Thanks for the info. Shouldn't the repositories only allow to merge reviewed code though? Which can also be bypassed, but then you need two compromised accounts instead of one.

I am trying to have a candid and polite discussion here. I asked a question, you didn't answer it, you insulted me and complained that I didn't answer your question. I don't understand the need to antagonise. I don't appreciate the double standard (you expecting me to answer your question while refusing to answer mine). Get angry all you want, but without me. Good job convincing nobody and making people want to disagree with you.

To be clear, I an not discussing legality but morality. Legality can change from one place to another, but we can still hope to reach an agreement on morality.

The claim I was responding to is that the person who put the prompt in and the warning in the release notes was "intentionally destroying somebody's data". I disagree with that claim and I use a relatable analogy of a simpler situation.

I believe my analogy is good, do you agree? If not, why not? If yes, in the analogy, did I throw a water balloon on your head?

Sure, but nobody did that here.

If I put a big red button and a sign that says "do not press the red button, it is not meant to be pressed", and you press the red button anyway, and that releases a water balloon that falls on your head, did I throw a water balloon on your head?

I agree that it's a dick move. I also don't think it should be illegal. Wishing financial ruin through litigation is also a bit of a dick move by the way. The guy is taking a stance and mentioning it in the release notes.

Regardless of intention, this is raising awareness, and the only way this causes more than a mild inconvenience is if you have zero backup and zero version control. Now imagine if, instead, the instructions had been about stealing private data. People should be aware that this kind of vulnerability exists and is essentially impossible to defend against if you give that kind of access to your agent.

It's one way to look at it. Another is that the person responsible for the damage being done is the person giving the LLM access to do said damage. The LLM is unpredictable. It's like letting a dog loose in a kindergarten and, if the dog bites one of the kids, saying that the harm to kids is done by the dog. Technically correct, but the person who brought the dog bears responsibility.

It only does in a certain system that encourages it. Tax processed food enough to pay for healthcare, give people access to fresh veggies (already the case in many places but not everywhere, can be subsidised), give people enough free time to cook by capping contract hours and setting a decent minimum salary, make cities walkable with grocery shops around the corner or easily reachable by public transport.

I'm just describing life most European cities by the way, even though things are changing slowly.

I'm sure you're familiar with the countdown protocol, all the pre-flight checks, etc. These power up a range of subsystems, motors, etc, so that everything can be verified prior to ignition itself. The complete sequence takes a very long time. Under normal flight conditions, you can't check for absolutely everything (instrumentation is mass, and mass is the enemy) but there's still a lot. However, during an engine test, you can pack a lot more sensors in.

This is where you'd want to be spotting loose connections, pumps that aren't quite even, pressures that aren't as steady as they should be, vibrations that shouldn't be there or do not match expectations, turbulent flows, and so on.

At ignition, it takes between 3-6 seconds to go from stopped to 90% thrust. For humans, that's near-instant. For a computer sensor that's operating a million samples per second, that's 3-6 million readings. A computer performing a billion calculations per second shouldn't have much difficulty in comparing 3 million readings against model predictions and determining if both the values themselves and the rate of change at each point such a sensor exists are all good. Emergency shutdowns during those first 3 seconds are perfectly viable.

Vibrations are the ones that are likely the most interesting, because those are likely to change before something breaks, not sure how fast you can make infrared sensors, but that's also an area where things are likely to alter before point of failure.

Lazy, yes. Bright, no. If you can't trust the average person to figure out that bread that's loaded with salt and sugar isn't healthy, then you can't trust the average person to figure out what a healthy balanced meal is.

The evolutionary pattern was created because food was unreliable and energy demands were unpredictable - but high, due to the large brain. (Possibly larger than it is today, but there seems to be conflicting data there.)

Now, rationing extreme energy foods is certainly one option, but it's not a particularly satisfactory one as the energy demands vary by profession and by time within a profession. You simply can't predict what people will need and there's no way to standardise this.

There is a second option. Intense focus is impossible for beyond about 45-90 minutes at a stretch, or for more than 3-5 hours in a day. Meetings degrade intelligence, according to psychological research, so you want to minimise those. After about 7 hours, work will mostly have negative value. If you increase the amount of high physical activity for at least an hour a day (and potentially longer if the amount of soft work is minimal in the job) then you will improve physical fitness and general health, without having to substantially alter diet. However, that still only gets you so far, because a poor diet still impacts physical and mental health, and can lead to brain decline. (It's a big factor in poor brain health in children in schools.)

A third option, then, is to actually improve meal quality in schools and for workplaces to work with the food industry to provide cheaper/easier access to high quality foods that actually taste good, not merely sensible energy foods. This would seem to be target solution, with in-work exercise to supplement it.

If that turns out to be the case, we get to find out something about what reshaping the brain means.

Whilst that is perfectly true, it is questionable as to whether it is useful or necessary. If a rocket is being tested, then logically it should be heavily instrumented. If it's heavily instrumented, and the instruments are themselves competently designed, there is no obvious reason why the engine can't be auto-cut when problems start to arise. And they will have arisen long long before the explosion.

The values may have independently been "within permitted range", but if the pattern of those values doesn't make sense, then something has gone wrong. There may well also have been subsystems that were insufficiently instrumented.

"They're the experts" is often an irrelevancy - we lost TWO shuttles and crews to political decisions, when the experts on the ground were ignored. DeHavilland lost endless Comets to basically the same blunder, when political decisions by management over the reality of metal fatigue overrode analysis by actual experts. Improper monitoring and inadequate computer controls will be from a burden of costs and time (both political constraints, not engineering constraints). As, indeed, will improperly manufactured parts, improper software (anyone rememebr Arianne IV's mishap due to buggy software?), improperly-defined constraints, and inadequate quality controls.

The experts are usually either well aware of mistakes or afforded no means of detecting them.

I see no reason not to think this was anything other than a management blunder.

Not sure why this is modded as "troll" when this is describing something that has happened.

Of course the question remains of whose definition we should use. But people would rather be divided by a word than actually talk.

Second link is incorrect and should be this https://en.wikipedia.org/wiki/...

I don't know why you try to minimise how evil the CIA is. Maybe you don't have all the info, or maybe you think it's fine to do evil things under the pretext of national security.

If it's the former, here is a starting point: https://en.wikipedia.org/wiki/...
I hear this book is also good: https://www.jeffsachs.org/news...

If the latter: can we please agree that arranging coups in distant countries has nothing to do with national security? When the US was attacked as a nation on its soil, it was by an organisation that had initially been supported financially by the CIA, arguably indirectly. So the CIA carries some responsibility for the only recent attack on national security in the US, that I can think of.