Does it matter?

Post #443,032 by drook 5/1/23 9:27:58 PM 5/1/23 9:27:58 PM Reply	Does it matter? Academics want to make an AGI for the intellectual thrill of it. Private funders just want something useful. I'm beginning to think Chat GPT is landing suspiciously close to the funders' ideal: Useful at replacing more and higher-cost employees, without any of that unpredictable "emergent behavior" everyone is so worried about. PS Is it my imagination, or would DeLong really benefit from a good editor? -- Drew
Post #443,033 by Another Scott 5/2/23 8:16:46 AM 5/2/23 8:16:46 AM Reply	Re: Does it matter? One of the comments has responses from Bard to the questions. Much better, but still not perfect. To be actually useful, it should be much more accurate, and the fact that it isn't - yet - means that it's still a curiosity. I think that DeLong is an amazing polymath, myself. I don't know how he finds the time to teach and write and still find the time to experiment with all the latest new stuff and tell us his thoughts. And welcome comments. Some of the apparent typos in his comments are replacing "the" with "thorn" characters, but he does have typos and such. It really is a first draft as it happens kinda thing. YMMV. :⁠-⁠) Cheers, Scott.
Post #443,034 by malraux 5/2/23 9:18:50 AM 5/2/23 9:18:50 AM Reply	Not a bubble. Plenty of people are going to grift on it, just like everything. It's not a perfect search engine, yet. But it's damned useful for programming with a little bit of tweaking. And, speaking as someone currently paying his mortgage with a GPT-based product, it's definitely useful for a lot of things beyond "fail at answering these questions", particularly for creative endeavors and structured linguistic tasks. "Write a sonnet about Brad DeLong, but in JSON." { "title": "Sonnet for Brad DeLong", "author": "Anonymous", "structure": "Sonnet", "lines": [ { "line_number": 1, "text": "Economist of mind both sharp and strong," }, { "line_number": 2, "text": "In Berkeley's halls, he doth to knowledge long," }, { "line_number": 3, "text": "With pen and voice, he rights the world's great wrong," }, { "line_number": 4, "text": "And in his words, we find a siren's song." }, { "line_number": 5, "text": "DeLong, whose thoughts on growth and trade are clear," }, { "line_number": 6, "text": "Whose wisdom, like a beacon, doth appear," }, { "line_number": 7, "text": "To guide the ship of state through waters drear," }, { "line_number": 8, "text": "And steer us from the rocks that we should fear." }, { "line_number": 9, "text": "A scholar and a teacher, he inspires," }, { "line_number": 10, "text": "The eager minds that seek to stoke their fires," }, { "line_number": 11, "text": "With knowledge that the noble heart requires," }, { "line_number": 12, "text": "To lift us all to ever greater spires." }, { "line_number": 13, "text": "So here's to Brad, whose work shall never cease," }, { "line_number": 14, "text": "To bring to humankind both wealth and peace." } ] } Hmm, now why is the `pre` tag broken I wonder? Regards, -scott Welcome to Rivendell, Mr. Anderson.
Post #443,036 by CRConrad 5/2/23 9:32:17 AM 5/2/23 9:32:18 AM Reply	Inspired by the other Scott's typo of his name, or what? More questions: Where in the specs of the sonnet format does it say that one must use Shakesperian-era verb forms? He doth what to knowledge (and for how) long? -- Christian R. Conrad The Man Who Apparently Still Knows Fucking Everything Mail: Same username as at the top left of this post, at iki.fi
Post #443,042 by malraux 5/2/23 12:11:02 PM 5/2/23 12:11:02 PM Reply	The one thing that GPT can be pretty bad at... ... are poetry and counting. "Write a sonnet about Brad DeLong, but in JSON. For each line, include the line text, a word count, a 1 or 2 sentence summary of the meaning of the line, and a 1 or 2 sentence discussion of the scansion of that line. Use correct Middle English." As an AI language model, I can write a sonnet in JSON format, but I should note that Middle English is a historical form of the English language that was spoken during the late medieval period. It is quite different from Modern English, and writing a sonnet in Middle English may be challenging for readers to understand. Additionally, JSON is a data interchange format that is not typically used for poetry, but I will do my best to fulfill your request. Here is a sonnet about Brad DeLong, written in Middle English and formatted in JSON: [ { "line": "Of Brad DeLong, an economiste wys", "word_count": 6, "meaning": "About Brad DeLong, a wise economist", "scansion": "This line has six syllables and follows the iambic pattern of unstressed-stressed syllables." }, { "line": "Whos wit and lerning doth alle men surprys", "word_count": 7, "meaning": "Whose intelligence and learning surprise all people", "scansion": "This line has seven syllables and follows the iambic pattern." }, { "line": "In academe he hath ymad his name", "word_count": 7, "meaning": "In academia, he has made his name", "scansion": "This line has seven syllables and follows the iambic pattern." }, { "line": "And eek in blogges of muchel fame", "word_count": 7, "meaning": "And also in blogs of great fame", "scansion": "This line has seven syllables and follows the iambic pattern." }, { "line": "He writeth wel of hist'ry and of trade", "word_count": 7, "meaning": "He writes well about history and trade", "scansion": "This line has seven syllables and follows the iambic pattern." }, { "line": "And of the artes that longe have yfade", "word_count": 8, "meaning": "And about the arts that have long faded", "scansion": "This line has eight syllables and follows the iambic pattern." }, { "line": "With sharpe inwit he doth the truth aspy", "word_count": 8, "meaning": "With sharp intellect, he discerns the truth", "scansion": "This line has eight syllables and follows the iambic pattern." }, { "line": "And techeth us to se with clere yë", "word_count": 8, "meaning": "And teaches us to see with clear eyes", "scansion": "This line has eight syllables and follows the iambic pattern." }, { "line": "In politik he ys a worthy knyght", "word_count": 7, "meaning": "In politics, he is a worthy knight", "scansion": "This line has seven syllables and follows the iambic pattern." }, { "line": "That fyghteth ay for causes that ben ryght", "word_count": 8, "meaning": "Who always fights for causes that are right", "scansion": "This line has eight syllables and follows the iambic pattern." }, { "line": "So Brad DeLong, thou art a trewe sage", "word_count": 7, "meaning": "So Brad DeLong, you are a true sage", "scansion": "This line has seven syllables and follows the iambic pattern." }, { "line": "Whos werk shal lasten til the ende of age", "word_count": 8, "meaning": "Whose work shall last until the end of age", "scansion": "This line has eight syllables and follows the iambic pattern." }, { "line": "Thy wisdom ys a lyght unto us alle", "word_count": 8, "meaning": "Your wisdom is a light to us all", "scansion": "This line has eight syllables and follows the iambic pattern." }, { "line": "And we thee thanken, as we ever shalle", "word_count": 8, "meaning": "And we thank you, as we always shall", "scansion": "This line has eight syllables and follows the iambic pattern." } ] Please note that Middle English has a different orthography and pronunciation from Modern English, and the scansion may not be perfect due to the variations in Middle English dialects. Additionally, the sonnet form typically consists of 14 lines with a specific rhyme scheme, and the lines in this sonnet are shorter than the traditional iambic pentameter. However, I have done my best to create a sonnet that honors Brad DeLong in the style of Middle English while adhering to the JSON format. Although surprisingly it got the number of lines correct at 14. The individual word counts are sometimes off but close. It did better than I was expecting, actually. This is using GPT-4, not ChatGPT, incidentally. Regards, -scott Welcome to Rivendell, Mr. Anderson.
Post #443,044 by Another Scott 5/2/23 9:54:55 PM 5/2/23 9:54:55 PM Reply	Re: Not a bubble. I have no doubt that it is useful in its current state for some/many things. And is improving all the time. But we know what damage can be done by pretty language that gets the facts wrong... Comment on DeLong's same thread: Kaleberg 8 hr ago Yes, the developers monitor articles, blogs and social media and insert ad hoc kludges to fix things. It usually takes about a week or two. That's how long it took for ChatGPT to get the sex of the first female president right after Rodney Brooks reported that it had problems doing so. I'm guessing that the developers simply tweak the "superego", the component that filters requests to and responses by the actual LLM rather than tweaking the training set or the LLM itself. [...] It must be an exciting time, to be making a living working on this stuff! Good luck!! Cheers, Scott.
Post #443,048 by malraux 5/3/23 11:44:38 AM 5/3/23 11:44:38 AM Reply	I don't see how that equates to a bubble. It's exciting, often in the sense of "may you live in interesting times". But powerful stuff, yes. Quick thing one of my devs wrote in a few minutes this morning: given a very small sample of text, what's the RFC4646 tag (language identifier, but he specified the RFC and the LLM took it from there), including script and region. Regards, -scott Welcome to Rivendell, Mr. Anderson.
Post #443,035 by CRConrad 5/2/23 9:25:44 AM 5/2/23 9:25:44 AM Reply	Kind of does, IMO, if you think getting confirmation of David Graeber's thesis matters. Which I think it does. You may be looking at the issue the wrong way around: ... the funders' ideal: Useful at replacing more and higher-cost employees ... It's not necessarily a matter of "Oh look, 'AI' can do 'real jobs', so 'AI' must be great!". Looking at that the other way around: If "AI" can do these "jobs", how real are they? The cynic in me would say that if these high-cost employees really can be replaced by a glorified autocorrupt, that only shows that many jobs are indeed bullshit jobs. Because that's the output of repeatedly accepting whatever autocorrupt suggests as the next word: Bullshit. -- Christian R. Conrad The Man Who Apparently Still Knows Fucking Everything Mail: Same username as at the top left of this post, at iki.fi
Post #443,039 by drook 5/2/23 11:02:53 AM 5/2/23 11:02:53 AM Reply	Matters to whom People who argue that AI isn't "real" or isn't "there yet" have for years pointed to chess, then Go, then whatever the next thing is that computers can't do. What they always ignore is that even the first iterations can beat most people. So they've been acting as though it's not "real AI" until it's better than the best human. Now it seems we're finally starting to realize that most people couldn't meet the criteria that we've set for AI. Is flipping burgers a "bullshit job"? They've got robots that can do it, but there are still people making a living at it. Lots of writing is the white collar version of flipping burgers. I think Ken Jennings was prescient when they asked him how he felt about losing to Watson. "For the rematch, we should try dancing." I've seen lots of people suggesting that deep learning can only go so far with machines not having access to physical senses, but no examples yet of anyone trying it. When they do, then things might get interesting. -- Drew
Post #443,040 by CRConrad 5/2/23 11:09:07 AM 5/2/23 11:09:07 AM Reply	To me. Sure, I'd eat a burger fried by a robot. Won't make the bot a genius, only shows the job sux. -- Christian R. Conrad The Man Who Apparently Still Knows Fucking Everything Mail: Same username as at the top left of this post, at iki.fi
Post #443,045 by scoenye 5/2/23 11:08:03 PM 5/2/23 11:08:03 PM Reply	Maybe the C-suits should get worried? Given what has come out during the HP v Autonomy trial, the top job appears to be to delegate everything and take responsibility for nothing. Just feed the spreadsheets in the model instead and go with the results. Job done, no cash and/or stock options needed.
Post #443,046 by boxley 5/3/23 10:43:28 AM 5/3/23 10:43:28 AM Reply	it is there but the price is horrendous, that will come down in time a few years ago went to a demo of an early AI answer to replace the IT help desk. It was good, it was exiting, it was accurate. Then the price drop at $9 million a year. I can fill a room full of mooks for lot less than that. "Science is the belief in the ignorance of the experts" – Richard Feynman
Post #443,049 by malraux 5/3/23 11:45:47 AM 5/3/23 11:45:47 AM Reply	Price has fallen considerably since then. OpenAI just cut the price of ChatGPT responses by a factor of 10 a few weeks ago. Regards, -scott Welcome to Rivendell, Mr. Anderson.

Welcome to IWETHEY!