nonfamous

The NYU Professor in question is Scott Galloway, better known now for his podcasts, books and general media appearances than as a professor.

My Dad did something like this for our pool in Australia in the '80s. He ran a bunch of black PVC pipe in a zigzag pattern on the garage roof and ran water through it from the pool filter. It was pretty effective at heating the pool!

I'm a member of Community Notes and while it's common for an unpublished note to be downranked by members, I have never before seen an example of a published note (visible to all Twitter/X users) later be withdrawn from visibility. In fact, I don't think there's even a process for Community Notes members to make that happen.

Fascinating, thanks for sharing! According to that article a second source has been discovered since the linked paper was published.

There’s also the recent film, “ Pinball: The Man Who Saved the Game (2022)”, which is a fun watch.

The underlying models have not changed. You can specify the specific model variant you want via the API. (The underlying model used by the ChatGPT application may be updated, but that's not what the linked paper discusses.)

Lot of confusion in the discussion between chatgpt the large language model (aka gpt-35-turbo) and ChatGPT the consumer application (the website where you type questions and you get a response beginning a conversation, which can be configured to use either the chatgpt or GPT4 models). To be clear:

* This paper called the models directly via the API not the ChatGPT application. This means that changes to the ChatGPT system prompt and other changes to the application aren't a factor here.

* The paper compared two variants each of the chatgpt and GPT4 models. The later variants are obviously different from the earlier variants in some way (likely having been fine-tuned)

* Any given model variant has not changed. You may continue to select the older model variant when using the API if you so wish

Lastly, and this one's my opinion, problems involving arithmetic and mathematics are not a good test of large language models.

AWS is listed as a partner: https://ai.meta.com/llama/#partnerships

Not that I'd be surprised if it were, but did you perhaps try to access that account from an unusual location or device?

I ask because I had the same thing happen to a little-used Twitter account of mine. Killing time while on vacation in a foreign company, I attempted to log in, and failed. Not sure if the failure was my fault or Twitter's. Nonetheless the outcome was the same -- the account was suspended for "breaking rules", despite not posting anything. An appeal failed, and I lost the account permanently.

More context here, but it seems impossible to read a Twitter thread outside of the app these days.

https://twitter.com/kathryntewson/status/1679245990187126785

Statistics.

Let’s ignore the time aspect, and assume that the test has only a 35% chance of returning positive for an infected but asymptomatic patient.

Repeat that test 3 times, and the chances of at least one returning positive jumps to 73%.

(That’s why there are two tests in most rapid test boxes. You’re supposed to test twice, 24 hours apart, to improve the chances of a positive result if you are indeed infected.)

Actual title of linked article: "Microsoft mitigates China-based threat actor Storm-0558 targeting of customer email"

I was a big cryptic crossword fan when I lived in the UK, and turned my nose up at American-style crosswords as being too vanilla. But I finally got into the NYT crossword after living in the US for a few years, and now I love it. Clever constructions and wordplay like this example are the reason why.

I was taken aback by the terminology too. Isn’t Retrieval-Augmented Generation the standard nomenclature for this pattern? https://arxiv.org/abs/2005.11401

You can create your own instances of the OpenAI models in Azure to keep proprietary data with the bounds of your tenant there.

Generics may be a low-margin business for manufacturers, but they are emphatically not so for retailers. The likes of CVS make a lot of profit from generics, which makes it surprising that this hasn’t spurred pressure from on the manufacturers to amp up supply.

You can opt out of the safety filtering, btw.

I had a similar contract when I lived in IL, and I also liked it and it made me more conscious of my energy usage patterns.

But there was always this niggling concern that the price could spike to 200x like it did in Texas, without me being able to react. Regulation to cap the peak rate would help allay consumer fears, IMO, while still offering ample profit potential to providers.

> ? It's a button above the mirror?

Yep. FWIW, Hazard Lights is the one control that is mandated to be a physical button by US law.

I don’t think it’s such an edge case at all. I imagine it would be disconcerting to the driver behind you if you slammed on the brakes on a wet road, locked the wheels into a skid, and the brake lights went out. (Shouldn’t happen with modern braking systems, but still…)

I got accused (more as a joke, but c’mon) by my boss the other day for using GPT to write a document. His justification: the grammar and spelling was too good. I had no idea that good writing was now a “tell” for generated content for some. I I just write good, honest!

Also so telling that every Bard comparison is to ChatGPT, when the correct comparison would be with Bing.

> OpenAI must be sweating profusely after viewing that I/O presentation. Not only do they not have the integration, scale and user base of Google ...

Good thing OpenAI has a distribution partner in Microsoft Azure, then.

The announcement was from the Maine AG. I presume these are all customers in Maine. It’s likely customers in other states were affected as well, but this has not been announced.

You mean, more efficient, with better outcomes for its customers?

Over 25 years ago I wrote a package in GNU Emacs (ielm) that, as I was chuffed to learn from a HN thread recently, is still in widespread use.

Yeah, but only as a means of transferring them to another device. Sure, you could abort the flow before the existing codes were deleted, but it was far from ideal.

I’m glad there’s finally real support for backing up codes.

The snipping tool does copy the image to the clipboard, so other than having to close the window the functionality of PrScr won’t change much.

This is a great guide!

This is a blast to see. I wrote IELM at least 25 years ago, when I was learning eLisp and was frustrated there wasn’t a REPL to help me try commands and learn it. I remember it being quite a thrill when the FSF wrote to me and asked me to assign the copyright so it could be bundled with Emacs. I’m so pleased to see it’s still in use!

I do remember I had to use an awful hack to make it work. Comint expects data from a pipe, but eLisp output is internal to Emacs. The details are hazy now, I but I think I pushed the eval result as stdin to a ‘cat’ process for Comint to ingest the output. I wonder if that ever got cleaned up…