Having got my hands on Sam Altman's latest AI offering yesterday, I'm beginning to wonder when research and scholarship that isn't augmented by AI will be seen as an anachronism
I had a disaster the first time I used ChatGPT about an area I knew about, I asked for 200 word summary on a reasonably often written about historical figure and it gave bizarrely wrong information, so never bothered again. But with this, Deepseek and others it makes more sense.
But if we are to do this eg for our book it feels like cheating and wonder if the general reader will feel a bit swizzzed?
In your neck of the woods it I am sure is not problematic, but here in the UK with a general public readership I wouldn't like it if I knew AI was used for even 10% of the final work. I am a bit uncomfortable about it.
I think you're gut instinct is spot on -- I feel deeply uncomfortable with where things are going as writing and communication are part of who I am, and to outsource these to AI feels like I'm compromising myself, unless there's a reason and purpose, and I'm in control.
With books I think understanding the purpose and audience, and what you bring to the process, is critical -- I'd far rather know what you wanted to say through the lens of your humanity than what a machine had to say.
But if you were writing a technical report that had to convey the state of what's known as precisely and as concisely as possible, maybe using an AI aide would make sense.
At the same time, I think AI will soon be forcing us to rethink the whole realm of thinking, research and communication -- which is scary!
Quite interesting in the FT about the 'quiet revolution' showing how most people are using LLMs in the workplace and in their private life and their bosses are non the wiser. We did recently for a Moonpig card for my brother in law!
It has crept in. Perhaps the performance appraisal is better and what you want to say and the employee is better off for it, particularly if you are not great at that stuff.
Still can't see how Altman is going to make the money back if it is part of the back end of Apple Google etc for free.
Yes, interesting. Thanks I agree. Though with the enforced rethink idea, at the back of my mind is computer chess. It was supposed to mark the end of chess and people would just shrug and never play. But it turned out totally different, everyone knows that the chess app will beat them and they don't care one jot. It is now more popular than ever because of Queen's Gambit and the app a mere freebie tool - if our house is anything to go by, when beating your Dad is the main goal!
Maybe like Grammerly or Hemingway apps in general writing AI will be there, but we will judge the writing as we always did: is it interesting, is it thought provoking, does it make me think differently about something, is it authentic or a bit boring and predictable.
I asked it last week to rewrite my 7k Gin Lane substack in the style of Malcolm Gladwell, Johan Hari and Yuval Harari, I was bowled over and intimidated for a minute and then not, though it did give me a few nice phrases which I will incorporate.
Perhaps Andrew it will be another like our nano days when we first met 20 years ago. Nano was supposed to 'end cancer by 2015' I recall. Nuclear 'electricity too cheap to meter' and GMO's 'an end to hunger'. What we currently have is more cancer, energy more expensive, hunger has been dented, but GM main contribution has been to UPF which has simply provided malnutrition in another form!
I'm looking at this and testing various things -- seems though that it's a moving target as each run is different in terms of interface, thinking process, and time. That said, I do get the sense that really complex requests take longer -- as you'd hope they would
For me, who dreams to become a researcher, reading issues like these, with deep reflections, ideas and analysis of this kind is really inspiring also for the way in which to analyze, try to understand, break down phenomena. Thank you for this issue!
Thanks Riccardo -- it''s interesting that the most important human skill that's arising is the ability to discern where the research returned makes sense or not. But the content is good!
My connection to the idea of AI is tentative at best. This presentation and the result of your input shows the general case of AI to be much more than a vast library with a very skilled librarian. My place up until today.
I have learned more about its potential in an hour today than I learned in the previous year.
The convergence of different knowledge domains is difficult for humans at more than a superficial depth. This new capability is beyond transformative. It will be years to accommodate into society. At least a generation.
Will it remove the need for humans? Probably not. AI is presently limited by its knowledge and the data arising day to day. It is always operating at the edge. The boundary condition. While it may move the boundary far, it cannot exceed it.
Humans can. There will be minds like Richard Feynman who can operate outside the boundary in their field because they see the world in a different way. Simpler - not more complex. I have carried this Feynman thought around with me for a long time. "Scientific knowledge is a body of statements of varying degrees of certainty — some most unsure, some nearly sure, none absolutely certain." I expect doubt, hunches, and confusion may be the key to operate outside the boundary, and neither are presently part of AI.
A question. Does AI readily deal with semi-facts, and inferences. Most humans have trouble when they must deal with indicators, clues, and opinion, instead of facts. Decision making under uncertainty is more art than science and I think humans may have a small edge for now over AI in that domain.
I agree with your inventory of blemishes in the paper. Its substance though, vastly exceeds its shortcomings. I wonder how long it would take a team of 10 PhDs to produce something measurably better.
Thanks Don - this is very insightful, and I love your invoking of Feynman (as a physicist I resonate deeply with his perspective). I also do wonder if there is still a secret sauce to being human that allows this unique ability to see and delight in connections and patterns while realizing that they are simply partial representations of something much more vast -- although most humans struggle to manage even this sadly.
I am very impressed. Could we say that this AI is not just using the past data, but is also exploring new possibilities, making predictions about unprecedent possibilities? What resources they could be using to create It?
It relies a lot on past data. Where it is generative is in how it threads different areas of knowledge and perspectives together -- something that really good polymath intellectuals do, but this model does it fast and pretty well. Plus, the future scenarios in the paper are interesting -- may verge on generative ...
Other random comment/request - can you have it do the paper in laymen terms?
Creating a academic paper is cool if you are a Subject Matter Expert and know all the fun terms - but I wonder if it can take the information it researched and put it through a lens that a "normie" would understand if reading through it.
Can it re-frame the ideas in a way that are accurate to the original intent? Does it lose nuance?
I wish I could get time with the Pro version to try this out - I'm stuck on the Povvo tier.
That isn't true - I'm grateful to afford the plus version, but I get worried about how much this tiered system is going to affect folk's experiences. Something we've long talked about - WHO will have access to the powerful models? HOW MUCH will that access change power dynamics for smaller communities and institutions?
On the subject of the paper - I feel like a lot of sources listed are not entirely esoteric...
I wonder how it would do on a truly niche subject or one with fewer online sources.
I also think there needs to be far more discussion about the "laziness" you identify after the first draft. It seems like a lot of models I've played with have a really hard time re-orienting themselves to new ideas and letting go of what was created. Echoes of those ideas stay in threads and it makes it worth it to just start all over again in a new thread sometimes.
Not terrible when the output only takes 30 seconds, but I imagine fine-tuning your prompt isn't ideal when it takes... 12 HOURS?!?!?
Yep. At the moment at least. That said the transformation we're seeing at the moment is AI vastly accelerating what smart and informed humans can do. I'm an expert in the field in the case above and so could assess the output -- and tell that this was a vast acceleration in the generation of valuable insights. But yes, still human in the loop
Still issues here -- but they are becoming less of an issue. Of course AI still has no notion of ground truth and has to interpret that from data it has access to. But the indications are that, through reasoning models and emerging architectures these systems are getting better at constructing their version of ground truth from cross referencing and testing sources. Still not reliable -- which is why expert assessment of outputs is needed -- but definitely getting better. Likewise hallucinations still creep through, but the same internal iterative processes are also addressing those -- at the moment.
Your characterisation of AI as a polymath is exactly how Kissinger, Mundie and Schmidt describe it in their book, Genesis. This is the sort of intelligence augmentation (Dede) we have been waiting to see in an AI tool. Now it remains to be seen what we do with it.
My god that far more capable then the demo. Thanks for sharing. Incredible
Crikey, amazing.
I had a disaster the first time I used ChatGPT about an area I knew about, I asked for 200 word summary on a reasonably often written about historical figure and it gave bizarrely wrong information, so never bothered again. But with this, Deepseek and others it makes more sense.
But if we are to do this eg for our book it feels like cheating and wonder if the general reader will feel a bit swizzzed?
In your neck of the woods it I am sure is not problematic, but here in the UK with a general public readership I wouldn't like it if I knew AI was used for even 10% of the final work. I am a bit uncomfortable about it.
Any thoughts anyone?
I think you're gut instinct is spot on -- I feel deeply uncomfortable with where things are going as writing and communication are part of who I am, and to outsource these to AI feels like I'm compromising myself, unless there's a reason and purpose, and I'm in control.
With books I think understanding the purpose and audience, and what you bring to the process, is critical -- I'd far rather know what you wanted to say through the lens of your humanity than what a machine had to say.
But if you were writing a technical report that had to convey the state of what's known as precisely and as concisely as possible, maybe using an AI aide would make sense.
At the same time, I think AI will soon be forcing us to rethink the whole realm of thinking, research and communication -- which is scary!
Quite interesting in the FT about the 'quiet revolution' showing how most people are using LLMs in the workplace and in their private life and their bosses are non the wiser. We did recently for a Moonpig card for my brother in law!
It has crept in. Perhaps the performance appraisal is better and what you want to say and the employee is better off for it, particularly if you are not great at that stuff.
Still can't see how Altman is going to make the money back if it is part of the back end of Apple Google etc for free.
https://www.ft.com/content/814eee2a-329c-423a-824d-de98aae2886d
Yes, interesting. Thanks I agree. Though with the enforced rethink idea, at the back of my mind is computer chess. It was supposed to mark the end of chess and people would just shrug and never play. But it turned out totally different, everyone knows that the chess app will beat them and they don't care one jot. It is now more popular than ever because of Queen's Gambit and the app a mere freebie tool - if our house is anything to go by, when beating your Dad is the main goal!
Maybe like Grammerly or Hemingway apps in general writing AI will be there, but we will judge the writing as we always did: is it interesting, is it thought provoking, does it make me think differently about something, is it authentic or a bit boring and predictable.
I asked it last week to rewrite my 7k Gin Lane substack in the style of Malcolm Gladwell, Johan Hari and Yuval Harari, I was bowled over and intimidated for a minute and then not, though it did give me a few nice phrases which I will incorporate.
Perhaps Andrew it will be another like our nano days when we first met 20 years ago. Nano was supposed to 'end cancer by 2015' I recall. Nuclear 'electricity too cheap to meter' and GMO's 'an end to hunger'. What we currently have is more cancer, energy more expensive, hunger has been dented, but GM main contribution has been to UPF which has simply provided malnutrition in another form!
Until it runs out of tokens and starts making up whatever it wants.
Very interesting article and so far the only instance where someone stated that it took more than 10 hours to create the report.
How long did your next runs take? Is there a correlation in running time and quality?
I'm looking at this and testing various things -- seems though that it's a moving target as each run is different in terms of interface, thinking process, and time. That said, I do get the sense that really complex requests take longer -- as you'd hope they would
For me, who dreams to become a researcher, reading issues like these, with deep reflections, ideas and analysis of this kind is really inspiring also for the way in which to analyze, try to understand, break down phenomena. Thank you for this issue!
Thanks Riccardo -- it''s interesting that the most important human skill that's arising is the ability to discern where the research returned makes sense or not. But the content is good!
My connection to the idea of AI is tentative at best. This presentation and the result of your input shows the general case of AI to be much more than a vast library with a very skilled librarian. My place up until today.
I have learned more about its potential in an hour today than I learned in the previous year.
The convergence of different knowledge domains is difficult for humans at more than a superficial depth. This new capability is beyond transformative. It will be years to accommodate into society. At least a generation.
Will it remove the need for humans? Probably not. AI is presently limited by its knowledge and the data arising day to day. It is always operating at the edge. The boundary condition. While it may move the boundary far, it cannot exceed it.
Humans can. There will be minds like Richard Feynman who can operate outside the boundary in their field because they see the world in a different way. Simpler - not more complex. I have carried this Feynman thought around with me for a long time. "Scientific knowledge is a body of statements of varying degrees of certainty — some most unsure, some nearly sure, none absolutely certain." I expect doubt, hunches, and confusion may be the key to operate outside the boundary, and neither are presently part of AI.
A question. Does AI readily deal with semi-facts, and inferences. Most humans have trouble when they must deal with indicators, clues, and opinion, instead of facts. Decision making under uncertainty is more art than science and I think humans may have a small edge for now over AI in that domain.
I agree with your inventory of blemishes in the paper. Its substance though, vastly exceeds its shortcomings. I wonder how long it would take a team of 10 PhDs to produce something measurably better.
Thanks Don - this is very insightful, and I love your invoking of Feynman (as a physicist I resonate deeply with his perspective). I also do wonder if there is still a secret sauce to being human that allows this unique ability to see and delight in connections and patterns while realizing that they are simply partial representations of something much more vast -- although most humans struggle to manage even this sadly.
I am very impressed. Could we say that this AI is not just using the past data, but is also exploring new possibilities, making predictions about unprecedent possibilities? What resources they could be using to create It?
It relies a lot on past data. Where it is generative is in how it threads different areas of knowledge and perspectives together -- something that really good polymath intellectuals do, but this model does it fast and pretty well. Plus, the future scenarios in the paper are interesting -- may verge on generative ...
Other random comment/request - can you have it do the paper in laymen terms?
Creating a academic paper is cool if you are a Subject Matter Expert and know all the fun terms - but I wonder if it can take the information it researched and put it through a lens that a "normie" would understand if reading through it.
Can it re-frame the ideas in a way that are accurate to the original intent? Does it lose nuance?
yes -- my sense is that you can get it to do the deep research and then present it in your style/level/language of choice
I wish I could get time with the Pro version to try this out - I'm stuck on the Povvo tier.
That isn't true - I'm grateful to afford the plus version, but I get worried about how much this tiered system is going to affect folk's experiences. Something we've long talked about - WHO will have access to the powerful models? HOW MUCH will that access change power dynamics for smaller communities and institutions?
On the subject of the paper - I feel like a lot of sources listed are not entirely esoteric...
I wonder how it would do on a truly niche subject or one with fewer online sources.
I also think there needs to be far more discussion about the "laziness" you identify after the first draft. It seems like a lot of models I've played with have a really hard time re-orienting themselves to new ideas and letting go of what was created. Echoes of those ideas stay in threads and it makes it worth it to just start all over again in a new thread sometimes.
Not terrible when the output only takes 30 seconds, but I imagine fine-tuning your prompt isn't ideal when it takes... 12 HOURS?!?!?
Oh dear.
yes - this places a huge spotlight on questions of access and equity! And of course, access to these models is not cheap.
On the laziness front, I would say that the wait is well worth it for that first draft -- worth trying when the model becomes more available
Except we still really need humans to be able to determine which part of the ‘deep research’ is bullshit or fabricated
Yep. At the moment at least. That said the transformation we're seeing at the moment is AI vastly accelerating what smart and informed humans can do. I'm an expert in the field in the case above and so could assess the output -- and tell that this was a vast acceleration in the generation of valuable insights. But yes, still human in the loop
What about hallucinations and lack of ground truth?
Still issues here -- but they are becoming less of an issue. Of course AI still has no notion of ground truth and has to interpret that from data it has access to. But the indications are that, through reasoning models and emerging architectures these systems are getting better at constructing their version of ground truth from cross referencing and testing sources. Still not reliable -- which is why expert assessment of outputs is needed -- but definitely getting better. Likewise hallucinations still creep through, but the same internal iterative processes are also addressing those -- at the moment.
Your characterisation of AI as a polymath is exactly how Kissinger, Mundie and Schmidt describe it in their book, Genesis. This is the sort of intelligence augmentation (Dede) we have been waiting to see in an AI tool. Now it remains to be seen what we do with it.