I wouldn’t say it’s totally legacy. (v)ram bandwidth does matter and while the Apple M-series chips do well with their unified memory architecture don’t forget it’s fixed because it’s part of the CPU chip.

I don’t doubt that locally hostable models will be important for agentic tasks but I suspect they are still going to be bigger than most can comfortably host for time being. I still think there will be a place for the super large models for more complex reasoning although how much will be due to the intrinsic knowledge in the weights and how much due to the plumbing around them remains too be seen.

And if course LLM’s are not going to be the end point of the search for AGI. Whatever their architecture they will still need copious amounts of compute.

Which is that? I’ve been experimenting with what it would take to auto detect ad reads in podcasts but I hadn’t come across any existing projects in my research.

I don’t think the UK wants the war in Ukraine to continue without end but the proscription by the writer in this piece seems to be Ukraine can’t win so we should stop giving them weapons to defend themselves and force Ukraine to capitulate.

There does seem to be a desire to get to a ceasefire but so far Russia hasn’t backed down from demanding territory they haven’t taken as a precondition to that. Does anyone seriously trust Russia wouldn’t just use the ceasefire to rebuild before taking the next chunk out of Ukraine? It takes two parties serious about peace to stop the fighting.

I wonder if there are any metrics for how Twitter pays it’s creators vs places like tiktok or YouTube? I guess their financials and burn rate are hidden in the spacex books now but I can’t imagine they ever got to the point of making money. If only premium members matter for monetisation that must be a small pool of mostly bots.

Plastics, waxes, solvents, synthetic fibres and even pharmaceuticals. There are second order products from cracking like hydrogen which are important for things like fertilizer but that is rather carbon intensive.

You can use organic oils for some of these. However organic oils are more complex so need to be broken down into simpler building blocks which all takes energy and processing. Of course you also need to grow oil bearing plants which also tend to require a lot of fertilizer.

My first year Chemistry dissertation was about this. Of all the useful things you can do with hydrocarbons burning them is the most wasteful.

On the other hand it’s legitimately one of the wonders of the digital world. A communally edited source of knowledge free for all to access. While it’s good to scrutinise foundations and how they are run I’ve not seen anyone propose a better solution so far. The data is freely licensed so forking is possible but it’s not really the technical details that make Wikipedia the resource it is today.

If you can get your reproducer in a one liner, e.g:

cd builds/bisect && make clean && make -j(nproc) && make run-failing-test

then it’s super easy to pass that to /bin/sh -c “…as above…” to feed to the git bisect command

What do you need/use that you find in modern editors that you can’t get in something like Emacs? While there are AI integrations for Emacs they are all optional.

Mine are all GPLv3 or forks of other repos and they are listed. I’m sanguine because the license allows for study and if it’s good enough for humans I don’t see why it’s not for clankers.

I’ve long been off the critical path because as a tech lead I have a lot more random stuff (and meetings) to deal with. I’ve been able to vibe code some non-production stuff like scripts to unify feature lists across JIRA, specs and the upstream docs which has helped free up time to hand craft more code on production.

I don’t care too much about the quality or maintainability of those scripts as long as they make my life a bit easier. I do care about the maintainability of the production code base.

The last big update to my git workflow was when I discovered --update-refs as it makes maintaining a stack of feature branches much easier. So far I really only have one “main” dev branch and then peel off the sub-branches at they need merging upstream.

However I shall have to investigate how the history commands are exposed in magit. I can see it being useful if you have long held branches that take a while to upstream but are useful to have in your trees.

In QEMU all of our CI environments are replicable locally as docker/podman images. However usually flaky CI is due to races exposed on overloaded runners so I often also run make -j(nproc) at the same time to simulate that. A retry script is also useful to get an idea of how stable a test is. Having sanitizer builds can also help.

I thought Episode IV was in the initial crawl text.

ETA I’ve just checked and apparently it was added in the special editions, the original 1977 crawl just started with the story.

It’s 3.

Perfect framing is often a giveaway. Most people’s phone footage wobbles and moves around. Background object permanence is another thing to look out for

The Corridor Crew have done a couple of videos where they explain a lot of the things to look for and the reason AI finds it hard. Some examples:

https://www.youtube.com/watch?v=M4TXO4kQwSQ

https://www.youtube.com/watch?v=6hI9T4jnrSI

Is this a sign of Lemmy’s popularity that we are now getting thirst spam?

Because running servers costs money. The project I work on gets donations towards it’s CI costs and it’s not insignificant.

I personally have email integrated into my editor (mu4e) so I can apply patches and search code directly from the email thread. It handles threads and searching really well.