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.

Issue triage, code exploration, extracting information from disparate sources, first pass code review. There are loads of use cases that it’s potentially useful.

For me it’s a lot better at extracting the requirements for a CPU feature from a 10,000 page architecture reference manual than I am.

I have API access at work because I don’t want to be tied to a UI. I’m very aware of the cost because I’m trying to see where it offers good value for money.

Of course things like the deep research and notebooklm are covered by the Google workplace fees which while including more than the personal plans are also a fair bit more expensive.

Your making a big assumption extrapolating from one particular study involving Java code and a static analyser.

How is that patch sloppy?

I feel the term slop is being overused to cover anything an LLM has touched. If I ask an agent to re-read a mail thread for me and apply the changes to my tree to review is that slop? Would you feel better about it if I copy and paste from email to code in my editor?

I’ve just been doing a bunch of bug triage which was mostly driven by the agent although I checked the issues where it had commented. Was that slop? Ironically a lot of the issues where AI generated although for the most part more complete than a lot of the purely human submissions we get. Are those bug reports slop? What about the poorly drafted human ones?

That’s not kernel policy but LF guidance. From the kernel’s point of view patches still have a high bar to pass to get merged and I don’t think we have enough data yet to see if LLM based submissions to the kernel have a higher or lower error rate than humans.

I certainly feel the uptick in LLM reports though - one of the projects I’m working on is seeing a deluge of them at the moment.

What’s wrong with backports and at a push snap/flatpak/appimg for apps? One of the main reasons I’m on Debian stable is I don’t want the underlying plumbing of my system changing radically.

Where you live maybe. The NHS is centrally funded through taxation.

If course you do - if the cost of treating the patient down the line is going to cost you more. Public health systems have a vested interest in healthier citizens.

The majority of my gaming is on the road too but I’ve found the Steam Deck hits that niche for me. I carry a thin Chromebook for work related things. Admittedly you don’t need as powerful a GPU for a small 720p display.

How big a niche is that - because when I think high end gaming a laptop has all sorts of trade offs to make anyway.

I’m going to take a swing at the moon reflects a relatively uniform spectrum of light from the sun but our varied atmospheric conditions can alter the refraction of that light.

If it’s mentioned up front and fixed then it’s fine. One way or another the restaurant needs to cover it’s costs and it’s either done via inflating the price of the food or a fixed service fee.

What I hate is a discretionary tip suggestion because suddenly I’m made to be responsible for how much the staff get.

On the potentially bright side maybe this will make people think harder about which model to use for which task. You don’t need to feed your entire code base into Opus when a Gemini Flash sub-agent can do a perfectly fine job running grep and compiling a summary for the main agent.

It will be fun watching those users who first make the jump to the new project.

If it’s finding valid vulnerabilities then it’s just another tool like static analysis, fuzzers and sanitizers. There definitely seems to be a difference in quality compared to earlier generations that were behind the sloppy avalanch of reports.

Funny 😂

They don’t have to be. They know what they asked the LLM to do. They know how much they adapted the output. You usually have to work to get the models to spit out significant chunks of memorised text.

No, that’s why the author asserts that with their signed-of-by. It’s what I do if I use any LLM content as the basis of my patches.

If the 2-10% is just boilerplate syscall number defines or trivial MIN/MAX macros then it’s just the common way to do things.