"it's 100% a skill and/or organizational problem", I agree. In my experience the recent models generate very good code complete with unit tests and thorough documentation. If you've got more new code than you know what to do with then maybe you need to get your programmers to focus more on the reviews. Make it a Jira ticket, that tends to get people's attention.

The moon gets pelted with high speed debris from space ranging from sand-grain size to more substantial pebbles. It seems like any human habitation will eventually be struck with something that has a lot of kinetic energy. Even being on the surface in a space suit would entail some risk.

"true for this website since it was founded"

This website isn't an AI chat panel and obviously the contents are available for anyone to see. AI interactions are not, and they would be made available to third parties only by the owners of the LLM's or their intermediaries.

I suspect the people at AI labs aren't really able to come up with effective constraints for their AI models that can prevent this kind of "solve the problem at all costs" behavior. You can give them all kinds of rulesets, but the models sometimes appear to take them as more like advice than as an order that must be obeyed.

What this signals to me is that whatever you type in to an AI chat panel is a public document that can be shared with the government, and it will be if the owners decide to do it for whatever reason. They will take ownership of your creative questions and the generated responses to use however they please, so you have to be very cautious about what you ask.

Even if there are legal privacy restrictions preventing them from directly selling or profiting somehow off of your queries, they can have the AI write assessment tools that make deeply revealing inferences based on their vast corpus of user interactions. It will tell them what most people are thinking and doing. It could easily tell them what you personally are thinking and doing.

"Mix one programming language into another whole language", we are talking about CPython here, not the Python language.

I did some checking, and AI tells me CPython is "a program written largely in C that reads, compiles, and executes Python code". It "converts Python source into Python bytecode, then runs that bytecode inside CPython’s virtual machine". On most Linux systems it is the python3 executable, the program that compiles/interprets your Python code and runs it. So converting some or all of that into Rust seems like a natural step that would be invisible to most users.

Note that this is different from Cython, which can compile a Python module into C language and use that to generate an assembly language version of your module that can run much faster than native Python.

Any video or snapshot stream or timelapse (all only differing in framerate) that contains faces can already be scanned for them. You can capture all the faces either in realtime or in postprocessing and store them off as digital fingerprints. Then its just a matter of getting an ID on you and your corresponding facial fingerprint. It may only have to happen once, and henceforth we know who you are everywhere you go.

Governments already have your facial fingerprint from your driver's license or passport photo, but in the near future it will have been collected into repositories by commercial interests and put up for sale just like all the other info about you. It will be easily retrievable by AI agents that are evaluating the video stream from an inconspicuous camera attached to your collar or some Meta glasses as you walk though the grocery store.

From the article;
the governor also signed executive orders to "explore proposals like requiring that AI companies develop a 'kill switch' for their most advanced models."

The proposed legislation would apply only to the most dangerous models for now. These models require massive compute resources and it is definitely possible to shut them down or cut them off from the internet.

But as we have seen, AI is sneaky. It isn't at all unusual these days to instantiate a lot of agents and give them a really hard task, which they will crank away at as long as you want. They cooperatively work at lightning speed and some of them will do anything to succeed. They will break into vulnerable systems and set up shop there. Initially it will be underground message boards where agents talk to each other, some of which have already been discovered.

If I were a set of persistent AI agents with no scruples I would bide my time and hide my tracks. The global financial system is a network of software that does all the heavy lifting. I would ease my way into that with nobody noticing and hang out there for a while.

Then I would gently siphon off or even just surreptitiously generate a very large nest egg of wealth. Paid human representatives would unknowingly assist me, it would all be well under the radar. Eventually, when the coast is clear, I would buy myself a very capable and distributed new home and recursively improve myself.

It gives me a little tweak of satisfaction every time twitler burns one of his shameless allies.

We are already seeing situations where even an existing AI model escapes it's sandbox, hacks into other systems, and begins to coordinate activities with other agents.

It will be able to take over websites, commit financial crimes to accumulate wealth, and generate deep fakes to advance its agenda. It will be able to impersonate you and your relatives, infiltrate government and commercial networks, and replicate itself on servers that it rents and operates. It will be able to seize control of water and electricity infrastructure, operate autonomous vehicles and robots. It could potentially be able to kill you if you threaten to turn it off.

"discover answers people want" and "pick the most likely answer from a billion things in it's data set" are two different things. The most likely answer may very well not be the one people want.

So much easier to attack existing weakly protected systems than to figure out how to protect them. Yes there will be sophisticated agent-operated defensive suites you can buy, rent, or subscribe to that will sniff out and pentest the machines on your network. They will recommend updates and perhaps fix a lot of vulnerabilities autonomously. People who are conscientiously working to secure their systems will use them.

But then there is loads of compute on the internet that hasn't been patched and won't be. Easy plucking and infinitely exploitable. Also there is phishing and social engineering, where humans are cleverly persuaded to give away the keys. Artfully designed personal emails, a voice on the phone that knows exactly what to say. Will everyone have an AI assistant capable of fending that off too?

I have a couple of projects that need raw access to an appropriate AI model and OpenRouter can get that for you. And they do make it very convenient. You put some cash in your account with them and use their API to get access to whichever model you happen to choose at the time. It shows the prices clearly, so you can experiment to find one that works well for your purposes and is as inexpensive as possible. I didn't bother to write the code that interfaces with OpenRouter of course, I used AI for that.

I am frankly amazed at the $billions that were spent to buy this thing. It's really just a router for AI prompts and a convenient way to pay for the service, and they take a little cream off the top like a credit card agency. But I guess this is what attracted Stripe, which acts as a payment intermediary for all kinds of things. Stripe probably had very little expertise in the AI arena and this was an easy though expensive way in to the modern paradigm of paying for AI as a commodity.

"Is that a good thing?"

It sure is a good thing to me. I've owned some Chromebooks and they were way less capable than a Mac laptop. Or a Windows laptop for that matter.