade of using Git, when I get into those 5% where something goes wrong, I have no idea how to actually solve it. None of it makes any sense in my head, and I have to search for a bunch of magical commands to run, or nuke things to get back to a good state.

With JJ, it just makes sense. It's intuitive and clear. I have been able to recover without any help. Just the undo alone is basically a life saver. Rebasing and merge conflict resolution is so much simpler.

Yes I probably won't use it to its full potential, but still, the little I do use makes 10x more sense than Git ever did.

I don't understand why the fact that it's an Airbnb really matters here? Unless a large number of apartments are meant to be empty, why does it matter if the tenant is short term or long term? Is it that short term users leave and enter their apartment more often?

I am of course just speculating (and I had a lot of waiting time to think about this), but there are a lot of people with a lot of luggage constantly going in and out. It makes the elevators be full faster (perhaps there is a weight based sensor to determine if the elevator is full, and the luggage tricks the system into thinking the elevator is not full yet, because a piece of luggage is lighter than a person but sometimes takes up more floor space than a person). Furthermore, people with luggage take longer to get in and out.

It's a tower in Kuala Lumpur btw. Since it is rather hot here and the elevators are not airconditioned, it's extra uncomfortable.

TIL that there are countries where the building code doesn't enforce the minimum number of elevators. In Russia it's one for buildings taller than 5 floors, two for taller than 9, and I believe three for taller than 16 (every 20-something-floor building I've ever been to had at least three). But we don't have many skyscrapers so I'm not sure how it works for those.

Amusing anecdotes and musings about your stay (well, amusing to read, probably not to live).

Since you’ve got a bunch of enforced idle time, maybe see if you can find the architect firm and see if you can get the backstory? Why three elevators in a 60 floor building? Just low level sadism, or what?

Because they are short-term and many are staying just for the day + bringing family (like kids, which is why they use AirBNB instead of a hotel). The check-in/check-out already add an entry-exit if the user is staying 24 hours only.

It’s not super common to stay in one place only 24h when traveling with kids. Other than the luggage, tourists and residents move in an out of the building just the same. People who live there also need to go to work or school every morning, and come back in the afternoon creating choke points. And these people are there every day, not a handful of seasonal tourists.

The problem is probably the same as every other super tall building. It was impressive for the potential buyers but designed poorly and this only became apparent after they lived there a while (crappy elevators, probably often defective, building sway, plumbing issues, etc). So the owners started to move out and it’s possible the building is mostly short term rental now.

I just tried it, and the answer is non-sense.

I asked it something simple, list some good indie puzzle games, and half the answers are games that don't exist. Imo quality > speed.

Isn't another issue that most successful open models are distilled from closed models, but closed models are putting more and better safeguards against distillation?

No? The only different between these robots and a washing machine is that in theory, the robot is generalist and can do many tasks, whereas a washing or laundry machine can only do one task. You can still in theory fully control what task the robot works on. Also, in theory, the robot would be the glue between all the other machines, like filling up the washing machine, then moving it to the dryer, etc. It deciding what to do isn't a "prerequisite" for the robot.