Never talk about fig....No, wait, that's management.

To do QA well is very very hard, but you have to absolutely hammer everything, test extremes, etc. If the QA people are playing it safe, it's because they're nervous that it's a really bad product. QA should haunt the dreams of developers and fill them with existential dread, to the point they don't make silly mistakes.

It is. The reviews and "we did this amazing pocket-protector design by AI" columns don't really tell us what AI can do, what the limits are, and where the problems lie. As long as AI is never used for anything that you'd actually want/need help with, this isn't something anyone can really answer.

Can AI design a piston engine? If the answer to that is yes, then it's obviously capable of handling incredibly intricate constraints and optimisation.

Chances are, the answer is no, but it'd be good to know. The other designs are lower complexity and have a much narrower band of areas they combine.

I would need to know how well these would work, but the ultimate goal is to develop a large set of such projects, each testing different permutations of abilities the AI would need to have to do anything useful at this level of complexity.

Once I have a heat-map of AI talent, then people can use AI where it's actually better. And we really don't know that right now.

Nobody comes up with interesting conspiracy theories, they're all boring and much of a muchness.

To combat this, I am working on some next-gen conspiracy theories which will improve this shoddy market.

1. Line noise was caused not by faulty connectors, but by a herd of buffalo that wandered onto the networks by mistake when someone left the back of a network cabinet open. What you were hearing was the buffalo grazing on passing connections.

2. Symphonic metal was invented as a way to smuggle the elder gods, disguised as musicians. to Earth without anyone noticing. This explains the typical themes of songs and why Odin and Loki have played bass for Nightwish.

3. The missing Doctor Who stories were all penned by actual Time Lords. The High Council found out and tried to erase them from existence.

This is great! Mythos = lots of eyeballs. Now tell us how many of each severity level were found, how many of those could be fixed automatically, and how many fixes both auto and manual then were found to introduce a vulnerability upon reanalysis. Though if there was even one critical severity bug found out of 271 that makes it worth it.

p.s. for people saying it is a good way to need rockets, okay. And I am definitely all for going to space. But you can also build megarobots that can shrug off Antarctica and be self-driven / self-walking. Personally they are both cool to me. And both would be useful for building a Moon base too.

Just wondering why not just build the data centers somewhere really cold? Antarctica has got to be easier than working in space.. at least you can get there without a rocket, and you can dig underground. I am not sure I understand the economic argument for space-based data centers as opposed to ground-based data centers. Space might have more solar flux than the poles but has anyone actually compared the costs?

Ok, to clarify a few things:

Current designs I've put up:

1. A modernised version of the DeHavilland DH98 and Merlin engine, where I basically fed ChatGPT and Claude with all of the known historic faults and some potential solutions to various problems, then let them run wild, feeding off each other to fix, refine, and clarify the various design. The premise here is that we're using known designs with known properties, changing only materials but doing so carefully so as to ensure that the balance is unchanged from the historic design. The aircraft is probably the least interesting part, as it would be very hard to make that safe, but a fully modernised Merlin that starts where Rolls Royce left off is something that could be built with minimal risk and could be quite interesting in its own right.

2. A High Dynamic Range microphone. This basically riffs off assorted physics technologies for measurement and the basic idea in many HDR schemes that you can split an input into the fine detail (essentially an equivalent of a mantissa) and a magnitude (essentially an exponent), producing a design that aught to permit (if it works) the same microphone with no adjustments handling everything from a nearby whisper to the roar of a jet engine -- but with all of the fine detail still captured from that engine.

3. An electric guitar that operates not by magnetic pickups but by accurate mapping of string behaviour in two dimensions via lasers, where this is then turned into an accurate representation of the sound in an external device. So it's not a synth guitar in the classic sense, it's actually modelling the waveform for each string in two dimensions precisely. The reason for doing 2D modelling is that this has the potential for novel behaviours but without an obligation for it to do so.

4. A synthesiser/wave processor that looks at everything that they knew how to do, and allows you to link it together arbitrarily. It is designed in two forms. The first is engineered to match the components, materials, and knowledge available in 1964, so it is something they could have built if sufficiently insane. The second is a modernised extrapolation of that, using modern digital electronics, where I can show that the modern version is a strict superset of any existing DAW, simply because I started with none of the assumptions and metaphors around which DAWs were subsequently designed.

5. Multiband camera. An attempt to build a digital camera that is far smaller and more compact than a 3CCD camera, but (like the 3CCD design) produces a far better picture than a conventional digital camera, where I don't stop at three frequencies but support many, albeit with the limitation that the time required for a photograph is abysmal.

Each design I've put up has a detailed hardware specification (including wiring where appropriate), validation/verification documents, and testing procedures. Software is defined by means of formal software contracts and occasionally Z-like forms. The designs are extremely detailed, although not quite at the level you could build them right there and then. However, the synthesiser is described right down to the level of individual transistors, diodes, and connectors, and the Merlin engine specifies precise materials, expected temperature ranges, material interactions (and how they're mitigated), and other such information.

Again, it's precise but not quite at the point where an engineer would feed comfortable feeding the specifications into an AI, having it order the bits online, and be sure of building something that works, but it's intended to be close enough that (provided the AIs actually did what they were supposed to) that an engineer would feel very comfortable taking the design and polishing it to working level.

If, however, an engineer looking at these designs comes to the conclusion that the AIs were utterly deluded, then obviously they can't handle something as simple as selecting candidate items from ranged data.

I post the AI-derived designs to my gitlab page:

https://gitlab.com/wanderingne...

I really should add this to my journal.

This is really creepy / nauseating to me, and also creates a high value target at World. You can reset leaked passwords, but you cannot reset your retina. If you choose to believe 100% in this service and willfully ignore implementing patterns to combat social engineering this could end up worse. From what I can gather (from Gemini), deepfakes take advantage of organizational social pressure, like a CEO demanding something instantaneously, or hackers being inside your email system for a long time. Perhaps this could be short-circuited by an organization actually requiring people to always call back officials on a secure phone number, confirm with shared personal knowledge, and never respond to a demand without out of band verification. I also wonder what if Zoom just calls the participants in such high-stakes meetings, instead of allowing participants to click on a link they trust because someone emailed it to them. Then the corporate security office can just verify the Zoom server. And iPhone/Macbook already have biometric sensors too but have Secure Enclave.. yes there is a big value in being able to identify someone for sure but putting it all in a single company's hands sounds like waiting for them to be attacked.

With the ESA supplying the spacecraft, most of the software is likely to be competently written and/or open-source. This will prove to the Martians that there is indeed intelligent life on Earth.

It's a good question and one I'm working on trying to get an answer to. By giving AI hard, complex engineering problems, and then getting engineers to look at the output to determine if that output is meaningful or just expensive gibberish.

By doing this, I'm trying to feel around the edges of what AI could reasonably be used for. The trivial engineering problems usually given to it are problems that can usually be solved by people in a similar length of time. I believe the typical savings from AI use are in the order of 15% or less, which is great if you're a gecko involved in car insurance, but not so good if you're a business.

If the really hard problems aren't solvable by AI at all (it's all just gibberish) then you can never improve on that figure. It's as good as it is going to get.

I've open sourced what AIs have come up with so far, if you want to take a look. Because that is what is going to tell you if good can come out of AI or not.

I would think it goes further. If the company has already failed (ie: no longer exists) then the action is not taken by the company but a former employee of that company (even if said former employee was the CEO). Former employees are not granted special authority over PII or over company-owned information.