lanza

I always find it really weird when somebody on the anonymous internet talks about local places as if we're all neighbors or something. Googling "Richmond Hill" gave me multiple pages of results that had nothing to do with the one that Attenborough lives at.

> In my view a spinor is even more familiar than a vector

Okay... Pauli and Dirac both received Nobel Prizes for discovering spinors. Nobody needed to discover pointing in some direction.

Same. I rarely use mine and find that they just about always still have battery. And even when they don't I'll just plug them in and go find something else to do for a few minutes and by the time I come back they are usable for hours. The talk about the power button is super strange to me.

> It fixates on one particular basis and it results in a vector space with few applications and it can not explain many of the most important function vector spaces, which are of course the L^p spaces.

Except just about all relevant applications that exist in computer science and physics where fixating on a representation is the standard.

> I own an M4 iPad Pro and can't figure out what to do with even a fraction of the horsepower, given iPadOS's limitations.

Literally everything you do gets the full power of the chips. They finish tasks faster using less power than previous chips. They can then use smaller batteries and thinner devices. A higher ceiling on performance is only one aspect of an upgraded CPU. A lower floor on energy consumed per task is typically much more important for mobile devices.

The difference in usefulness between ChatGPT free and ChatGPT Pro is significant. Turning up compute for each embedded usage of LLM inference will be a valid path forward for years.

That's a JIT. It uses the same compiler infrastructure but swaps out the AoT backend and replaces it with the JIT backend in LLVM. Notably, this blog post is targeting on-device usage which a custom JIT is not allowed. You can only interpret.

Because the usefulness of an AI model is reliably solving a problem, not being able to solve a problem given 10,000 tries.

Claude Code is still only a mildly useful tool because it's horrific beyond a certain breadth of scope. If I asked it to solve the same problem 10,000 times I'm sure I'd get a great answer to significantly more difficult problems, but that doesn't help me as I'm not capable of scaling myself to checking 10,000 answers.

Without reading an entire novel's worth of text, do they explain why they picked these dates? They have a separate timeline post where the 90th percentile of superhuman coder is later than 2050. Did they just go for shock value and pick the scariest timeline?

Only gripe I have with the tool is that once you've gotten a country right a few times it zooms in too far. I still had no clue where Eritrea was after getting it right like four times. Just got lucky.

But now that the map only shows me three possible countries I can trivially remember which one it was. Ask me again tomorrow while only showing me the full map and I might guess it's in South America.

> I really don't understand what their endgame is here.

To not lose. History is full of stories of incumbents not wanting to cannibalize themselves and dying because of it.

> Nobody up to this day has been able to give a formal mathematical definition of intelligence, let alone a proof that it can be reduced to a computable function.

We can't prove the correctness of the plurality of physics. Should we call that a dead end too?

> The argument is that, instead of hiring a junior engineer, a senior engineer can simply produce enough output to match what the junior would have produced and then some.

...and that's just as asinine of a claim as the original one

> Llama 3.1 405B can currently replace junior engineers

lol

When shopping for a new car to take to the race track on the weekends did you stop and point out that the Honda Odyssey's suspension is too soft?

If I see a single ad I'm uninstalling it and switching to Apple maps.

I guess this is the new "Be maximally evil" Google. Visual ads while driving that force you to look at your phone.

Ditto. They'll use it now while they stand to benefit and in 3 years they'll be lambasting OpenAI publicly for not being private enough with data and pretend that they never had anything to do with them.

It's in the earnings. It was $5.2B this quarter against $15.7B repurchased.

If you're measuring a compiler you need to post the flags and version used. Otherwise the entire experiment is in the noise.

> Mostly the `final` keyword serves as a compile-time assertion. The compiler (sometimes linker) is perfectly capable of seeing that a class has no derived classes

That's incorrect. The optimizer has to assume everything escapes the current optimization unit unless explicitly told otherwise. It needs explicit guarantees about the visibility to figure out the extent of the derivations allowed.

I work on clang and don't know go yet still find the go version easier to read.

Sounds like shorthand terminology in your subfield is okay with the usage of "magnetism." But that's local slang, as a (former) high energy theorist I interpret the phrase "a new form of magnetism" as a new gauge theory or something equally spectacular. So I just rolled my eyes at the headline.

I love the ideas of jj but `jj split` is just an awful workflow. Splitting commits is too important of a feature to get right and jj gets it really wrong.

The delay between the physical keystroke and rendered text insertion + search filtering is uncomfortably high. Compared to FZF's instantaneous rendering of the same this is a UX downgrade.

> You don't need any of it. > History is useful enough to exist as a feature, I up-arrow routinely, but it doesn't actually matter when it doesn't exist.

That's absurdly naive to think the simplistic constraints of your own workflow is a general rule.

I love the idea of Atuin but it's just way too slow with large history files. I've synced my history on my own for the past decade and have like 170k lines and the history search the ctrl-r search just crawls.

I don't need most of the history, but there's 0 chance in hell I'm auditing that many lines to decide what I need and what I don't need.

PyTorch and Triton don't use CUDA. An extraordinary volume of big tech compute only cares about the PTX “isa” for NVIDIA.

sshing into a server only happens within one shell. Thus you can open tmux there and that's the only level of granularity you can have.

iterm window->iterm tab->ssh->tmux windows->tmux panes

With `tmux -CC` it lets you hoist that chain upwards and map the entire ssh session into an iterm window.

iterm window (ssh session) -> iterm tabs (tmux windows)

The biggest benefit for me is that you can then open another tmux session within the CC one and use tmux as normal

iterm window (ssh session) -> iterm tabs (tmux windows) -> tmux2 windows -> tmux2 panes

And then the normal feature of tmux is that it persists across ssh sessions. This also means that the entire iTerm2 window and all its tabs and splits are also maintained across sessions. I think my current layout on my main development machine that I remote into is over 3 months old.

Another point of view is that it decouples the state of your tmux tabs/panes from your terminal emulator itself. e.g. if you use this locally instead you could then terminate the entire iterm program and reopen it and restore the same exact window/tab/pane splits you previously had. This is a nice perk for restarting/updating iTerm.