jarbus

Thank you for posting this, I didn't realize just how good 4.1 flash was. it far surpases gpt-luna in the limited testing I did: luna would take so many tokens that implementing a new feature would be nearly as expensive as if i just did it with sol 6.1, but 4.1 flash can do it for basically 1/6th the price, if not less. Obviously depends on the feature, but very very impressed.

The ego behind the frontier labs is growing evermore concerning

This sounds so interesting and fun. I’d love to do something like this but I lack the experience in construction to actually pull it off. Also, I like to host, and I feel that this kind of lifestyle is incompatible with the average person, who probably wouldn’t want to visit for more than an hour.

Maybe just me, but I have a feeling that we are currently stuck at 100-200k effective context window due to a parameter limitation, specifically the hidden dimension of the model, which would be an incredibly expensive dimension to increase vs adding optimizations elsewhere. I don’t think we are done scaling parameters yet.

Praying that fairphone somehow makes a foldable one day. My next phone will probably be a fold 8 because I really want a nice reading experience on the go, but I love the philosophy and love my framework laptop.

Personally I think we will just become more ambitious with our abstractions since the LLMs can generate more complex stuff at the lower level. To take the article's examples, LLMs will write kernels that fuse 3+ layers of a model, so the abstractions can enable more complex forms of kernel fusion.

I really want to love mobile Linux, and have given it multiple shots before. But I just don't see how it will catch up with iOS/Android in the same way it caught up with Windows/Mac:

- OEMs have poured an enormous amount of resources into the camera software, Linux cameras are years behind from what I understand. - Android/iOS have thoroughly optimized the UX in all sorts of subtle ways, the big one being their keyboards. I literally can't type a single correct sentence on any of the mobile linux keyboards that I've tried because there's so much proprietary magic needed under the hood to figure out what the user is trying to say (autocorrect mostly, but I recall something about how apple also virtually adjusted the sizes of the keys to make it easier to type) - GPS, banking, and now keys for your house and car all need to work.

I’m still a novice on the material science behind it, but what is QuantumScape lacking here if anything, and why is it a dealbreaker?

As someone who generally liked the article, this is perhaps the strongest argument against it. Prose like this requires even native speakers to really focus. I wouldn't write this way for anything intended to be read by an unknown person.

I was unaware of alternative IRs for julia (albeit I started learning much more about compilers once I left the julia ecosystem). What other IRs is this supposed to replace? And why not just go with MLIR directly?

Thanks for sharing your experience, I'm glad to hear that mobile linux is practical for some people. What do you do for GPS? And in day-to-day use, doesn't WayDroid drain the battery?

Tap to pay isn't itself a dealbreaker, for me it's more a representation of the status-quo and what is to come, specifically how more and more of the modern world will rely on integrations with proprietary software and hardware on phones. Tap to pay with credit cards isn't too big of a deal, but the wallet as a whole (i.e. boarding passes for airplanes, gift cards, ChargePoint tap-in) is a major feature.

Linux has long been the most practical laptop OS for me, but I can't see it ever being competitive with mobile OSes, and that's coming from someone who wants it to succeed (I've installed postmarketOS on a OP6T). I just don't see how it will overcome the various issues (app support, tap-to-pay, camera quality, etc).