ch_sm

I agree, this is important context. They call it out explicitly. However, even identifying clusters that specialize in certain fields automatically and then finding and fixing learning “errors” is incredibly impressive to me. Especially for medium sized local models between 20B an 70B parameters, this could yield huge capability and efficiency improvements, not to speak of the potentially enormous downstream effects that come with interpretability.

Hm, that’s doubtful. I think even one instance of DS 4.1 Flash alone at ~500B parameters needs a full node (8 GPUs).

Very Steampunk.

Historically, Apple used real photos and video for marketing material, only lit and polished to perfection to create a sort of hyper-real effect. Nowadays, who knows.

Ah, pull-down menus. I love them, use them all the time. Some modern web products actually implement them. Linear for example, and some UI libraries I’ve seen. I always add that interaction at the companies I work for, unprompted haha.

That Anthropic bet is dubious, when such credible and often far cheaper alternatives exist. There‘s barely any lock in or stickiness here.

> a function with 100 lines

there‘s your problem right there

How does it prove the performance improvement? Does it A/B test in place on a production app?

That is pretty adventurous, but it sounds really cool! Will it be open source?

A pervasive feeling at this moment in time for me.

Not sure about that. This was structurally the same before AI: the author of the code wrote the tests. The tests still helped.

Very interesting! I think I can‘t try it since I‘m on an M2 Max, but I would love to see some real-world numbers from the community here!

Sure, it’s not easy, but GALILEO shows projects like this are possible and can be successful in the E.U.

I‘m a long-time Safari user, and I still prefer it, but I nonetheless concur with everything you said in this post.

Slightly OT, but it‘s interesting and nice to see that Anthropic‘s own blog posts don‘t have many AI prose tells (or Claude-isms) in them.

What you‘re saying, critically, is only half the story. The fight was between Russia backed separatists and the Ukraine gov. The shelling went both ways. This has been endlessly discussed and investigated by the OHCHR.

Great! We need any and all open source alternatives to products that people obviously find useful but do not respect their privacy. Looks good! I‘d reconsider the name.

i get the point about the traditional UI being possibly obsolete for a large group of users. But I can‘t help wondering:

> don't know much Linux, but I've had no problems putting together a capable, well-configured, and secure Linux box by using Claude Code

How can you be sure of it being well configured and secure? Especially if you don‘t know much about Linux?

So direct manipulation is dead?

Thanks for that!

I seemed to recall that Tog made that point back then. Note that both Tog‘s text and Keystroke-Level are from the very early days of GUIs, so, grain of salt.

Omanson, OTOH actually contradicts them a bit: expert users are generally faster with well-learned shortcuts – a widely-held view point and the whole reason for shortcuts.

I mean maybe I‘m an outlier but I always use the physical buttons or knobs to control volume on my devices. I think that is by far the fastest, most exact and overall most user friendly and unobtrusive option, even when you have an agent that‘s always listening (which is a future I don‘t want). The point about being across the room is fair.

About that study: I‘m interested to read it, as it seems to contradict established literature on UX (not to mention lived experience) - do you have a link for that?

> "set volume to 20%"

haha oh god no please don’t.

EDIT:

Sorry for being dismissive. What I mean is, volume is something people generally don’t think about in percentages, but by feel/ear. It’s also adjusted so often and impulsively that doing it via chat would be extremely slow and unergonomic. So yeah, I don’t think that should happen.

there‘s another option. It could be both. The labs are never able to recoup their massive investments, eventually causing a huge loss on the stock market, AND the efficiency gains from tech they pioneered but failed to capitalize on wipes out millions of jobs.

In Claude Code (at least inside VSCode) thinking traces are hidden by default. You can reenable them via JSON. They’re not the real thinking traces though, they are something like thinking summaries.

Yes! You’re describing a fundamental dichotomy in approaches to GUIs in general! Historically, we can see the difference and some oscillation very clearly:

The Xerox Star, the Canon Cat and Apple Lisa’s OSes were largely document oriented, and placed no emphasis on the application at all. The Operating System owned the creation and editing aspects of the document. Developers could only add "tools" that slotted into the existing interfaces the OS provided. You’d never launch an App on any of those OSes, you’d "tear off" a new doc from the stationary pad on a Lisa (that sounds ridiculous as I type it).

That paradigm really goes against the idea of "Software as a separate product", which is an older idea, starting in the 1960s at IBM (they unbundled HW and SW, partly under anti-trust pressure). The document-centric GUIs/Shells don’t really allow for a separate software market.

The 1984 Mac was kind of a middle ground, in my mind. It was a platform for discrete Applications (sold separately) but it also had the Finder which let you organise, open and delete documents separately from the Application that edited them.

Apple and later MS tried to bring the document-first concept back with OpenDoc and OLE respectively. OpenDoc never went anywhere then Jobs killed it, and OLE turned out much smaller in ambition and scope (and I think was never very popular with devs).

I think ultimately the intuition and perceived benefit behind document centric GUIs is pretty clear: the OS knows what kinds of artefacts it can handle, and presents a consistent and controlled interface to the user. File interoperability problems are minimised. Devs are allowed to extend functionality to an extent, but they can never take over the system or invent a new kind of interaction.

The App-first paradigm OTOH places more power in the Application developers hands, gives them the means to innovate, and allows for a larger, more diverse, more self-important software market to exist (see iOS).

Mixing the two, as the Mac does, comes with some confusion and complication for the end user, but I think is ultimately the right balance for everyone.

to be fair, nobody in 1990 decided between NeXT and Linux, because (A) Linux didn’t exist yet and (B) they were / would have been completely different markets and totally different capabilities.

I can imagine that the new vapour chamber they’re introducing helps or is necessary to run this chip at high speeds.

I don’t know if i want to expose my child to auto-censorship (or online gaming anyway).

unpopular opinion apparently, but i don‘t understand the finder hate. as a real-life, non-software-developer user, i couldn‘t care less about library folders, root folders, or a unix-tradition file system at all. i want my folders with my documents, pictures, etc. to me that‘s far more user-friendly and comprehensible than what explorer exposes to normal people.

I’m not an expert, but my understanding is that MTPs are smaller LLMs fine-tuned to "mimic" / predict a specific model’s response. It’s possible that the MTP you’re using isn’t trained well enough on Qwen 3.8. What accept rate are you getting?