The AI bubble is the only thing keeping the US economy together, Deutsche Bank warns: When the bubble bursts, reality will hit far harder than anyone expects techspot.com/news/109626-ai-bu…
bad news "AI bubble doomers". I've found the LLMs to be incredibly useful and reduce the workload (and/or make people much, MUCH more effective at their jobs with the "centaur" model).
Is it overhyped? FUCK Yes. Salespeople Gotta Always Be Closing. But this is NOTHING like the moronic Segway (I am still bitter about that crap), Cryptocurrency, which is all grifters and gamblers and criminals end-to-end, and the first dot-com bubble where not NEARLY enough people had broadband or even internet access, plus the logistics systems to support shipping products was nowhere REMOTELY where it is today.
If you are expecting this "AI bubble" to pop anytime soon, uh.. you might be waiting a bit longer than you think? Overhyped, yes, overbuilding, sure, but not remotely a true bubble any any of the same senses of the three examples I listed above 👆. There's something very real, very practical, very useful here, and it is getting better every day.
If you find this uncomfortable, I'm sorry, but I know what I know, and I can cite several dozen very specific examples in the last 2-3 weeks where it saved me, or my team, quite a bit of time.
@codinghorror @mikemccaffrey @JamesHMcLaren It’s the overhyping that’s the problem; when the hype bubble bursts, it’s not just going to take down the no-value “solutions”, it’s also going to crater the actually-valuable propositions because those products aren’t profitable without the hype money coming in.
(My wife was a pre-IPO employee of Amazon; I watched the dot-com collapse in extreme close-up.)
It's not the data centers. It's the overvaluation of "AI" stocks that pushes bubble money into pension funds
When that bubble valuation goes away, the idled data centers will be there like the feet of Ozymandias, and the pension funds and 401(k)s will be the barren desert all around
@lbcp @doragasu @trochee (a) the cost of a building isn't in construction but in ongoing maintenance and operation, and (b) data centres aren't convertible into some other, more useful, facility. So it's highly focussed economic activity that is essentially pointless (hence wasted).
We share the same fate as Trump's casinos. Welcome to "government run like a business" where profit for the CEO is the only goal and your human rights and well-being are a "cost center" to be mitigated.
I'm not sure it will burst that quickly because a lot of the LLM use is incredibly practical and effective, as long as you understand what it is good at. Compare to first tech bubble (pets.com), not enough people on the internet, wildly inadequate shipping locgistics. Compare to crypto bubble: it's all gambling and grifters and bullshit. Is it being over hyped? Yes. is it practical TODAY and getting more practical EVERY DAY? Also yes!
@codinghorror I think the "bubble" talk is more about whether it's a financially sustainable business model, pouring billions of dollars into training and running the back end systems. For example I read OpenAI would need about 10x their current revenue to be profitable.
@codinghorror Some of it is useful, but note the point in the article: it is not adding anything to the GDP in terms of productivity. I’ve seen people use LLMs as slightly better interfaces to documentation. Yes, it’s useful, but it isn’t a killer app. People use the, for summarisation but they’re really bad at it and they often miss the key point. Using that in decision making lets you move faster, but in the wrong direction. It takes a while for that to appear and then it’s a problem. People are ‘vibe coding’ business apps, but the ones that work are the kind that you could build in less time with MS PowerApps and no programming ability beyond Excel formulae. Producing moderately bad subtitles is better than having no subtitles. Low-stakes translation was already pretty good pre-LLM and is now better (but still makes ludicrous mistakes and so can’t be trusted in situations where being misunderstood costs money or lives).
And all of these use cases are massively subsidised. How many people are willing to pay ten times OpenAI’s current price for LLMs? Inference costs dropping relies on the ML-tailored GPUs having large economies of scale, which relies on large numbers of consumers. Increase the price and that cycle ends. Training costs are increasing and must be amortised over more users, increase the costs and this stops, so you must increase the costs more.
I’ve seen estimates that, without VC money being set on fire, a ChatGPT subscription would cost at least $1,000/month. For a thing that doubles employee productivity, that’s an easy sell. For a modest increase, it is an easy ‘no’.
And LLMs become obsolete fast. A coding assistant that doesn’t know about new language features and which always uses deprecated or removed APIs is a productivity drain, not a help. A search assistant that has no knowledge of current events past 2025 will rapidly become useless. For translation, this is slower, but language evolves, both in terms of new words and new meanings for words, so these need retraining. For some use cases, RAG helps patch over the limitations of a stale model, but it also drives up the inference costs.
And all of that is assuming that the large-scale plagiarism for training these models is legal. At the moment governments are believing the ‘this will add massive amounts to your GDP’ hype from big tech and so are willing to throw creative industries under a bus in aid of better AI models. When they realise that the hype is built on lies, this will change.
"People use the, for summarisation but they’re really bad at it and they often miss the key point" not our experience at Discourse with extensive testing, so I question your whole statement on first principles. You have built a strawman argument with zero supporting data.
@codinghorror There are benchmarks in research papers evaluating this, but let’s assume that you are correct and, in your use case, it works well. If OpenAI charged 10x their current amount, would it still be worthwhile? Are you seeing enough revenue from customers that paying that amount would be worth it? If other people decide it isn’t and you’re now paying a larger fraction and the cost is 15x, is the same true? If so, great, you have a use case for LLMs that is sustainable after the bubble bursts.
@codinghorror Substitute someone else for OpenAI if you want. The compute requirements for teaming an LLM are huge. So huge that the economies of scale mean that whoever has the most paying customers can charge the least (once they’ve finished burning investor money). That’s a textbook example of a natural monopoly.
@codinghorror Gemini is produced by a three-trillion-dollar company that is ‘all in’ on AI and is burning their cash reserves (and money raised by pumping their stock price as part of the bubble and issuing more stock) to fund development.
There are a handful of big tech companies doing the same. There are a handful of companies with tens of billions of investor money doing the same.
I’d strongly dispute ‘tons’. Can you list even ten companies producing LLMs that are in a similar class to OpenAI? Can you list one that is reporting an operating profit from their LLM offering?
In our testing, if we start by asking the LLM to summarise, what comes back looks pretty good. But if we start by doing our own summary, then ask the LLM, then compare, we find the LLM summary usually misses what we believe to be important. Our conclusion was that the LLM doesn’t summarise with purpose. It has no concept of what is important to us, so it ends up producing, well, generic slop. @codinghorror @nixCraft @david_chisnall
@BenAveling @codinghorror Summarisation is a staggeringly hard problem. English schools used to tech précis, because it’s a skill that’s difficult. Identifying which parts of a document are the key ideas and then expressing those in fewer words is something requires understanding not just the text of the document but also the purpose of the summary. Summarising a technical document for someone making a business decision requires removing most of the technical detail and providing the information that leads to a cost-benefit analysis, whereas summarising the same document for someone evaluating experimental rigour requires pulling out the methodology and discarding much of the rest.
There are a bunch of classical NLP approaches to summarisation that rely on part-of-speech tagging and removing words (and sometimes entire sentences) that have a low probability of contributing to the end result. These often work well, but they have awful failure modes. For example, they will usually strip superlatives or modifiers on adjectives, but sometimes these are the critical information in the sentence.
LLMs are not summarising, because they do not have any of this context. They are doing a text-to-text transform to make the text shorter, in a way that mirrors what other summaries look like in their training data. If you have formulaic boilerplate, LLMs are great at removing that. For a lot of corporate documents, this kind of structure repeats and so simply removing it is easy as a statistical transform, though this doesn’t usually save much time because the boilerplate usually exists to make it easy for people to learn to navigate a class of documents quickly and find the points that are relevant for their particular use. If you have text that was expanded from a handful of bullet points by another LLM, they are pretty good at reversing that transform.
@david_chisnall @BenAveling David, I cannot emphasize enough how NONE of my (extensive) personal experience using LLMs to summarize at Discourse, and for my own summarizing use (extensive) completely contradicts your statements here. Let's just agree to disagree, because I have absolutely no clue what data you are looking at that says something COMPLETELY THE OPPOSITE of mine. It's basically "my data can beat up your data" and, well..
When I ask an LLM to create an abstract for a grant proposal I have just written, it's laughably off the mark, highliting irrelevant points.
When I ask an LLM to summarize a scientific manuscript, the result is concerning in that some of the points aren't made in the manuscript at all.
When I ask an LLM to pick key points from a long document, be it a book or a white paper, the choices are its own, not the ones I would have chosen. Resisting its choices is hard, setting up a concerning anchoring effect that I'd rather avoid.
Last I tried was this last August.
For all the above I don't use LLMs. For my domain of knowledge LLMs fall very short.
@albertcardona @david_chisnall @BenAveling it can definitely depend on the domain, also, I've seen a LOT of REALLY poorly written academic documents ;)
@codinghorror @david_chisnall @BenAveling Asking an LLM to fix a poorly written scientific manuscript is also a failure: makes the language smoother, even grandiloquent, but the content isn't clearer, primarily because it doesn't know – it can't know – what's missing. Otherwise the work wouldn't be at the horizon of knowledge, i.e., out of distribution to use the CS lingo.
@albertcardona @david_chisnall @BenAveling I would only ask it to simplify, as broadly speaking most writers use way way too many words because they think they need to.
Indeed and that highlights another major point: LLMs aren’t generic know-it-all tools, rarher, like all tools, there’s a learning process towards their proper, sensible use. Including, and particularly, being aware of its limitations as a tool and the domains were they apply at all.
Major issue for me is that unless one is an expert in the knowledge domain of the content of the text being generated, there’s no way to assess it for correctness or completeness, yet the neat eloquent language misleads.
These limitations go against current business models of growth and more growth so it isn’t surprising they aren’t more known.
@codinghorror @albertcardona @david_chisnall @BenAveling Having taught physics to different levels and ages, may I suggest the following. When you see a mess from GenAI, its like that which physics teachers see from students. Same depth or lack of grasp in meaning with full "I know it" aura. Student's goal is to get grade. GenAI is possessed by prompt, and just completes it. As a human, it is a humbling sight.
@codinghorror @david_chisnall Summarization is a well defined task and it covers multiple layers depending on the source material to sump up. Strong value on conversation (or QA discourse) does not mean it will be good on other content. LLMs been proven (solid evaluation protocols) not to be great to summarize complex written documents. It might appear I'm saying you ate both right... which I am. This confirms LLMs are not the generic problem solvers providers want us to believe.
@codinghorror@infosec.exchang @david_chisnall @nixCraft I share that view. Especially now since we are barely starting to know what are the best uses for these new set of tools.
Test it, try it, proceed with caution.
(Also don't ask me to review the code of your coding assistant, that's your job)
@codinghorror @gdupont @david_chisnall I trust that "living it" is a controlled experiment in which operators' biases and survivors effects are taken into account. I'm only explicitly mentioning it because these are common in evaluations that don't follow a carefully designed experimental protocol. So in that case one is left with a *personal opinion* which could be heavily biased by a number of phenomena.
@joe_vinegar @gdupont @david_chisnall ultimately the market will decide. Is it worth it? What percent productivity boost? And it will depend on the domain / content.
> ultimately the market will decide There is a strong bias here, given the incentive to make everyone use it. It might take time for the dust to settle to realistic usage (ie not marketing driven).
@gdupont @joe_vinegar @david_chisnall ironically I think programming and coding will end up being one of the worst, least reliable uses. It's the loosey-goosey human language stuff it will fare better at (legal docs, generic PDFs, happy marketing talk, etc)
@codinghorror @gdupont @joe_vinegar @david_chisnall May I suggest, while gravy flows, and some of you may have it, or may influence distribution, please direct it to cimc.ai Those guys have a point. 🙏
One of the use cases I thought was reasonable to expect from ChatGPT and Friends (LLMs) was summarising. It turns out I was wrong. What ChatGPT isn’t summarising at all, it only looks like it…
@david_chisnall @codinghorror are you sure about the numbers? I remember reading that an average user cost Altman 2.5x the subscription cost. I get that the cost are rising fast, but I have not seen the 10x.
I have been trying to decrease my exposure to tech this whole year, and it's actually *hard* to do now, because a lot of tech companies are officially classified as something else. An "ex-tech" index fund still has Meta and Google in it because those are considered advertising companies, not "tech". A "social" fund has, I think, all seven of the profoundly antisocial "Magnificent 7".
@wollman the classification of Meta and Google as advertisement companies is on point, because that's where their main profit comes from. It is still hard for me to imagine how the bubble burst would impair the advertisement activities (to which profiling is instrumental). I foresee no reduction of usage of non-AI surveillance tools, quite the contrary. And the cost of labor will be cheaper in a crisis.
I've never seen "parabolic" used in economic discussions before — how interesting! I know what it means, though in computer science we'd probably say "polynomial". But "parabolic" is more specific...
And, per the article, it's not even the tech keeping the bubble afloat, just the construction of data centers. Presumably, we could get the same economy-floating effect by shifting investment to, say, affordable housing initiatives, which would have the added result of keeping more people sheltered when the downturn comes.
I feel so stressed out now I can feel it building in waves physically inside me all the time now. I was never a fan of Bush 2.0, and I don’t want him back either, but I never felt like this in the 2000s.
Sixth-century mosaic of a bird from the floor of basilica in Thessaloniki, #Greece. This is one of several birds in the remaining fragments. Another fragment features a deer. Both are housed in the Museum of Byzantine Culture. 📷🇬🇷 www.flickr.com/photos/60661...#photography#Greece#archaeology
Sixth-century mosaic floor found in a basilica north of the Church of Archangels in Thessaloniki, Greece. This is one of several birds in the remaining fragments.Another fragment features a deer. Currently housed in the Museum of Byzantine Culture.
#droni "E' in corso un'escalation e la #Russia ne ha la piena responsabilità", ha dichiarato l'ammiraglio Giuseppe Cavo Dragone, presidente del comitato militare della #Nato.
“The disenfranchised men of the manosphere disdain women, and yet women continue to be asked to feel pity and concern for them.” lrb.co.uk/the-paper/v47/n16/em…
A imagem é um diagrama ilustrativo que explica a ordem lógica de execução de uma consulta SQL. No topo, há duas tabelas, t1 e t2, representadas por quadrados coloridos. Uma seta azul conecta t1 e t2, indicando um JOIN. A consulta começa com a cláusula SELECT, que especifica as colunas a serem recuperadas (column_a, column_b). Em seguida, a cláusula FROM indica a tabela principal (t1), seguida pela cláusula JOIN que une t1 e t2 com base em uma condição (t1.column_a = t2.column_a). A cláusula WHERE filtra os resultados com uma expressão de restrição. Em seguida, a cláusula GROUP BY agrupa os resultados por uma coluna específica. A cláusula HAVING filtra os grupos com uma expressão de restrição. A cláusula ORDER BY organiza os resultados por uma coluna em ordem ascendente ou descendente. Por fim, a cláusula LIMIT limita o número de linhas retornadas. Cada cláusula é conectada a uma tabela correspondente por uma seta azul, ilustrando a ordem lógica de execução da consulta.
Fornecido por @altbot, gerado localmente e de forma privada usando Ovis2-8B
Profitez des vidéos et de la musique que vous aimez, mettez en ligne des contenus originaux, et partagez-les avec vos amis, vos proches et le monde entier.
Condiviso dall’utente Emilia su un gruppo telegram (DataKnightmare fan club) grazie per aver condiviso.
Thiagarajan, T. C., Newson, J. J., & Swaminathan, S. (2025). Protecting the Developing Mind in a Digital Age: A Global Policy Imperative. Journal of Human Development and Capabilities, 26(3), 493–504. doi.org/10.1080/19452829.2025.…
L’episodio, durato alcune ore, è stato confermato oggi dalla polizia locale. «Non possiamo rilasciare informazioni sulla provenienza dei droni», ha dichiarato l’ufficiale di servizio Simon Skelsjaer, precisando che i velivoli non sono stati abbattuti.
L’ufficiale di servizio Simon Skelsjaer ha precisato che i velivoli non sono stati abbattuti. Non sono state rilasciate informazioni sulla provenienza dei droni
La mayor prueba de que los empresarios no tienen ni puta idea de lo que es trabajar, es que piensan que pueden sustituir a sus trabajadores por una IA.
Ich glaube, dieses Foto hatte ich hier schon einmal geteilt, es zeigt eine gelbe Blüte eines Klees, leider hatte ich letztens meinen vorherigen Account gelöscht, sorry.. #fotografie #photography #natur #nature #pflanzen #plants
This week's LaRouche show features Lyndon LaRouche’s remarks to a June 18, 1997 policy forum titled “Never Again! Stop British-Backed Genocide in Africa.” Fr...
Will Sexton, of Duke University Libraries, shared some very astute characterizations of David Heinemeier Hansson when announcing their decision to stop using Basecamp.
I also appreciated the thoughtful digression on the history of systems of oppression, and reflecting on the harms caused as it relates to Duke University. Well worth reading.
Profitez des vidéos et de la musique que vous aimez, mettez en ligne des contenus originaux, et partagez-les avec vos amis, vos proches et le monde entier.
Profitez des vidéos et de la musique que vous aimez, mettez en ligne des contenus originaux, et partagez-les avec vos amis, vos proches et le monde entier.
"In some cases, this reporting requirement poses real safety risks," Dara Lindenbaum writes.
"I worry about that — especially because my agency, the Federal Election Commission, oversees the requirement that gives rise to these risks." wapo.st/4mBDyum
Serie A,5° turno: Juventus-Atalanta 1-1 20.00 Divisione della posta allo 'Stadium': tra Juventus è Atalanta è 1-1. Bianconeri a 11 punti, orobici a 9. Subito palo della Juve con la zuccata di Kalulu.Adzic chiama Carnesecchi alla parata, a ruota tiro di Thuram deviato da Pasalic sfiora il montante. Adzic da fuori,Carnesecchi blocca.Il montenegrino perde palla, Sulemana si invola e fa secco Di Gregorio (46'pt).Ripresa. Bremer buca,Krstovic spreca.Pasalic di testa impegna Di Gregorio.Pari bianconero di Cabal appena entrato su liscio di Kossounou (78'). Rosso a De Roon (80'). #televideo #ultimaora
Profitez des vidéos et de la musique que vous aimez, mettez en ligne des contenus originaux, et partagez-les avec vos amis, vos proches et le monde entier.
Aaron Longchamps
in reply to nixCraft 🐧 • • •Em reshared this.
James H McLaren
in reply to nixCraft 🐧 • • •Lesley Carhart reshared this.
Jeff Atwood
in reply to James H McLaren • • •@JamesHMcLaren that's just completely absolutely untrue, I'm sorry, but.. infosec.exchange/@codinghorror…
Jeff Atwood
2025-09-28 03:06:32
Mike McCaffrey
in reply to Jeff Atwood • • •Jeff Atwood
in reply to Mike McCaffrey • • •George Mitchell
in reply to Jeff Atwood • • •@codinghorror @mikemccaffrey @JamesHMcLaren It’s the overhyping that’s the problem; when the hype bubble bursts, it’s not just going to take down the no-value “solutions”, it’s also going to crater the actually-valuable propositions because those products aren’t profitable without the hype money coming in.
(My wife was a pre-IPO employee of Amazon; I watched the dot-com collapse in extreme close-up.)
doragasu
in reply to nixCraft 🐧 • • •Jeremy Kahn
in reply to doragasu • • •@doragasu
It's not the data centers. It's the overvaluation of "AI" stocks that pushes bubble money into pension funds
When that bubble valuation goes away, the idled data centers will be there like the feet of Ozymandias, and the pension funds and 401(k)s will be the barren desert all around
@nixCraft
@cstross
doragasu
in reply to Jeremy Kahn • • •lbcp 🦖
in reply to doragasu • • •Charlie Stross
in reply to lbcp 🦖 • • •The Incredible Laser
in reply to nixCraft 🐧 • • •Knightmare
in reply to nixCraft 🐧 • • •Mariner
in reply to nixCraft 🐧 • • •Andreas K
in reply to nixCraft 🐧 • • •Nazo
in reply to nixCraft 🐧 • • •JP
in reply to nixCraft 🐧 • • •Galad
in reply to nixCraft 🐧 • • •xs4me2
in reply to nixCraft 🐧 • • •Karsten Johansson
in reply to nixCraft 🐧 • • •Isn't that another way of saying the US will do anything to keep propping it up, no matter the cost?
This is a race that is China's for the taking. USA will let that happen?
Huge sacrifices are about to be made to keep up the illusion as long as possible.
Kierkegaanks regretfully
in reply to nixCraft 🐧 • • •Astrius
in reply to nixCraft 🐧 • • •Alan Hicks
in reply to nixCraft 🐧 • • •Al & Val's Modern Homesteading
in reply to nixCraft 🐧 • • •Scotty Trees
in reply to nixCraft 🐧 • • •Strefa Linux
in reply to nixCraft 🐧 • • •acm
in reply to nixCraft 🐧 • • •Alex Lohr (RIP Natenom)
in reply to nixCraft 🐧 • • •It is only AI in PowerPoint. In Python, it's large models. In truth, it is creative marketing around statistical models tuned to predict the answer.
And the answer is: don't waste so many resources on predicting answers using models that human brains are far better at giving.
The attempt to replace human brains was doomed to begin with for any task that required one in the first place.
SpaceLifeForm
in reply to nixCraft 🐧 • • •When Deutsche Bank speaks, people listen.
They know it is a clusterfuck. Now, they are wondering what are going to get in return for their investments.
Maybe they are thinking to pop the bubble before they lose more money.
#AI #Insanity #Money
RRB
in reply to nixCraft 🐧 • • •contrasocial
in reply to nixCraft 🐧 • • •deedlebaum
in reply to nixCraft 🐧 • • •samiamsam
in reply to nixCraft 🐧 • • •Lesley Carhart
in reply to nixCraft 🐧 • • •Jeff Atwood
in reply to nixCraft 🐧 • • •pcyx
in reply to Jeff Atwood • • •David Chisnall (*Now with 50% more sarcasm!*)
in reply to Jeff Atwood • • •@codinghorror
Some of it is useful, but note the point in the article: it is not adding anything to the GDP in terms of productivity. I’ve seen people use LLMs as slightly better interfaces to documentation. Yes, it’s useful, but it isn’t a killer app. People use the, for summarisation but they’re really bad at it and they often miss the key point. Using that in decision making lets you move faster, but in the wrong direction. It takes a while for that to appear and then it’s a problem. People are ‘vibe coding’ business apps, but the ones that work are the kind that you could build in less time with MS PowerApps and no programming ability beyond Excel formulae. Producing moderately bad subtitles is better than having no subtitles. Low-stakes translation was already pretty good pre-LLM and is now better (but still makes ludicrous mistakes and so can’t be trusted in situations where being misunderstood costs money or lives).
And all of these use cases are massively subsidised. How many people are willing to pay ten times OpenAI’s current price for LLMs? Inference costs dropping relies on the ML-tailored GPUs having large economies of scale, which relies on large numbers of consumers. Increase the price and that cycle ends. Training costs are increasing and must be amortised over more users, increase the costs and this stops, so you must increase the costs more.
I’ve seen estimates that, without VC money being set on fire, a ChatGPT subscription would cost at least $1,000/month. For a thing that doubles employee productivity, that’s an easy sell. For a modest increase, it is an easy ‘no’.
And LLMs become obsolete fast. A coding assistant that doesn’t know about new language features and which always uses deprecated or removed APIs is a productivity drain, not a help. A search assistant that has no knowledge of current events past 2025 will rapidly become useless. For translation, this is slower, but language evolves, both in terms of new words and new meanings for words, so these need retraining. For some use cases, RAG helps patch over the limitations of a stale model, but it also drives up the inference costs.
And all of that is assuming that the large-scale plagiarism for training these models is legal. At the moment governments are believing the ‘this will add massive amounts to your GDP’ hype from big tech and so are willing to throw creative industries under a bus in aid of better AI models. When they realise that the hype is built on lies, this will change.
Jeff Atwood
in reply to David Chisnall (*Now with 50% more sarcasm!*) • • •David Chisnall (*Now with 50% more sarcasm!*)
in reply to Jeff Atwood • • •There are benchmarks in research papers evaluating this, but let’s assume that you are correct and, in your use case, it works well. If OpenAI charged 10x their current amount, would it still be worthwhile? Are you seeing enough revenue from customers that paying that amount would be worth it? If other people decide it isn’t and you’re now paying a larger fraction and the cost is 15x, is the same true? If so, great, you have a use case for LLMs that is sustainable after the bubble bursts.
Jeff Atwood
in reply to David Chisnall (*Now with 50% more sarcasm!*) • • •David Chisnall (*Now with 50% more sarcasm!*)
in reply to Jeff Atwood • • •Substitute someone else for OpenAI if you want. The compute requirements for teaming an LLM are huge. So huge that the economies of scale mean that whoever has the most paying customers can charge the least (once they’ve finished burning investor money). That’s a textbook example of a natural monopoly.
Jeff Atwood
in reply to David Chisnall (*Now with 50% more sarcasm!*) • • •mms
in reply to Jeff Atwood • • •David Chisnall (*Now with 50% more sarcasm!*)
in reply to Jeff Atwood • • •@codinghorror
Gemini is produced by a three-trillion-dollar company that is ‘all in’ on AI and is burning their cash reserves (and money raised by pumping their stock price as part of the bubble and issuing more stock) to fund development.
There are a handful of big tech companies doing the same. There are a handful of companies with tens of billions of investor money doing the same.
I’d strongly dispute ‘tons’. Can you list even ten companies producing LLMs that are in a similar class to OpenAI? Can you list one that is reporting an operating profit from their LLM offering?
Ben Aveling
in reply to Jeff Atwood • • •Joe Vinegar reshared this.
David Chisnall (*Now with 50% more sarcasm!*)
in reply to Ben Aveling • • •@BenAveling @codinghorror
Summarisation is a staggeringly hard problem. English schools used to tech précis, because it’s a skill that’s difficult. Identifying which parts of a document are the key ideas and then expressing those in fewer words is something requires understanding not just the text of the document but also the purpose of the summary. Summarising a technical document for someone making a business decision requires removing most of the technical detail and providing the information that leads to a cost-benefit analysis, whereas summarising the same document for someone evaluating experimental rigour requires pulling out the methodology and discarding much of the rest.
There are a bunch of classical NLP approaches to summarisation that rely on part-of-speech tagging and removing words (and sometimes entire sentences) that have a low probability of contributing to the end result. These often work well, but they have awful failure modes. For example, they will usually strip superlatives or modifiers on adjectives, but sometimes these are the critical information in the sentence.
LLMs are not summarising, because they do not have any of this context. They are doing a text-to-text transform to make the text shorter, in a way that mirrors what other summaries look like in their training data. If you have formulaic boilerplate, LLMs are great at removing that. For a lot of corporate documents, this kind of structure repeats and so simply removing it is easy as a statistical transform, though this doesn’t usually save much time because the boilerplate usually exists to make it easy for people to learn to navigate a class of documents quickly and find the points that are relevant for their particular use. If you have text that was expanded from a handful of bullet points by another LLM, they are pretty good at reversing that transform.
Jeff Atwood
in reply to David Chisnall (*Now with 50% more sarcasm!*) • • •Albert Cardona
in reply to Jeff Atwood • • •@codinghorror @david_chisnall @BenAveling
A few anecdotal cases:
When I ask an LLM to create an abstract for a grant proposal I have just written, it's laughably off the mark, highliting irrelevant points.
When I ask an LLM to summarize a scientific manuscript, the result is concerning in that some of the points aren't made in the manuscript at all.
When I ask an LLM to pick key points from a long document, be it a book or a white paper, the choices are its own, not the ones I would have chosen. Resisting its choices is hard, setting up a concerning anchoring effect that I'd rather avoid.
Last I tried was this last August.
For all the above I don't use LLMs. For my domain of knowledge LLMs fall very short.
CDCastillo reshared this.
Jeff Atwood
in reply to Albert Cardona • • •Albert Cardona
in reply to Jeff Atwood • • •Asking an LLM to fix a poorly written scientific manuscript is also a failure: makes the language smoother, even grandiloquent, but the content isn't clearer, primarily because it doesn't know – it can't know – what's missing. Otherwise the work wouldn't be at the horizon of knowledge, i.e., out of distribution to use the CS lingo.
CDCastillo reshared this.
Jeff Atwood
in reply to Albert Cardona • • •Albert Cardona
in reply to Jeff Atwood • • •Indeed and that highlights another major point: LLMs aren’t generic know-it-all tools, rarher, like all tools, there’s a learning process towards their proper, sensible use. Including, and particularly, being aware of its limitations as a tool and the domains were they apply at all.
Major issue for me is that unless one is an expert in the knowledge domain of the content of the text being generated, there’s no way to assess it for correctness or completeness, yet the neat eloquent language misleads.
These limitations go against current business models of growth and more growth so it isn’t surprising they aren’t more known.
Ben Aveling
in reply to Jeff Atwood • • •Mikalai
in reply to Jeff Atwood • • •Having taught physics to different levels and ages, may I suggest the following.
When you see a mess from GenAI, its like that which physics teachers see from students. Same depth or lack of grasp in meaning with full "I know it" aura.
Student's goal is to get grade.
GenAI is possessed by prompt, and just completes it.
As a human, it is a humbling sight.
ggdupont
in reply to Jeff Atwood • • •@david_chisnall Summarization is a well defined task and it covers multiple layers depending on the source material to sump up. Strong value on conversation (or QA discourse) does not mean it will be good on other content. LLMs been proven (solid evaluation protocols) not to be great to summarize complex written documents.
It might appear I'm saying you ate both right... which I am. This confirms LLMs are not the generic problem solvers providers want us to believe.
Jeff Atwood
in reply to ggdupont • • •ggdupont
in reply to Jeff Atwood • • •@codinghorror@infosec.exchang
@david_chisnall @nixCraft
I share that view. Especially now since we are barely starting to know what are the best uses for these new set of tools.
Test it, try it, proceed with caution.
(Also don't ask me to review the code of your coding assistant, that's your job)
Joe Vinegar
in reply to Jeff Atwood • • •Jeff Atwood
in reply to Joe Vinegar • • •ggdupont
in reply to Jeff Atwood • • •@codinghorror @joe_vinegar @david_chisnall
> And it will depend on the domain / content.
1000%
> ultimately the market will decide
There is a strong bias here, given the incentive to make everyone use it. It might take time for the dust to settle to realistic usage (ie not marketing driven).
Joe Vinegar reshared this.
Jeff Atwood
in reply to ggdupont • • •ggdupont
in reply to Jeff Atwood • • •I wouldn't place any bet.
time will tell...
Mikalai
in reply to Jeff Atwood • • •May I suggest, while gravy flows, and some of you may have it, or may influence distribution, please direct it to cimc.ai
Those guys have a point.
🙏
mkroehnert
in reply to Jeff Atwood • • •@codinghorror
@david_chisnall @nixCraft
This is a good text going over the summarizing argument.
ea.rna.nl/2024/05/27/when-chat…
When ChatGPT summarises, it actually does nothing of the kind.
R&A IT Strategy & Architecturemms
in reply to David Chisnall (*Now with 50% more sarcasm!*) • • •Jeff Atwood
in reply to mms • • •Garrett Wollman
in reply to nixCraft 🐧 • • •Joe Vinegar
in reply to Garrett Wollman • • •Tofu Golem
in reply to nixCraft 🐧 • • •Bob Thomson
in reply to nixCraft 🐧 • • •a7ndrew
in reply to nixCraft 🐧 • • •Toni M.
in reply to nixCraft 🐧 • • •Mike Torr
in reply to nixCraft 🐧 • • •I can't say I'm surprised.
I've never seen "parabolic" used in economic discussions before — how interesting! I know what it means, though in computer science we'd probably say "polynomial". But "parabolic" is more specific...
fhekland
in reply to nixCraft 🐧 • • •⁂ L. Rhodes
in reply to nixCraft 🐧 • • •