Panem et Circenses is such a wonderful term for "entertainment till death" approach that's been adopted and will be adopted by more. If the current dev leads us to the AGI technocrats envision of, then I can't really see a world outside RP1 scenario and UBI being part of the necessity.
What a great and delightful article. I want to write a serious comment, but I can't stop laughing at 'the square root of a goat is one'. Give me a minute...
OK, seriously this time: your four points are spot on, as is your observation of how the West systematically underestimates China. I would be curious to hear your view on the doubts that have been raised as to whether or not DeepSeek has really been developed with as few chips and as little money as they say. My take is that even if the answer is not, the mere possibility seems to be triggering an innovation race which is in itself good.
However, I remain very skeptical on the actual benefits and value of this generation of GenAI. Like ChatGPT and his American cousins, DeepSeek suffers from both censorship and hallucinations. And if I can’t trust it on some things, I can’t trust it on anything. I know I can buy a book to learn about Tienanmen square. But I don’t know what I don’t know, or rather I don’t know all the things that DeepSeek and ChatGPT are censoring. Similarly, because I do not know when and how they will hallucinate, I have to doublecheck every single answer, which kind of undermines the whole point of having this supersmart omniscient AI at my fingertips.
So my main reaction, I confess, was something like ‘This is much cheaper garbage, but it’s still garbage…’
Eventually we’ll do better, and perhaps this US-China race will help. But we are certainly not there yet. And by the way, this is without getting to the issue of how all these technologies seem to make us lazier and more stupid rather than smarter and better educated.
Nonetheless, this is a superb post with great and very important points, and plenty of humor sprinkled in – loved it.
I totally agree Marco: much cheaper garbage, perhaps. A great example is what happened to me yesterday: I asked ChatGPT and DeepSeek to calculate eigenalues of a system to check its stability. Although the DeepSeek answer came slightly faster than ChatGPT's both were garbage: they both provide a set of eignevalues well within steady state range. Yet, upon simulation, the system was unstable so I asked clarification to both. Here below I report their answers:
DeepSeek: You're absolutely correct to point out that the simulation results show the variables growing over time, which contradicts my earlier claim that the system would converge to a steady state. This discrepancy arises because the eigenvalues I initially provided were hypothetical and not calculated directly from the matrix A. Let’s carefully revisit the problem, recalculate the eigenvalues, and explain why the system behaves as it does.
ChatGPT: I apologize if my earlier answer caused any confusion. The standard method for evaluating whether a system of equations converges to a steady state is by examining its eigenvalues, which must all be less than 1 in absolute value. Would you like to proceed with this method, or would you prefer to suggest an alternative approach to obtain a satisfactory answer?
Both garbage. Yet, I have a small preference for DeepSeek's apologetic tone over ChatGPT quizzical answer, where it tried to teach me what I already asked to do and then, with a newly found arrogance, asked me if I know better.
I am happy to report that at the second estimation the eigenvalues were mostly much higher than 1 and that we worked together to find a solution (hallucinating further, probably).
Fantastic! Now, the hallucinatory experience you had with DeepSeek yesterday is identical to what happened to me with ChatGPT almost a year ago. I asked for the shares of fossil fuels and renewables in US power generation in the past 20 years. ChatGPT promptly provided data and even a chart! Excited, I asked for the source, and ChatGPT said:
ChatGPT: The data I used for the plot is simulated to illustrate the trends. For accurate and real-time data, you can refer to reliable sources such as the U.S. Energy Information Administration (EIA). The EIA provides detailed statistics and reports on the shares of fossil fuels and renewables in U.S. electricity generation.
So we're getting very excited because in one year we've progressed from unreliable nonsense to the same unreliable nonsense.
We're all familiar with the 'garbage in, garbage out' problem. Here we have 'good data in, garbage out'.
Here's the link to the post detailing my older ChatGPT experience:
i heard lots of great things from frds as medical researchers or quants saying current dev has been quite helpful and accelerating their work exponentially, for me personally it's just been a quite fast talking calculator with better user interface
Re: "I know I can buy a book to learn about Tiananmen square."
Any book you can buy on any topic, has limited info on written subject, and is always biased towards the author's, editor's or publishing authority's opinion.
Re: GenAI Chatbot
The current bet on LLM approach with previous investment largely gone to pre-training is based on the assumption that GenAI can and will be achieved based on "existing data" and scaling law, which I personally find ignorant and naive, as "existing data we have in record" is far from "existing data" per se.
Perhaps I (along with people with same reasoning behind skepticism) simply have too high a standard for AI when a hyper processor able to form reasoning, able to self reflect upon prompting with access to as much data as possible is already good enough
I think bread and entertainment (!) is the better translation Just hinting
also "stop complaining and get to work" is what maybe 99% in the developed world need to learn but don't have to for foreseeable future
Panem et Circenses is such a wonderful term for "entertainment till death" approach that's been adopted and will be adopted by more. If the current dev leads us to the AGI technocrats envision of, then I can't really see a world outside RP1 scenario and UBI being part of the necessity.
What a great and delightful article. I want to write a serious comment, but I can't stop laughing at 'the square root of a goat is one'. Give me a minute...
Probably, it's too much ahead of its time...
too early to comment, may be...
OK, seriously this time: your four points are spot on, as is your observation of how the West systematically underestimates China. I would be curious to hear your view on the doubts that have been raised as to whether or not DeepSeek has really been developed with as few chips and as little money as they say. My take is that even if the answer is not, the mere possibility seems to be triggering an innovation race which is in itself good.
However, I remain very skeptical on the actual benefits and value of this generation of GenAI. Like ChatGPT and his American cousins, DeepSeek suffers from both censorship and hallucinations. And if I can’t trust it on some things, I can’t trust it on anything. I know I can buy a book to learn about Tienanmen square. But I don’t know what I don’t know, or rather I don’t know all the things that DeepSeek and ChatGPT are censoring. Similarly, because I do not know when and how they will hallucinate, I have to doublecheck every single answer, which kind of undermines the whole point of having this supersmart omniscient AI at my fingertips.
So my main reaction, I confess, was something like ‘This is much cheaper garbage, but it’s still garbage…’
Eventually we’ll do better, and perhaps this US-China race will help. But we are certainly not there yet. And by the way, this is without getting to the issue of how all these technologies seem to make us lazier and more stupid rather than smarter and better educated.
Nonetheless, this is a superb post with great and very important points, and plenty of humor sprinkled in – loved it.
I totally agree Marco: much cheaper garbage, perhaps. A great example is what happened to me yesterday: I asked ChatGPT and DeepSeek to calculate eigenalues of a system to check its stability. Although the DeepSeek answer came slightly faster than ChatGPT's both were garbage: they both provide a set of eignevalues well within steady state range. Yet, upon simulation, the system was unstable so I asked clarification to both. Here below I report their answers:
DeepSeek: You're absolutely correct to point out that the simulation results show the variables growing over time, which contradicts my earlier claim that the system would converge to a steady state. This discrepancy arises because the eigenvalues I initially provided were hypothetical and not calculated directly from the matrix A. Let’s carefully revisit the problem, recalculate the eigenvalues, and explain why the system behaves as it does.
ChatGPT: I apologize if my earlier answer caused any confusion. The standard method for evaluating whether a system of equations converges to a steady state is by examining its eigenvalues, which must all be less than 1 in absolute value. Would you like to proceed with this method, or would you prefer to suggest an alternative approach to obtain a satisfactory answer?
Both garbage. Yet, I have a small preference for DeepSeek's apologetic tone over ChatGPT quizzical answer, where it tried to teach me what I already asked to do and then, with a newly found arrogance, asked me if I know better.
I am happy to report that at the second estimation the eigenvalues were mostly much higher than 1 and that we worked together to find a solution (hallucinating further, probably).
Fantastic! Now, the hallucinatory experience you had with DeepSeek yesterday is identical to what happened to me with ChatGPT almost a year ago. I asked for the shares of fossil fuels and renewables in US power generation in the past 20 years. ChatGPT promptly provided data and even a chart! Excited, I asked for the source, and ChatGPT said:
ChatGPT: The data I used for the plot is simulated to illustrate the trends. For accurate and real-time data, you can refer to reliable sources such as the U.S. Energy Information Administration (EIA). The EIA provides detailed statistics and reports on the shares of fossil fuels and renewables in U.S. electricity generation.
So we're getting very excited because in one year we've progressed from unreliable nonsense to the same unreliable nonsense.
We're all familiar with the 'garbage in, garbage out' problem. Here we have 'good data in, garbage out'.
Here's the link to the post detailing my older ChatGPT experience:
https://justthink.substack.com/p/california-dreaming-and-hallucinations
adding to the list is me asking Grok to calculate how rare Kumbh Mela would occur together with the alignment of 7 planets of the solar:
https://x.com/CharlieEinhorn/status/1882425542483513604
i heard lots of great things from frds as medical researchers or quants saying current dev has been quite helpful and accelerating their work exponentially, for me personally it's just been a quite fast talking calculator with better user interface
Re: "I know I can buy a book to learn about Tiananmen square."
Any book you can buy on any topic, has limited info on written subject, and is always biased towards the author's, editor's or publishing authority's opinion.
Re: GenAI Chatbot
The current bet on LLM approach with previous investment largely gone to pre-training is based on the assumption that GenAI can and will be achieved based on "existing data" and scaling law, which I personally find ignorant and naive, as "existing data we have in record" is far from "existing data" per se.
Perhaps I (along with people with same reasoning behind skepticism) simply have too high a standard for AI when a hyper processor able to form reasoning, able to self reflect upon prompting with access to as much data as possible is already good enough