hedgehog

Maybe enabling MTP causes some weights to be displaced to host memory? MTP itself doesn't do anything during prefill so that should be exactly unchanged, decode will vary depending on settings but with 2-4 proposals depending on workload I've never seen an overall slowdown.

edit: I recommend building recent llama.cpp from source, I've been updating about once a week, as there has been a fair amount of work related to MTP recently. If you're running a lot of tool calling on Qwen you might also benefit from one of the bugfixed chat templates like the Froggeric version.

MTP has no effect on prompt processing.

Yes that's what I've read. As far as I know the approach should transfer well to hybrid model architectures like modern Qwen and sizes like 27B by using multiple chips. LoRA-steered Qwen 27B at 10K+ tokens per second would be transformative for some workflows.

No the weights are in the metal layers, they cannot be updated.

The library does respect the client-side reduce motion setting (by disabling all animated effects).

For our purposes it has the mining/crafting/construction bits the kids care about so in that dimension there's no difference. From my perspective it's been simple to operate, and having all the source makes a good base for building tools and mods (some new pet/creature entities, simplified launcher, screenshot tools).

Luanti with Mineclonia plug-in works well if you are willing to move entirely to Mac/Windows/Linux clients. The whole stack is open source so it's easier to make changes etc if you want. No hacky binary patching of Java bytecode. Originally I was going to try Minecraft proper first since I'd already paid for that, but Microsoft doesn't honor old licenses and in the end the Luanti path probably works better.

https://content.luanti.org/packages/ryvnf/mineclonia/

The VAIO P and some Panasonic rugged handhelds have those chips, both interesting machines, I think it would be possible to write a modern Mesa driver for those but it would be a lot of work.

I auto compact around 200k tokens both due to this and because the cached read cost really escalates when sessions have more tokens than that (too short and you pay a lot in per-compact re-reading of state)

I don't know about the overall breakdown but in my case longer runs are prototyping, bug hunting, reverse engineering, etc. For example Gnome Remote Desktop didn't work in my configuration due to a combination of hardware and codec bugs and settings. One drive to make it work, another to backport the current upstream packages to Debian stable, restack the patches, and push to my machines. Another sequence of long runs was writing a new client for a closed-source conferencing service I use to allow fixing some particularly irritating bugs. Exploit development has the same shape although that's not something I do personally. From what I've seen the amount of useful hands-off run time is directly related to how clearly it's possible to specify a concrete, verifiable standard by which to judge the outcome. For some tasks that might be days or weeks, for forward engineering on a software product that for me is usually under an hour.

I don't think the current subscription price is intended to be a money maker. It's the loss leader to get people invested in the companies' tooling, and make those people more willing to start paying higher enterprise rates as they grow.

It's what some more experienced users want and the companies are following the well-trod path of optimizing heavily for power users at the expense of complexity, only now it's gotten easier to add absurd amount of code to a project. Not necessarily to make it work right. Personally I have some tasks where sessions between one and five days are typical so I appreciate that it's possible.

Once native apps became important Microsoft didn't have a chance, the basis for Apple's success was in place back in 2003-2005. Apple had a big head start on what is now the standard phone template (always-on cell data, capacitive touch, etc), and based on the experience and success of the iTunes / iPod platform they found leverage to negotiate a unique deal where they controlled OS updates and the app store (instead of the carriers). They also had a lot of experience with OpenGL composited UI from OS X which translated to a (relatively) good GLES implementation on iPhone even though they didn't originally have any kind of public SDK. Very useful for games. Now I think there are some groups that are hoping LLM-based agents will upset the dominance of the app store model... I'm skeptical but we'll see.

For the sake of accuracy Claude Code has a "fork" feature which does what you'd expect and based on traces I skimmed this morning it seems under some circumstances you get fork behavior for subagents as well.

If you are not issued body armor and K&R insurance FDE seems like the wrong term. (the use of "engineer" for non-PEs is... a fair debate)

I don't think that's quite accurate, the screens of the time were more like 480px wide and I don't think most people had a strong aesthetic opinion. Just look at Android. The major problem is that whatever app you cared about, whether it was your bank, Facebook, sport news, or Uber, or Google Maps, or whatever, it was on iOS and maybe Android. So at least in rich countries the decision tree was: buy iOS if you can afford it, buy Android if you can't. Apple basically got extremely lucky that the native app thing took off instead of their original vision of everything via the browser.

It wouldn't necessarily have been better, a major reason the Windows phone stuff failed is it didn't have market share to justify app development. Android barely made it work as a well-funded #2. Palm WebOS, MeeGo, there were various efforts that were better than Android and even iOS in a lot of ways but app availability seems to have been the biggest factor in the lack of platform diversity.

Edit: And consistent with sibling comment Microsoft was even paying companies to build apps for their platform, and it _still_ wasn't enough.

That is true of Anthropic's implementation but not inherent in sub-agents in general.

I don't have the experience, how high are Haskell compile times such that a switch like this is worth it? I have some experience with generated Python, Rust, and TypeScript, and I have not found that compile time is enough of a concern that it would offset additional safety of a stronger type system. Like some of the siblings I'm now starting to do experiments with Lean and other tools to get even stronger assurance about system behavior.

What gets more?

Any article with headings "Eye invasion" and "Lumbar puncture" is bound to be a good time.

It may have gotten a nibble but empirically I still have a finger :)

I wanted to see some pictures, this paper has good ones:

https://onlinelibrary.wiley.com/doi/full/10.1002/ece3.10332

If you put your finger in front of a garden slug it may try to eat it, it's a very odd sand-paper sensation but I never knew why.

To start the transition you can build your own tooling, in this case maybe start with the whole app stack including browser in an emulator, emulator controlled over a socket (write a harness that exposes all the inner debuggers, framebuffer, snapshotting, etc). Then generate a component inventory and likely failures for each, and generate pixel-perfection + internal state checks for each. Then migrate one component out (this may be quite a large project due to all the glue you'll need to make this possible). Then do the rest one at a time.

The big problem with doing it this way is you end up with something structurally the same as what you started with, but potentially more code if you e.g. end up carrying your own reimplementation of Web Forms.

Looks like now it shows up but it's not available to try on a free plan.

Heads up to anyone else curious, I installed the Devin CLI and SWE-1.7 is not currently available there.

The "magic here somewhere" in the car is in the design that reference aspects of animal anatomy (facial features, stance) and in the millions of dollars of advertising that prime the public with expectations about how they'll feel driving it, or how to see other people in the car. There's a direct connection there to packaging LLMs as chatbots, it gives them a recognizable shape and behavior that a lot of people interpret as consciousness and personality.

They were (maybe still are?) around 20% of sq feet overall and a lot of that was leases in buildings far from SLU with no obvious Amazon branding.

I'll add a few bits. Commercial leases are typically "triple net" so taxes are passed pretty much directly through to tenants and land lords don't need to worry too much about them. A very visible part of the "dead downtown" effect is due to small businesses that have terrible margins, high fixed costs (including rent), and don't survive losing 20% of their customer base. And finally, anyone paying attention saw that Seattle core downtown is a highly concentrated bet on office rental to the exclusion of almost any other use of space or reason to go there.

A few years back I did an art installation in one of the storefronts at the 2+U building and in the process got to study up on some of the issues and talk to a few people, the general theme was that everyone had a vested interest in focusing on possible causes that were external and fixable within a short time. I don't think that's reality.

It would be really interesting to see how the Qwen 3.6 35B model compares to the 27B on your benchmark.