I think that the biggest loss will be in first-level customer service jobs. They rely on a limited amount of relevant documents, _and_ they are notoriously not very reliably good. This seems like a perfect area for AI, since any errors will be nothing new.
I would agree with this, but iirc an airline company tried this and was sued when the AI provided incorrect information. They couldn't say the AI was liable, since it wasn't a person, so the airline company itself was liable for damages.
A change in laws could ameliorate this though. Or I think a simple fine-print statement like "statements from the chatbot may be incorrect" may have sufficed.
Ironically, the airline company tried to make a legal case for AI personhood in their defense - possibly the first in history so far.
LLMs are text compression algos. They are very good at memorising & retrieving text related to the user input. They can even pass bar exams based on that - some misinterpret that as intelligence.
However, nothing (even scaling) can change a text memorisation algo into a symbolic reasoning or composability algo, both of which are necessary for progress towards AGI.
Until research moves on from LLMs, AGI will remain elusive.
It's fascinating to read this piece. Because just today I am seeing a plethora of articles in the media purporting how AGI and ASI are nearly upon us. Forbes is especially bad for that right now.
OK before I close this window on my computer, I may as well post this here...
I just renewed my subscription to Microsoft 365 Family, expecting that it would include at least SOME GAI features. No such luck. So I figured this was something that you had to download, and looked around. No such luck. There are plenty of pages on Microsoft touting these features, but apparently it's only for the confusingly named Microsoft 365 product ... which is apparently only for corporation and bulk purchasers. Since I couldn't find a straight answer myself, I tried Microsoft's Help/support chat window. I was fourth in line for an agent, expecting it to get me to one shortly. I gave up about two hours later (seriously - and the number DID go down)!
So I gave up and asked the chatbot. Here's the conversation:
"Hi, I'm a Store Assistant powered by Al. I can help you with shopping and support. I can also connect you with a Microsoft Sales Associate. What can I do for you?"
Me: "Can I get the free version of copilot in my 365 family subscription?"
It: "Yes, you can access the free version of Copilot with your Microsoft 365
Family subscription. The free version of Copilot allows you to find information, create original content, and complete tasks quickly with features like text, voice, and image capabilities. For more details, you can visit the Microsoft 365 Copilot page [link]."
This is false.
Conclusion: (1) Microsoft's chatbot hallucinates about their own products! (2) Their product naming and marketing materials are, to put it mildly, less than competent.
We just, and might I add reluctantly, replaced our Windows 7, internet connected machine with a Windows 11 one and it's been a terrible experience. Some semblance of usability has finally been achieved by ripping out all the "AI" "features" that could be uninstalled or disabled, most especially Micro$oft Autopilot. My former colleague Tony Andrews coined a term that applies to 11's features: Deprovements.
Only after seeing this did it occur to me the LLMs are trained on as much or more marketing material, which tends to be vacuous, and as often as not deceitful, about what a product can do, as it is trained on technical manuals and help files.
No wonder these things provide text that is confidently wrong! :)
technical moat would be like “OpenAI [or some other company X] knows how to build something nobody else can build in the next 2 or 3 years, so there is a clear technical reason to favor them”
The challenge with any prediction about work force replacement is: How do you know? We have media-celebrated cases like Klarna, but suspicious motives generate suspicious data. A recent Adecco article says that 13% of workers in 2024 “reported having lost their jobs” because of AI — without reporting a source. What is most likely is that there will be a few clear examples (customer support, freelance creators), and a lot of messy examples (Microsoft *may* hire fewer programmers — but tech companies have already laid off tons of people before genAI gained widespread use).
What is needed are clearer reporting guidelines by employers, which should be a key part of any company's AI ethics rules. -gB
Really excellent thoughts as always - thanks for continuing to provide a common sense grounding in these topics amidst the endless hype. And Happy New Year!
* "Predtictions" is misspelled (should be "Predictions") in the heading before the "25 Predictions for 2025" section.
* "is still not generally available; o3 has been previewed but not released likely will likely be quite expensive" has a doubled "likely" — it should probably read "...but not released and will likely be quite expensive" or similar.
* "Moreover Metaculus is the process setting up a webpage" appears to be missing "in" — should read "Moreover Metaculus is in the process of setting up a webpage"
"You, you people!, are all making fundamental errors in understanding the technologies and where their boundaries lie. Many of them will be useful technologies but their imagined capabilities are just not going to come about in the time frames the majority of the technology and prognosticator class, deeply driven by FOBAWTPALSL, think.
"But this time it is different you say. This time it is really going to happen. You just don’t understand how powerful AI is now, you say. All the early predictions were clearly wrong and premature as the AI programs were clearly not as good as now and we had much less computation back then. This time it is all different and it is for sure now."
The prediction about a lack of consumer protections for US citizens is a bit blue, but not wrong either. Unless there is a radical shift in US politics in the next year, other countries are going to have to pull their weight in protecting consumers and citizens from the negative effects of AI deployment, no matter how many shareholders and Twitter bluechecks pout about it.
For fun, I translated some of these predictions as if Gary Marcus was an AI bull:
* AGI could arrive as soon as 2026!
* Up to 4 of the AI 2027 Marcus-Brundage tasks could be solved by a single system in 2025, and possibly more by multiple systems working together.
* Chip-making companies will continue to do well.
* Stifling regulation will be minimal in the US, allowing more AI development and progress than in Europe.
* AI Agents will be a popular topic throughout 2025, and they could function reliably in certain use cases.
* Humanoid robotics will be a popular topic as well, and their motor control might be impressive. However, they won't be quite as good as the fictional robot from The Jetsons.
* Truly driverless cars will be used in several cities, and semi-driverless cars could be even more widely available. However, for the time being, human drivers will still make up a large part of the economy.
* AI companies will continue to scale up their electric power infrastructure and capabilities.
* Up to 5% of the work force could be replaced by AI, and possibly even up to 10%. Many more jobs will be modified as people begin to use new tools.
* People will intially be amazed at o3, and it will work best in domains like math problems.
* Advancement in AI technology will remain competitive globally, rather than a single firm or country becoming a monopoly.
* Companies will continue to experiment with AI, with tentative adoption to production-grade systems scaled out in the real-world.
* Neurosymbolic AI will become much more prominent.
* There is real potential for a "GPT-5 level” model (meaning a huge, across the board quantum leap forward as judged by community consensus) throughout 2025.
* Even without a GPT-5 level model, we may see models like o1 that are quite good at many tasks for which high-quality synthetic data can be created.
This is a great first read for the start of the year! Felt compelled enough to write a comment. From my experience as a game designer and playing around in Unity. I'm not surprised about the issue with physics in world generation with Sora. Everything is predictive but with rote memorization and varying levels of confidence in recalling visual information, but it's not calculating and adapting to real-time changing physics like a game engine does. When it can do that some other issues with reasoning and factuality can be solved imho.
As always thought-provoking and astute. I would add - and have put this in my own much more humble technology predictions for 2025 - that cybercrime will continue to track upwards in part because AI gives the criminals more tools to work with. Thanks for these.
As somebody who is worried by unemployment caused by generative AI, less than 5 to 10 percent unemployment would still be massive in just one year.
I think companies will blame AI for layoffs as cover, even when it has nothing to do with AI. 🤷♀️
When OpenAI lays off personnel, will Sam blame it on AI?
Can’t say, but there will be an “awe shucks” flavor to the delivery and he will somehow skirt responsibility 🙃
He’ll probably blame it on Grok.
I think that the biggest loss will be in first-level customer service jobs. They rely on a limited amount of relevant documents, _and_ they are notoriously not very reliably good. This seems like a perfect area for AI, since any errors will be nothing new.
I would agree with this, but iirc an airline company tried this and was sued when the AI provided incorrect information. They couldn't say the AI was liable, since it wasn't a person, so the airline company itself was liable for damages.
A change in laws could ameliorate this though. Or I think a simple fine-print statement like "statements from the chatbot may be incorrect" may have sufficed.
Ironically, the airline company tried to make a legal case for AI personhood in their defense - possibly the first in history so far.
Doesn’t 0% actually fall within that category?
LLMs are text compression algos. They are very good at memorising & retrieving text related to the user input. They can even pass bar exams based on that - some misinterpret that as intelligence.
However, nothing (even scaling) can change a text memorisation algo into a symbolic reasoning or composability algo, both of which are necessary for progress towards AGI.
Until research moves on from LLMs, AGI will remain elusive.
It's fascinating to read this piece. Because just today I am seeing a plethora of articles in the media purporting how AGI and ASI are nearly upon us. Forbes is especially bad for that right now.
Forbes has a contributor model with little oversight. Contributors get paid per click. It has no credibility at this point. Tell your friends. https://www.niemanlab.org/2022/02/an-incomplete-history-of-forbes-com-as-a-platform-for-scams-grift-and-bad-journalism/
OK before I close this window on my computer, I may as well post this here...
I just renewed my subscription to Microsoft 365 Family, expecting that it would include at least SOME GAI features. No such luck. So I figured this was something that you had to download, and looked around. No such luck. There are plenty of pages on Microsoft touting these features, but apparently it's only for the confusingly named Microsoft 365 product ... which is apparently only for corporation and bulk purchasers. Since I couldn't find a straight answer myself, I tried Microsoft's Help/support chat window. I was fourth in line for an agent, expecting it to get me to one shortly. I gave up about two hours later (seriously - and the number DID go down)!
So I gave up and asked the chatbot. Here's the conversation:
"Hi, I'm a Store Assistant powered by Al. I can help you with shopping and support. I can also connect you with a Microsoft Sales Associate. What can I do for you?"
Me: "Can I get the free version of copilot in my 365 family subscription?"
It: "Yes, you can access the free version of Copilot with your Microsoft 365
Family subscription. The free version of Copilot allows you to find information, create original content, and complete tasks quickly with features like text, voice, and image capabilities. For more details, you can visit the Microsoft 365 Copilot page [link]."
This is false.
Conclusion: (1) Microsoft's chatbot hallucinates about their own products! (2) Their product naming and marketing materials are, to put it mildly, less than competent.
We just, and might I add reluctantly, replaced our Windows 7, internet connected machine with a Windows 11 one and it's been a terrible experience. Some semblance of usability has finally been achieved by ripping out all the "AI" "features" that could be uninstalled or disabled, most especially Micro$oft Autopilot. My former colleague Tony Andrews coined a term that applies to 11's features: Deprovements.
Only after seeing this did it occur to me the LLMs are trained on as much or more marketing material, which tends to be vacuous, and as often as not deceitful, about what a product can do, as it is trained on technical manuals and help files.
No wonder these things provide text that is confidently wrong! :)
Could you describe in more detail what you mean by technical moat?
Found this hilarious:
“shovels will continue to sell well throughout the gold rush.” 🤣🤣🤣
technical moat would be like “OpenAI [or some other company X] knows how to build something nobody else can build in the next 2 or 3 years, so there is a clear technical reason to favor them”
As opposed to a “tricknical mote”
The challenge with any prediction about work force replacement is: How do you know? We have media-celebrated cases like Klarna, but suspicious motives generate suspicious data. A recent Adecco article says that 13% of workers in 2024 “reported having lost their jobs” because of AI — without reporting a source. What is most likely is that there will be a few clear examples (customer support, freelance creators), and a lot of messy examples (Microsoft *may* hire fewer programmers — but tech companies have already laid off tons of people before genAI gained widespread use).
What is needed are clearer reporting guidelines by employers, which should be a key part of any company's AI ethics rules. -gB
Really excellent thoughts as always - thanks for continuing to provide a common sense grounding in these topics amidst the endless hype. And Happy New Year!
A few small typos:
* "NVidia" is misspelled (should Nvidia).
* "Predtictions" is misspelled (should be "Predictions") in the heading before the "25 Predictions for 2025" section.
* "is still not generally available; o3 has been previewed but not released likely will likely be quite expensive" has a doubled "likely" — it should probably read "...but not released and will likely be quite expensive" or similar.
* "Moreover Metaculus is the process setting up a webpage" appears to be missing "in" — should read "Moreover Metaculus is in the process of setting up a webpage"
You meant “Predations” was misspelled, right?
For comparison, here is Rodney Brooks' scorecard, 2025 January 01 https://rodneybrooks.com/predictions-scorecard-2025-january-01/
"You, you people!, are all making fundamental errors in understanding the technologies and where their boundaries lie. Many of them will be useful technologies but their imagined capabilities are just not going to come about in the time frames the majority of the technology and prognosticator class, deeply driven by FOBAWTPALSL, think.
"But this time it is different you say. This time it is really going to happen. You just don’t understand how powerful AI is now, you say. All the early predictions were clearly wrong and premature as the AI programs were clearly not as good as now and we had much less computation back then. This time it is all different and it is for sure now."
The prediction about a lack of consumer protections for US citizens is a bit blue, but not wrong either. Unless there is a radical shift in US politics in the next year, other countries are going to have to pull their weight in protecting consumers and citizens from the negative effects of AI deployment, no matter how many shareholders and Twitter bluechecks pout about it.
The author of the cartoon is Glauco Villas-Boas.
I came here to say the same and found you.
Glauco was an outstanding cartoonist from Brazil.
https://www.google.com/search?hl=pt-br&q=glauco+charges
I predict that the hype will continue to increase exponentially
Or maybe hyperbolically
For fun, I translated some of these predictions as if Gary Marcus was an AI bull:
* AGI could arrive as soon as 2026!
* Up to 4 of the AI 2027 Marcus-Brundage tasks could be solved by a single system in 2025, and possibly more by multiple systems working together.
* Chip-making companies will continue to do well.
* Stifling regulation will be minimal in the US, allowing more AI development and progress than in Europe.
* AI Agents will be a popular topic throughout 2025, and they could function reliably in certain use cases.
* Humanoid robotics will be a popular topic as well, and their motor control might be impressive. However, they won't be quite as good as the fictional robot from The Jetsons.
* Truly driverless cars will be used in several cities, and semi-driverless cars could be even more widely available. However, for the time being, human drivers will still make up a large part of the economy.
* AI companies will continue to scale up their electric power infrastructure and capabilities.
* Up to 5% of the work force could be replaced by AI, and possibly even up to 10%. Many more jobs will be modified as people begin to use new tools.
* People will intially be amazed at o3, and it will work best in domains like math problems.
* Advancement in AI technology will remain competitive globally, rather than a single firm or country becoming a monopoly.
* Companies will continue to experiment with AI, with tentative adoption to production-grade systems scaled out in the real-world.
* Neurosymbolic AI will become much more prominent.
* There is real potential for a "GPT-5 level” model (meaning a huge, across the board quantum leap forward as judged by community consensus) throughout 2025.
* Even without a GPT-5 level model, we may see models like o1 that are quite good at many tasks for which high-quality synthetic data can be created.
This is a great first read for the start of the year! Felt compelled enough to write a comment. From my experience as a game designer and playing around in Unity. I'm not surprised about the issue with physics in world generation with Sora. Everything is predictive but with rote memorization and varying levels of confidence in recalling visual information, but it's not calculating and adapting to real-time changing physics like a game engine does. When it can do that some other issues with reasoning and factuality can be solved imho.
hi gary; you might find this 'appreciation' of your current predictions for 2025 generated from OpenAI GPT 4.o mini (via attap.ai) fun.
https://attap.ai/chat/s/2b9b905a-5928-464b-80f5-005881701829/
As always thought-provoking and astute. I would add - and have put this in my own much more humble technology predictions for 2025 - that cybercrime will continue to track upwards in part because AI gives the criminals more tools to work with. Thanks for these.