@lpryszcz and the internet of things, and virtual reality, and the metaverse, etc... bubbles upon bubbles, this one being shockingly big
Discussion
Loading...
Discussion
@gabrielesvelto we regularly have to push back on this at work. Why spend a fortune on an AI agent when a simple regex will do.?
@gabrielesvelto In areas not entirely technological many solutions are also already well known and haven't been realized because monneeh. Politicians just rather throw an app at people with e.g. mental health issues than finance more actual therapy opportunities, or hope an AI fixes the climate crisis instead of changing the industries causing it.
Now with AI and LLMs around, the people with the power for meaningful change can just point at that and say "Look, we did something! Now shut up!" Neither the solutions we already have are being used, nor are people looking for new ones, as you say.
@gabrielesvelto THIS! Whenever anyone proposes solving a probem with AI I try to present it to them like this:
You have a choice between spending a few days now to hash out a procedure that eliminates this problem. It's a one-time cost.
OR you can try to use AI. You'll have a ongoing cost forever, which could increase with higher needs, you'll have to send your data to an untrusted third party, it's not guaranteed to work reliably and it exposes you to cybersecurity threats.
@jmaris @gabrielesvelto so from the perspective of the system being built, implementing it properly is CAPEX, implementing it with AI is OPEX with a supplier who can jack up the prices when they want.
@gabrielesvelto many of the engineers building AI are not computer scientists. Does this matter? Yes, even more than the fact that they know next to nothing about cognitive science.
@gabrielesvelto
But... But... But.... If I don't use "AI" how will I ever demolish the environment while churning out mountains of garbage based on stolen content and forgetting how to actually do anything?!?!?!
"I used AI to....", is nothing more than, "Listen I'm not an asshole but....", for the 21st Century.
@gabrielesvelto there's also a bunch of allegedly "AI" things which turn out, when you scratch them, to actually be simpler things like Bayesian stats or k-means clustering or just random guesses filtered by the user.
@nickzoic yeah, and in some cases they even make sense. One of the casualties of the LLM craze are all the other areas of artificial intelligence research.
@gabrielesvelto Quite often, I struggle to see why the thing they are calling "AI" isn't just normal 'computation'. Buzzwords must sell?
@gabrielesvelto ⬆️ ⬆️ ⬆️ ⬆️ ⬆️
Absolutely agree. In preparing some upcoming interview for a group grant, we realised we did not mention AI once in 50 pages...and now we are afraid to get even strong critical questions concerning why we neglect this all-mighty-all-solving-magic , as if we underlooked something fundamental (we did not, for our science, but we're already at the stage in which this might look "retro")
@gabrielesvelto ah, but the AI version makes more profit, allegedly.
@gabrielesvelto this sounds very similar to trying to solve everything with #blockchain a few years ago... #aiBubble
@lpryszcz and the internet of things, and virtual reality, and the metaverse, etc... bubbles upon bubbles, this one being shockingly big
@gabrielesvelto I've been wondering what the next bubble will be. Could it be #quantum #computing?
https://www.theguardian.com/technology/2025/sep/04/quantum-computing-firm-reaches-10bn-valuation-as-investor-interest-builds
https://www.theguardian.com/technology/2025/sep/04/quantum-computing-firm-reaches-10bn-valuation-as-investor-interest-builds
@lpryszcz maybe? I guess it depends on how much money venture capitalists are willing to splash on it once this particular bubble bursts. They might even rethink about burning more money in the tech sector and move elsewhere, which would be a net win for humanity most likely.
@gabrielesvelto Speaking of secondary inefficiencies, last night I encountered a vibe code call which pings a service to check if the supplied string is longer than the supplied max length.
This stuff is just useless.
@gabrielesvelto Someone suggested to me I can run an article through an LLM vor spell-checking.
That said, they ment grammar-checking. And even that is a more or less solved problem with word processors. Why had that to be an LLM?
@inw indeed. Plus a regular spell- and grammar-checker will use language rules for that instead of relying on a probabilistic system, so they're not perfect but they're less likely to make unpredictable mistakes. And they require a minuscule amount of resources in comparison.
@gabrielesvelto tax billionaires out of existence for a start and then, as a natural consequence, all the rent-seeking gatekeepers of those sweet, sweet NVIDIA-based data centres will also die on the vine.
Simples.
@gabrielesvelto Well this is not only related to AI, but a general trend. It's called the "Crapularity". Any task in IT takes more and more resources (both CPU and personal) to complete and gets less reliable. So eventually we're going to have huge development teams, developing trivial applications, that no longer work. We currently see that, for example, on the WWW.
@casandro that's true, but AI just took the regular inefficiency of the tech sector and cranked it up to 11
@gabrielesvelto oh yes, this precisely. the "AI" solution was aggressively marketed as *automatically* better than any other method of solving problems. Why it's self-evidently better! It's _learning_! It's superintelligent, how can it possibly be anything less than optimal?
Well...
@gabrielesvelto You mean… like reading books?
@gabrielesveltohttps://en.wikipedia.org/wiki/Fast_Fourier_transform is almost always the algorithmic alternative to use for specific sets of data. (Really, an LLM is just doing this same thing, at massive scale, on unrelated and unfiltered data, only to give you the wrong answers)
@gabrielesvelto Until someone comes up with an energy efficient LLM then this form of AI is completely bottlenecked. It’s utterly useless for new, nuanced, and novel tasks… IE, most situations. LLMs will never ever have enough data to replace workers, no matter how much the new aristocracy wishes it to. The most it can do is make existing workers a little more productive.
@gabrielesvelto I'm reminded of someone making a calculator that uses a LLM as a joke. There were processors that ran in the kiloherts (not megahertz!) range in the 1970s that could do basic calculations more accurately than a LLM, yet here we are where there actually are people who non-jokingly ask a LLM to do calculations...
There are some things a LLM can be good at, but everyone is focused on using them for the wrong things instead. Admittedly there isn't a lot they're good at, but it's kind of sad that between the con artists and people caught up in the hype, all the focus has been on using them as wrongly as possible instead of on those few good things.
I'm hoping when the bubble pops people don't kneejerk 100% away from the mechanisms but instead they become a niche thing.
@gabrielesvelto To me the most disheartening fact is that the stuff where LLMs are good at, a (big) if-then-else would probably suffice and requiring a fraction of the energy (but fair enough, more robotic).
The stuff where an if-then-else is not enough, the more novel approaches you need, the less effective LLMs start to become as they can only do probability calculations on existing knowledge.
In those cases, a rubber ducky probably produces the same result, but LLMs are simply better at bias confirmation, so more comforting and less confronting as critical thinking
The stuff where an if-then-else is not enough, the more novel approaches you need, the less effective LLMs start to become as they can only do probability calculations on existing knowledge.
In those cases, a rubber ducky probably produces the same result, but LLMs are simply better at bias confirmation, so more comforting and less confronting as critical thinking
@gabrielesvelto As someone who's actually designed new [efficient] algorithms, this hurts.
OTOH, it's always been that way. We just have a new most-obvious (and even more inefficient) way in people's heads.
@gabrielesvelto you mean, like using an index to search for things? Nah, that's crazy talk, this would never work! Trust us bro: using a bullshit machine in the hope it regurgitates something vaguely resembling what you're looking for from its training set is really the only way to solve that problem!
@gabrielesvelto
I 100% agree. When all you have is a hammer (LLM) then everything looks like a nail. I see future systems combining probabilistic and deterministic programs, making best use of their complimentary capabilities. And smaller more efficient generative models trained for specific domains.
I 100% agree. When all you have is a hammer (LLM) then everything looks like a nail. I see future systems combining probabilistic and deterministic programs, making best use of their complimentary capabilities. And smaller more efficient generative models trained for specific domains.
@mattjhayes narrowly trained small models work very well for specific use-cases, but we've known that since we started using neural networks
@gabrielesvelto Sure. I do wonder, with pretty much every org on the planet running an AI transformation programme (often as a solution looking for problems), it would be interesting to see data on usage of general purpose vs domain specific generative AI models. My hypothesis is that most enterprise AI implementations make inefficient use of GP models, augmented with prompting techniques such as RAG, as the barrier to model training is perceived as too high, or it is not even considered. And, as per your original post, a certain percentage could have been solved more effectively with non-AI solutions.
@mattjhayes yes, or even "different kind" of AI solutions. I'll give you two examples we've had success with within Mozilla: translations done using a very small scale local model (a few MiB in size). Not perfect, definitely lower quality than large models, but can be run on any hardware and trained in a few hours on a regular desktop machine.
@mattjhayes the other one automatic crash detection. We've got a bot using deep learning techniques that automatically sifts through our crash reporting data and files a bug when it finds something suspicious. Not perfect, has both false positives and false negatives, but it greatly reduced the need for manually triaging large amount of crash reports. Again it's a very simple solution that doesn't need an LLM in spite of using similar underlying technology.
@gabrielesvelto
Interesting use cases. I imagine there are a lot of architectural and engineering considerations when updating browser code. Everything has to be fast, stable, secure, privacy-preserving, etc. Does this create contention with product teams, who want to add new features?
Interesting use cases. I imagine there are a lot of architectural and engineering considerations when updating browser code. Everything has to be fast, stable, secure, privacy-preserving, etc. Does this create contention with product teams, who want to add new features?
1+ more replies (not shown)
@gabrielesvelto
The vast investment in AI was intended to derail innovative climate solutions.
AI has sucked billions of investment dollars away from real problems & real solutions.
The source of all that funding is a red flag.
Petrostate despots are using it as cover for an international state surveillance platform.
Putin, OPEC & Koch Network intend to keep captive oil consumers or else.
https://www.cnbc.com/2025/08/27/saudi-arabia-wants-to-be-worlds-third-largest-ai-provider-humain.html
@Npars01 yes, that's very true. Without the LLM craze CO² emissions would have already peaked a couple of years back. The bubble is keeping the fossil fuel industry on life support, and the fossil fuel industry is happy to oblige
@gabrielesvelto I did product management for an AI solution, and it was so boring because the answer to every problem was "Let's ask an LLM!"
Now I'm back doing product management on exceedingly boring business software and it's so much more interesting because every change has to be carefully architected.
@gabrielesvelto Not only algorithmic solutions. The other “use cases” of LLMs also: People use them because search engines did not give them good answers, good (re)sources for their questions (either anymore or never did), because they did not find good tutorials and boilerplates for their programming questions, because they were drowned in low-quality, ad-infested click-bait instead of getting well-presented articles with expert knowledge and links to dig further.
It would be, would have been so much more valuable to develop further in those directions, make good original content more omnipresent, better discoverable, … instead of letting “AI” do the discovery and synopsis in a very suboptimal way.
@gabrielesvelto I recently saw a pitch from an AI HR company showing off that AI could easily show the boss that people who left the company made significantly less money than those who stayed, and do they really not have a spreadsheet made 30 years ago that tells them that? That’s a 1990’s spreadsheet problem.
@gabrielesvelto I'm partly involved in our company's marketing activities and happily label everything "AI" regardless of how it was implemented. Fool's gold.
@gabrielesvelto I keep asking the question, "what does this (LLM built idea) do better than (existing way of solving thing)" and no one knows how to answer
@gabrielesvelto Theory: CS fundamental research has always declined at a rate, inversely proportional to IT profits.
And the reason why I'm absolutely certain of this is the volume of investments sunk into artificial intelligence. We're talking about 500+ billion $ in 2024 alone. That kind of money finds an alternative solution to *every* problem. You don't even need to think about the technical aspects, with enough funding we can make pigs fly.
@gabrielesvelto
Now do return on investment into internet.
For shits and giggles.
@gabrielesvelto
When you give money to people, it can be a little unclear what they'll do with it.
Hardware & Energy, you can take a good guess what stock to buy, so you can buy in all the way along the supply chain before the loans are issued.
Is there a better algorithm than an LLM to generate code? I bet there is. Is there a better algorithm to do fuzzy search? You bet. Pattern matching? Machine vision? Even natural language processing. I'm absolutely sure there are better algorithms and 500 billion $ are more than enough money to find them.
FYI this bit of news is very relevant to my post from yesterday:
CoreWeave is one of the companies at the center of the AI bubble, and insiders are dumping their stock options as fast as they possibly can. Nothing tells you "this industry is toast" more than insiders dumping shares on the market.
@gabrielesvelto
All true!
There are better SPECIFIC algorithms.
LLM may be shitty and inefficient...but nearly universal.
A sonic screwdriver of the toolbox.
@n_dimension @gabrielesvelto ah, this must be why you see carpenters, electricians, dentists, painters, mechanics, deep sea divers, astronauts all using a leatherman multitool as their first tool of choice.
@gabrielesvelto i'm very interested in fuzzy search in particular and it's one of the subjects i would be working on for a phd
@gabrielesvelto ended up not getting accepted for a phd thanks to The Political Situation
@hipsterelectron sorry to hear that. Yeah, fuzzy search is interesting and the computing power of modern machines make it very appealing. ripgrep can tear through data at 10s of GiB per second on my box. Could that be used for more structured search across different file types? I'm sure it's possible, and it would surely beat sending stuff back and forth to a gigantic data center on the other side of the world.
@gabrielesvelto my emacsconf presentation last year https://emacsconf.org/2024/talks/regex/ mentions my research to develop a database structure that can execute NFAs directly against the index without touching the source. ripgrep is actually remarkably efficient against large monorepos like we had at twitter because of SIMD prefilter optimizations but finite automata are actually insufficient on a theoretical level to enable using SIMD after the automaton is invoked (so
"a.*b" can't use a SIMD operation to match b). text search is also completely unable to parallelize across threads, and my model addresses this and more
2 more replies (not shown)
@gabrielesvelto The question is ill posed because an LLM doesn't generate code. It copies code the model was overfitted to and obfuscates the source to avoid getting sued.
Can an algorithm generate code? In some cases yes, but when that can be done reliably in a way you can prove is reliable, it's indicative that you're writing code at the wrong level. This is the problem high level languages have been addressing for nearly half a century.
@gabrielesvelto
there is a dark/ humorous corollary to the problem you've identified:
everything getting labelled #AI
even things not AI
when the decision makers are in the throes of mania, to get approval or funding: it just gets called AI, whatever it is, and green lights
call it malicious compliance to mass hysteria without grounding in technical acumen
"we need a new door"
"don't have the funding"
"the door is AI: it senses when someone approaches, and slides open"
"approved!"
@gabrielesvelto yes, but... Llm can do all that _and_ link one with the other. All that being accessible to anyone, which is really the big new thing.
Sometimes I wonder if in the resistance to AI (and I include myself in it so don't take it badly) there could be some computer science aristocracy corporatism.
@julienw @gabrielesvelto dude, wait til someone wants a return on their investments. This stuff is only free right now to capture the market. It will not stay free. Have you seen what training costs?
@julienw @gabrielesvelto accessible is going to be very relative when it costs 200 bucks per month.
@julienw @gabrielesvelto It’s not accessible to anyone, though. When you don’t know what questions to ask and what to do with the answer, it wouldn’t help even if you had access to the most reliable question/answer machine. And these machines are not even reliable.
@gabrielesvelto computers are great at generating code in deterministic ways. After all, that's basically what a compiler is! Leapfrogging over DSLs and metaprogramming for a shitty jackpot.
It also drives me nuts because I've always advocated for the end-user and for adding the cloud (and thus making the app unusable if the company folds) only as a last resort. Now it's so much harder to do this. The AI bubble is user-hostile.
bonfire.cafe
A space for Bonfire maintainers and contributors to communicate
Automatic federation enabled