I suspect that even of that 1/5 that are getting "good" results, they are unlikely objectively good. So many are impressed by LLMs eloquently crafted wording that they completely miss the nonsense it is spewing.
The only people that might be getting really good results are those who have jobs making spam content all over the web. It is decently good at that.
Glad you asked. Absolutely! The whole idea is preposterous. Only someone with no interest in, respect for, or knowledge of biology, evolution, and how intelligence actually works could have come up with such an idea. That so many, who ought to know better, seem to be smitten, if not bitten, by the same bug, is shocking.
Just the basic premise that approximations of incomplete information, extracted by inaccurate methods, using a sliver of sensory data, producing irrational, inaccurate, and downright bogus results, ought to be sufficient warning to wake anybody up, but instead, the poor methodology, along with the frequent, but constant, mismatch of results to reality, is just brushed aside, in the belief that it will work next time. ("Then a miracle happens.")
Humans have a long history of magical thinking, for example, religions, and alchemy, are two of many based on such thinking. To the true believer, unfortunately, there is no real-world corrective.
I do wonder what the is the aftermath of the reseaech when the genai bubble bursts. and what will the new hyped reseaech direction be and how long will something with a comparable glamour to llms take to emerge.
The new hype appears to be "neurosumbolic AI". Incremental improvements may continue but the term is used more optimistically than evidence so far suggests. People underestimate the time and effort required for symbolic AI in narrow specialized domains, to say nothing of general domains. I don't see the combination making significant leaps in capability without a significant investment in research, if then.
Well said. See also CYC (https://en.wikipedia.org/wiki/Cyc). Neuro-symbolic is the great new hope (actually, it's been the great new hope for decades. Still not there but there's always hope.
Yes, Cyc is a great example of both worthwhile R&D as well as a worthwhile case study in how difficult symbolic reason is and is likely to remain for some time. We need better knowledge acquisition tools for symbolic reasoning as well as better tools for building and combining reasoners.
About five years ago (maybe more since time is melding together), I remember telling my boss — the dean of the college of science of technology — how most AI was nothing more than machine learning with better marketing and the rest was useless. Now I know how Cassandra felt.
I use Claude 3.7 Sonnet as a (baka-tensai - "stupid genius") research assistant. Despite its obvious limitations, I find it genuinely useful, the only caveat being of course that I have to keep a very close eye on what it's telling me. Even though it's extremely unlikely IMO that any system that relies on LLMs for any significant component of its cognition will ever achieve reliable human-level AGI, Claude and its cousins provide genuine utility (at least for this user). That said, although I'm willing to pay £18/month for a subscription, there's no way I can afford the new $100-200/month plans. Whether or not the AI labs will ever be able to turn a profit selling baka-tensai remains to be seen.
Tech companies are selling a solution in search of a problem, so no surprise ROI ain't great. They're convinced AI is going to be world-changing, but in what way, exactly... eh, that remains to be seen. So let's just shove it into everything and find out where it works!
AI differs from technological breakthroughs of the past in that it is amorphous. It's hard to define, different people mean different things when they say it, and the optimism for what it will do in the future comes mostly from imagining some version of it that's "better" than what we have today. And what does "better" mean? All anyone seems able to come up with is "able to do all the things human minds can do."
It's hard enough to even define what counts as AI right now. LLMs and image generators, we're agreed those are AI. Self-driving cars... yes, AI. How about recommendation algorithms (TikTok, YouTube, Netflix, Amazon...)? If those are AI today, were they also AI for the couple of decades before we decided to call them AI? When MS Word underlines stuff I wrote to suggest changes... AI? When the announcers on a football game tell you a team's "win probability", that AI? Are the ghosts that chase Pac Man around AI?
At some point, we're just taking "automated thing the computer does" and calling it AI, and then using our imaginations to extrapolate into the future: maybe we can automate a computer to write hit pop songs! Maybe we can automate a computer to look at MRI scans and diagnose disease! Maybe we can automate a computer to make scientific discoveries and find a novel solution to the problem of climate change! And so on.
This isn't analogous to the historical breakthroughs that AI optimists put forth. It's not like the discovery of fire or electricity, and it's not like the internet revolution. Those things were concrete. AI is whatever you dream it to be. Of course it's always going to disappoint.
LLM bots are all over X/Twitter. They are all over social media. They are operating scams all over. LLMs are helping hack. Israeli accounts are reporting serious problems on X being throttled by mass reporting of their accounts.
LLMs are good at information (or disinformation) warfare.
AI's "intelligence" has limits, but its business potential is real. Making money with AI means building the right solutions and finding the right market. The three year ROI might be small, but the gains could be big 10 years out. Like the Internet.
I was interested to read Google’s announcement on new hardware for inference. They seem to be doing massive amounts of inference already and are sensibly driving down the cost.
Dozens of hardware and software adjustments to come. Interesting times, if massively over hyped.
Your distinction between LLMs and what they evolve into is understandable.
Gary, you have been dead-on about the limitations of LLMs, but you still cheerlead baselessly about future "AI" developments. Yes, they will (probably) come eventually. But no one today has any clue about how to start writing code for them. New breakthroughs are needed, and the timing of those is completely unpredictable.
It's strength seems to be in marketing applications. But, the way it is being effectively used just seems so unethical and disturbing to the integrity of information.
For example, take a look of the clipped video I have here:
Where LLMs really excel is at what some refer to as “hallucinating” which is actually just “fabricating”.
But “hallucination” sounds better than “fabrication” (which sounds too much like simply lying), so the AI companies have called LLM propensities to spout falsehoods” hallucination”
The value is in automation. Some automation works better with classic ML, some might need gen AI, but most quite simply need stabillity and robustness. The biggest challenge to automation is organization inertia.
Those mocking "wasted" AI spending miss the historical pattern. When factories first adopted electric motors, they simply replaced central steam engines and saw minimal gains. It took decades before companies realized the true innovation wasn't the motor itself, but distributing smaller motors throughout the factory floor, revolutionizing layouts and workflows.
We're at that same stage with AI. Most organizations are "electrifying their steam engines" - dropping AI into existing processes. Meanwhile, the successful 21% aren't just deploying AI - they're experimenting with reimagining their data architecture and processes from the ground up. The AI revolution isn't years away - it's happening now, but only for those thinking beyond simply replacing their metaphorical steam engines.
It’s been very dependent on what is being measured. We have been relying on projections of real business impact based on benchmarks or other less-relevant metrics for some time now:
One-fifth of the AI buyers getting good results are in the company of those like Spotify, Salesforce, and Netflix. Four-fifths are at risk for stagnating and going out of business.
Interesting news! I wonder if that is why my medical editing is picking up. For a year and a half, I'd be lucky to get one paper every other week; I had five this week.
I suspect that even of that 1/5 that are getting "good" results, they are unlikely objectively good. So many are impressed by LLMs eloquently crafted wording that they completely miss the nonsense it is spewing.
The only people that might be getting really good results are those who have jobs making spam content all over the web. It is decently good at that.
Also porn. If there is a "killer app" for AI, it's that. Except it can't possibly make enough money to pay for what it really costs.
PornAIgraphy: the only AI app that reveals the naked truth
Yes, anything in the category of engagement farming as well.
Do you think AI is bunk? Let us know below!
Glad you asked. Absolutely! The whole idea is preposterous. Only someone with no interest in, respect for, or knowledge of biology, evolution, and how intelligence actually works could have come up with such an idea. That so many, who ought to know better, seem to be smitten, if not bitten, by the same bug, is shocking.
Just the basic premise that approximations of incomplete information, extracted by inaccurate methods, using a sliver of sensory data, producing irrational, inaccurate, and downright bogus results, ought to be sufficient warning to wake anybody up, but instead, the poor methodology, along with the frequent, but constant, mismatch of results to reality, is just brushed aside, in the belief that it will work next time. ("Then a miracle happens.")
Humans have a long history of magical thinking, for example, religions, and alchemy, are two of many based on such thinking. To the true believer, unfortunately, there is no real-world corrective.
LLM: (10) Lords-a-Leaping Model (aka, the Ten Commandments)
I do wonder what the is the aftermath of the reseaech when the genai bubble bursts. and what will the new hyped reseaech direction be and how long will something with a comparable glamour to llms take to emerge.
The new hype appears to be "neurosumbolic AI". Incremental improvements may continue but the term is used more optimistically than evidence so far suggests. People underestimate the time and effort required for symbolic AI in narrow specialized domains, to say nothing of general domains. I don't see the combination making significant leaps in capability without a significant investment in research, if then.
Well said. See also CYC (https://en.wikipedia.org/wiki/Cyc). Neuro-symbolic is the great new hope (actually, it's been the great new hope for decades. Still not there but there's always hope.
Yes, Cyc is a great example of both worthwhile R&D as well as a worthwhile case study in how difficult symbolic reason is and is likely to remain for some time. We need better knowledge acquisition tools for symbolic reasoning as well as better tools for building and combining reasoners.
The aftermath of an LLM is s calculator.
finally the hallucinations ended
If calculators could talk:
“Man , these LLMs can talk the talk but sure can’t calc the calc”
Raise your mouse finger if you trust a chatbot to do your taxes.
IRS auditor: “what’s this claimed business credit here for 3 billion travel miles?”
Taxpayer: “Damned if I know. TaxGPT did my taxes. I’m not responsible for preparer hallucinations, right?
About five years ago (maybe more since time is melding together), I remember telling my boss — the dean of the college of science of technology — how most AI was nothing more than machine learning with better marketing and the rest was useless. Now I know how Cassandra felt.
I use Claude 3.7 Sonnet as a (baka-tensai - "stupid genius") research assistant. Despite its obvious limitations, I find it genuinely useful, the only caveat being of course that I have to keep a very close eye on what it's telling me. Even though it's extremely unlikely IMO that any system that relies on LLMs for any significant component of its cognition will ever achieve reliable human-level AGI, Claude and its cousins provide genuine utility (at least for this user). That said, although I'm willing to pay £18/month for a subscription, there's no way I can afford the new $100-200/month plans. Whether or not the AI labs will ever be able to turn a profit selling baka-tensai remains to be seen.
Tech companies are selling a solution in search of a problem, so no surprise ROI ain't great. They're convinced AI is going to be world-changing, but in what way, exactly... eh, that remains to be seen. So let's just shove it into everything and find out where it works!
AI differs from technological breakthroughs of the past in that it is amorphous. It's hard to define, different people mean different things when they say it, and the optimism for what it will do in the future comes mostly from imagining some version of it that's "better" than what we have today. And what does "better" mean? All anyone seems able to come up with is "able to do all the things human minds can do."
It's hard enough to even define what counts as AI right now. LLMs and image generators, we're agreed those are AI. Self-driving cars... yes, AI. How about recommendation algorithms (TikTok, YouTube, Netflix, Amazon...)? If those are AI today, were they also AI for the couple of decades before we decided to call them AI? When MS Word underlines stuff I wrote to suggest changes... AI? When the announcers on a football game tell you a team's "win probability", that AI? Are the ghosts that chase Pac Man around AI?
At some point, we're just taking "automated thing the computer does" and calling it AI, and then using our imaginations to extrapolate into the future: maybe we can automate a computer to write hit pop songs! Maybe we can automate a computer to look at MRI scans and diagnose disease! Maybe we can automate a computer to make scientific discoveries and find a novel solution to the problem of climate change! And so on.
This isn't analogous to the historical breakthroughs that AI optimists put forth. It's not like the discovery of fire or electricity, and it's not like the internet revolution. Those things were concrete. AI is whatever you dream it to be. Of course it's always going to disappoint.
LLM bots are all over X/Twitter. They are all over social media. They are operating scams all over. LLMs are helping hack. Israeli accounts are reporting serious problems on X being throttled by mass reporting of their accounts.
LLMs are good at information (or disinformation) warfare.
AI's "intelligence" has limits, but its business potential is real. Making money with AI means building the right solutions and finding the right market. The three year ROI might be small, but the gains could be big 10 years out. Like the Internet.
Some form of AI will hold big rewards; not so sure LLms will be central
I was interested to read Google’s announcement on new hardware for inference. They seem to be doing massive amounts of inference already and are sensibly driving down the cost.
Dozens of hardware and software adjustments to come. Interesting times, if massively over hyped.
Your distinction between LLMs and what they evolve into is understandable.
"Some form of AI will hold big rewards"
But when?
Gary, you have been dead-on about the limitations of LLMs, but you still cheerlead baselessly about future "AI" developments. Yes, they will (probably) come eventually. But no one today has any clue about how to start writing code for them. New breakthroughs are needed, and the timing of those is completely unpredictable.
It's strength seems to be in marketing applications. But, the way it is being effectively used just seems so unethical and disturbing to the integrity of information.
For example, take a look of the clipped video I have here:
https://www.mindprison.cc/p/dead-internet-at-scale
Where LLMs really excel is at what some refer to as “hallucinating” which is actually just “fabricating”.
But “hallucination” sounds better than “fabrication” (which sounds too much like simply lying), so the AI companies have called LLM propensities to spout falsehoods” hallucination”
That’s why they are so apt for marketing because most marketing involves fabrication
The term hallucination when applied to LLMs is total BS, of course
Frightening.
AI's "intelligence" has limits, but its business potential is real“
That’s because the public’s intelligence has even more severe limits
The value is in automation. Some automation works better with classic ML, some might need gen AI, but most quite simply need stabillity and robustness. The biggest challenge to automation is organization inertia.
You can’t spell AI ROI without AIR
To err is human but to AIR is AI
We are nowhere near AGI.
However, the recent models, Gemini 2.5 Pro in particular, are remarkably better than what was there last year.
I think this year we will see more progress on tune-up and deployment and figuring out where they work well enough and where not.
Historically, the path to adoption is bumpy and lengthy.
Those mocking "wasted" AI spending miss the historical pattern. When factories first adopted electric motors, they simply replaced central steam engines and saw minimal gains. It took decades before companies realized the true innovation wasn't the motor itself, but distributing smaller motors throughout the factory floor, revolutionizing layouts and workflows.
We're at that same stage with AI. Most organizations are "electrifying their steam engines" - dropping AI into existing processes. Meanwhile, the successful 21% aren't just deploying AI - they're experimenting with reimagining their data architecture and processes from the ground up. The AI revolution isn't years away - it's happening now, but only for those thinking beyond simply replacing their metaphorical steam engines.
It’s been very dependent on what is being measured. We have been relying on projections of real business impact based on benchmarks or other less-relevant metrics for some time now:
https://open.substack.com/pub/mohakshah/p/genais-true-worth-measuring-what?r=224koz&utm_medium=ios
Gary, is there a way to contact you and/or Ernie Davis for a brief about a small project working on a compositional non-LLM architectural approach.
Gary, do you have any take on the evolution of image generations by current models and possible copyright violations ?
One-fifth of the AI buyers getting good results are in the company of those like Spotify, Salesforce, and Netflix. Four-fifths are at risk for stagnating and going out of business.
"There were possible..."
Interesting news! I wonder if that is why my medical editing is picking up. For a year and a half, I'd be lucky to get one paper every other week; I had five this week.