Peter wrote:
Claude just writes the stuff as needed.
I agree, it’s one of the most impressive things that it does. It’s exceptionally good at implementing well documented RFC standards, HTTP, the TCP state machine, etc. But I would still not trust it because of security issues, and for such critical things as TLS I still vastly prefer an actual (vendor originated, even if it’s nonprofit and open source) library. But that equation is obviously being shaken today because noone actually looks at the code of open source libraries anymore (only AIs do).
Even Linus is now accepting Claude code into linux, if that doesn’t mean we’re doomed, I don’t know what will…
Linus is an old guy now 
One problem with OS libs is that if your product is sold to customers who know what is inside (e.g. MbedTLS v3.3) and the MbedTLS people push out v3.4, your customer will be onto you demanding to know why you are shipping an “obsolete version”. And that crowd was doing new versions more often than the proverb about knickers 
Whereas if you get Claude to knock up application specific code, which rolls up e.g. LWIP and MbedTLS but only the actually required functionality, you avoid that problem. I was massively impressed at the way it knocked up WFM200 code for the stuff I mentioned above, and it was all working on just one Sunday afternoon, compared with 2-3 months’ work doing it conventionally. But I can still understand the code if I look at it.
Pretty obviously it picked up all this around the internet, github, etc.
And “aviation” apps are appearing almost daily now
We don’t approve such signups. Regular contributors can advertise of course but these people just want to chuck in an advert…
This is “AI” generated speech.
https://share.google/lMsP1vQQat5G7ZN2w
I am glad “she” manages only a working class accent otherwise we would all be doomed 
Impressive but the London/Estuary accent in the “oh my god” one is a bit wobbly. You probably wouldn’t notice if you weren’t listening for it though.
Just removed a spammer who looked like a guy from a Czech flying school, who used chatgpt or some such to generate posts for appending to a load of old threads.
It was blatent advertising (and most aviation business people do not participate usefully before dropping in adverts) but in general is a problem a lot of forums face.
The appended posts are not obviously fake unless read by someone familiar with the topic. So those people we have had here who chucked their toys out of the pram because a mod/admin was also participating in the forum are talking nonsense, because a non aviation knowledgeable mod/admin would never spot this dross.
This was one of the posts (with the advert deleted):
Claude can now draw almost reasonable circuits. It could work out what they do a while ago.

Althought to be fair it did see my circuit first and it used that as a guide to the layout.
I would suggest making it use scripting tools or formal languages to describe the circuit, instead of outputting an image directly. I’m pretty sure it would work wonders there, and it could audit its own circuit by making the script run. Then you can generate the image using standard exporters.
I was considering making it conceive a Kicad PCB from scratch starting from a specification, via python scripting. Claude would simply make the declarative / creative scripts and then export to Kicad via its Python API. Or make it complete a 16-bit custom ISA CPU I made for fun on FPGAs – pretty sure the VHDL code base is big enough that it should do it no problem.
Drawing circuits is not easy. It takes good spatial perception as well as intuitive understanding of circuit operation.
It has been worked on for probably 40 years now.
Doing a messy job is easy.
Claude will use libraries if you tell it to. I have an embedded project with a library implementing some of the C lib functions (e.g. malloc) which are wrappers around what already exists in ROM so they don’t have to exist in libc, I had Claude generate a hardware testing tool and it used the library (and the specs file that modifies the linker process to ensure my library functions get put into the static binary rather than the newlib ones) because I told it that it should.
I’m pretty impressed what Claude will do given a decent spec, e.g. I needed an emulator for a hardware project, and given that the Verilog code already generates synthesisable hardware, Claude could work with it and spit out a sufficiently accurate emulator. It also found a couple of bugs in the Verilog in the process (which I fixed myself, and then had Claude verify they were fixed).