"And it tried to stop us in no man's land. Insane. Anyone seeing this behavior would have thought I was drunk. So next time something like that happens if I'm in FSD mode, I'm going to cancel it by turning the steering wheel slightly and taking over fully."

I won't say this happens regularly, but it certainly has happened to me several times over nearly five years of Tesla ownership. You always have to be ready to take over control of the car in FSD in an instant. It is nowhere near being a license to ignore the road, traffic, and surroundings. I hope to see that someday, but we're not there just yet.

tags: teslacarsai

posted by matt in Sunday, September 10, 2023

"Starting in November, Google says advertisers must include a disclosure when an election ad features 'synthetic content' that depicts 'realistic-looking people or events.'"

Fine print will never be the same.

tags: ai

posted by matt in Wednesday, September 6, 2023

Basically what we're trying to do is amplify what's considered good behavior and then de-amplify what's considered bad behavior.

Sounds good. But, um, 'considered' by whom?

tags: aitechlawpolicy

posted by matt in Sunday, April 30, 2023

Thaler told the Supreme Court that AI is being used to innovate in fields ranging from medicine to energy, and that rejecting AI-generated patents "curtails our patent system's ability - and thwarts Congress's intent - to optimally stimulate innovation and technological progress."

Congress can change this at any time. I bet we'll have Congressional hearings on AI inventorship over the next few years, which will be fascinating. For me, at least.

tags: lawpatentsaisupremecourtcongress

posted by matt in Monday, April 24, 2023

A strapline added: "it sounded deceptively real”. Inside, it emerged the quotes had been produced by AI."

This is interesting. Sounds like the magazine didn't hide the fact that the interview was AI generated (although they likely deemphasized it). The magazine essentially used Schumacher's name and his likeness without his permission. I suspect they'll regret the "deceptively real" descriptor.

tags: ailaw

posted by matt in Thursday, April 20, 2023

"...no legal basis to justify the mass collection and storage of personal data for the purpose of 'training' the algorithms underlying the operation of the platform."

Italy is taking an interesting angle on guard railing AI—privacy.

tags: techai

posted by matt in Friday, March 31, 2023

"Midjourney is putting an end to free use of its AI image generator after people created high-profile deepfakes using the tool."

And they're surprised by this? Really? Give me a break.

tags: techai

posted by matt in Friday, March 31, 2023

I've played around with ChatGPT a bit over the last few days. I'm wondering how it might be useful for Daystream. I can't get past the basic inaccuracies in the system, though.

For example, today I asked it to "Describe some events or happenings that occurred on March 29, 1983." I thought API calls using a prompt like that might be an interesting way to generate user-independent content for historical days. I'm not sure why I picked 1983, but 40 years ago today seemed like a decent test. I was 13.

ChatGPT quickly gave me a list of seven events that, according to the AI, "occurred on March 29, 1983." The first event on the list caught my eye because I have specific memories of it: "The final episode of the television series 'MAS*H' aired on CBS, drawing a record-breaking 125 million viewers in the United States."

Unfortunately, ChatGPT got this one wrong. Basic web research using Wikipedia and IMDB reveals that the final episode of MASH, Goodbye, Farewell and Amen, actually aired on February 28, 1983, 40 years and 1 month ago.

This is an easy one to get right, too. The last episode of MASH is generally considered to be one of the most-watched scheduled television episodes of all time. It's the GOAT of episodic television. If AI got details wrong on something that is so easy to verify, what might it get wrong on things that aren't as easy to check? Or that can't be checked? What about a historical event or item that has such a low level of general interest that people only touch its details once a generation, once every other generation, or less? Human verification of AI generated historical content isn't just something that should be done, it's something that must be done to avoid a quiet rewriting of the details of human history. Anything that isn't human-verified should be labeled as such, and treated accordingly.

ChatGPT, and AI generation of content generally, still intrigues me and I think there might be a place for it in Daystream at some point. But, in light of errors like this that are revealed with basic fact-checking, I currently have no confidence in using it to assert that something actually happened on a particular day in the past, or that a list of various things occurred on a particular day.

tags: daystreamtechai

postposted by matt in Wednesday, March 29, 2023

"AI labs and independent experts should use this pause to jointly develop and implement a set of shared safety protocols for advanced AI design and development that are rigorously audited and overseen by independent outside experts," it continues. "These protocols should ensure that systems adhering to them are safe beyond a reasonable doubt."

We need Asimov's Laws of Robotics for AI. Frankly, I'd feel better if that came in the form an actual law, not an agreement among companies. An international treaty would be best.

tags: techailawpolicy

posted by matt in Wednesday, March 29, 2023

I don't fully understand the math involved, but I appreciate the significance of the advance made here. And the potential practical applications are interesting - the author sees a role for it in weather predictions...and climate change

tags: mathaiweatherclimatchange

posted by matt in Thursday, November 5, 2020