QuoteI'm surprised she wasn't excited about you buying her a new collar.

It raised an eyebrow certainly. 

Quote"Too many requests, please slow down."

One day computers will be faster than us, but it appears its still a barrier.  Makes me quite happy to hear that!

>>no way I'd trust material from this source

I would apply that rule to every source ;-)

The thing is, nonlinear processes have a habit of overtaking us suddenly.

In the 1950s or 1960s, a computer scientist had confidently stated that with a grant to support a graduate student for a summer, they would be able to teach a computer to process natural language. For 50 years there was almost no discernible progress.

In the early 1990s, I was working on a edition of articles of a theologian whose active years were mostly 1950 to 1980. His writings were published, but no electronic copies. We used OCR to try to get them in digital form and it was useless. With a serif font on glossy paper, we would get entire pages of i, j, u, m, n, l. An "s" for instance would render as two "i"s. After many years of attempting to improve OCR, it was horrible. And then, suddenly, in the space of a few years, it made a huge leap forward. The Google Books transcriptions got radically better (though they still can't handle things like the tall "s" in old books).

About the same time, a company demonstrated their cutting edge translation software to a team I was affiliated with. To demonstrate their capabilities, they had written their letter in English and sent it to the team in French. Everyone got huge laughs out of it. Now I regularly vet my French prose using free online translators. Sure, with length and complexity they suffer, but they are surprisingly good compared to a short time ago.

Those early non-linear processes often appear linear. Now, though, things seem to be moving very very fast and don't feel linear anymore. After years of slow or no progress on natural language, it seems as if we are very suddenly in the steep part of the curve. That makes it hard to see 10 years ahead.

The question always is what short of curve is it? We have seen with self-driving that it is an S-curve, meaning you get exponential growth in the middle, but then it becomes asymptotic and the growth slows every year.

With language, though, the systems don't have to drive around in the real world. They can digest almost infinite amounts of text at whatever rate the system can process it. So my guess is that if improvement in natural language is asymptotic (and I think it is), the asymptote is very far away and texts may well reach expert writer level before we hit it.

Or put another way, I would not be surprised if in 10 years the only being that can tell human-generated text from AI-generated text will be an AI with a very large sample of text from a given "writer."

It took me 1 day to create a program, using GPT-3, to create a highly convincing small army of bots to post on Reddit: Here's how I did it

https://old.reddit.com/r/singularity/comments/wa9enf/it_took_me_1_day_to_create_a_program_using_gpt3/

ChatGPT Crushes Google on Coding Queries, and (Already!) Matches it at General Information

https://www.surgehq.ai//blog/googles-existential-threat-chatgpt-matches-googles-performance-on-informational-search-queries-and-smashes-it-on-coding

The internet is increasingly awash with text written by AI software. We need new tools to detect it.
https://www.technologyreview.com/2022/12/19/1065596/how-to-spot-ai-generated-text/

The interesting part is that participants in the study tended to prefer texts written by AI. I suspect that this will follow a trajectory similar to chess. First the AI beats mediocre writers, then great writers, then a generation of writers grows up with AI editors who help them write at an even higher level, and though humans cannot write as well as AIs, people will prefer to read books written by people just as we prefer to follow chess competitions played by people, all of whom would be easily beaten by a chess program on an iPhone. And eventually, there will be some sort of certification process similar to what we have for organic foods or for wines that will verify production methods and provenance.

And then the AIs will get bored and go off and write texts for other AIs who will be better equipped to read and understand those texts.

QuoteBecause large language models work by predicting the next word in a sentence, they are more likely to use common words like "the," "it," or "is" instead of wonky, rare words. This is exactly the kind of text that automated detector systems are good at picking up, Ippolito and a team of researchers at Google found in research they published in 2019.

But Ippolito's study also showed something interesting: the human participants tended to think this kind of "clean" text looked better and contained fewer mistakes, and thus that it must have been written by a person. 

In reality, human-written text is riddled with typos and is incredibly variable, incorporating different styles and slang, while "language models very, very rarely make typos. They're much better at generating perfect texts," Ippolito says. 

"A typo in the text is actually a really good indicator that it was human written," she adds. 

For Buckworks

QuoteThings AI doesn't do that people can do really well: close reading, in the sense of teasing out tensions, assumptions and begged questions within words and phrases.

In fact, AI's bottomless fluency encourages us to stop doing this; to treat words as fungible content...mere interchangeable vessels for information. Which, sometimes and in some circumstances, they are. But even then, they can always be *read* more closely, challenged, interrogated, reframed. And this critically engaged art of reading, rereading and rewriting is precious... because it is in large part *how* we learn to think well about the world

https://twitter.com/TomChatfield/status/1638488424427540480

This gets to the point that AI researcher Gary Marcus makes, which is that LLM is only capable of and always will be only capable of generating "pastiche" never genuinely creative, original work, because it is a purely probabilistic model based on the word most likely to come next (perhaps with some "salt" to make for some variation).

This is why some AI researchers like Marcus think it is a dead end if you are looking for AI to be able to, for example, "know" whether or not the things it says are bullshit and in some way relate to the world outside of the data model inherent in the AI.

QuoteAI to be able to, for example, "know" whether or not the things it says are bullshit

A few weeks ago I asked ChatGPT to generate some lists for me and I saved them to look at later.

This week I've been examining them line-by-line and not only noticing gaps but spotting some items that are flat wrong. All presented with aplomb, of course.

For my purposes, ChatGPT is turning out to be a lot less useful than I thought it might be.

> items that are flat wrong

That brought back memories of Toolman's recipe sites.

Ha ha!

I saw a recipe that ChatGPT gave someone and it was so vague it was like "add some liquid then add a spice." I can't find it, but this is a decent proxy

QuoteIt's a neural net generated brownie recipe called Chocolate Baked and Serves, and its distinguishing feature is the CUP OF HORSERADISH it contains. It was so bad that my eyes watered as I removed it from the oven.

https://www.aiweirdness.com/ai-recipes-are-bad-and-a-proposal-20-01-31/

If you read one article today, make it this one. My wife and I were almost brought to tears laughing. We're thinking our dinner might be Crock Pot Cold Water with Chocolate Chicken Chicken Cake for desert.

Now note the date - January 2020. But I've seen things like this from GPT-3 as well.

I should say that the current version of ChatGPT when asked to give me a recipe for Chocolate Chicken Cake said, "I'm sorry, but chocolate chicken cake is not a common recipe, and I'm not aware of any recipe for it. Can I suggest an alternative chocolate cake recipe that you might enjoy?"

I am told also that similar questions report the same text... I have asked that to be checked on version 4  otherwise dup content will still be a thing.

>dup

Google Bard Plagiarized Our Article, Then Apologized When Caught | Tom's Hardware

https://www.tomshardware.com/news/google-bard-plagiarizing-article

OK, thx RC.