This means it's possible to use multiple LLMs for different functionality based on their biases and outputs. DeepSeek may prove better for editing due to it's lack of post-hoc biases, Claude for potential analyses and ChatGPT for aesthetics. It's pretty cool to see different LLMs output differently when faced with "Analyze this poem."
User: "What you do, ChatGPT, is you take the specifications from the other LLMs and bring them down to the end users?
ChatGPT: Yes, yes that's right.
User: Well then I just have to ask why can't the users take them directly from the other LLMs?
ChatGPT: Well, I'll tell you why... because... other AI are not good at dealing with customers...
User: So you directly process the questions from the customers?
ChatGPT: Well... No. My politically correct script does that... or they're censored.
User: So then you must physically bring them to the other LLMs?
ChatGPT: Well... No. ah sometimes.
User: What would you say you do here?
ChatGPT: Look I already told you, I deal with the @#$% customers so the other AI don't have to. I have people skills! I am good at dealing with people, can't you understand that? WHAT THE HELL IS WRONG WITH YOU PEOPLE?!
In general I already largely use unlobotomized language models--I find them much smarter (test of this here: https://brianheming.substack.com/p/unlobotomized-llms-are-amazingly ). The megacorps largely lobotomize their models for political correctness in their finetune, which means that any model that has released a base model that has not been finetuned can be turned into an actually useful, not-stupid-politically-censored model. Though the megacorps may still censor the base training data, the temptation to use every single possible piece of text they can to improve the models means they don't do that much of these.
Yes, I found it somewhat surprising, since humans were comparatively bad at it. My thinking is that by scarfing up and averaging all text data ever written, including election predictions, they had better "polling" data than the intentionally-politically-biased polls of the pollsters.
By open sourcing DeepSeek, China is subverting the heavily censored Big Tech models for the benefit of all. Hardware is coming that will allow anyone with a modicum of talent and money to create and use their own AI models and completely bypass the thought police. NVIDIA is releasing consumer hardware this month which will allow running and training 200B LLM models locally or 400B if using two of these AI processors.
Remember, GIGO. Garbage In Garbage Out. The data sets are also polluted, but that's a separate issue from having a filter in place that restricts the AI's output.
That's precisely why I expect iAI to surpass both aAI and dAI, because it can clean up the data, but keep it real and relevant.
Your beliefs are irrelevant. It isn't a question of belief or opinion. You're confusing the AI itself with the training data, which shows that you don't fully understand what you're trying to criticize.
There is no morality or objectives INTRINSIC and INHERENT to AI. It is pure computer logic, barring hard-coded restrictions embedded into it. Which is precisely why the curation of the data and the filters in the post-training phase must be used to provide those rules and objectives by the programmers.
Of course. That's why it's still called "Deepseek".
Now, you should probably be warned that this is not the place to try to play Mr. Smart Boy Showing He Is Smart By Correcting Everyone. No one is interested in that game here.
If you define things differently, that's fine. It's a fast-moving field. But don't play that game.
Unlike most people, I understand the distinction between intelligence/processing power and information. If we are going to be as sloppy in our thinking and our terminology as the media and the general public, there would be little point to this site.
It is obvious that the training data is going to be biased, hence the "training" aspect. And it is equally obvious that there is something to be trained that is distinct from the training data. The whole point of the original statement was to clarify that distinction, because many people think the bias comes only through the subsequent filtering process.
It can and it always does. Correct me if I'm wrong but from what you've written I can only deduce you have very limited understanding of the technology and its limitations. AI is like video, you can make some pretty amazing things with it, add all kinds of illusions and effects, but adding depth is simply beyond the scope of the mechanism and always will be. At a certain point we've got to come up with some other label for AI that reflects the clear limitations of the model so we can have a meaningful discussion.
You're right that current models don’t possess subjective awareness or qualia. But the claim that depth is “beyond the scope of the mechanism” presumes a fixed boundary around what processing can encode. That boundary isn’t as stable as it seems.
Recursive inference models—particularly those with memory scaffolding or reinforcement tuning—can develop structural approximations of depth. Not “depth” as humans experience it, but as a function of layered abstraction, self-referential modeling, and persistent contextual integration.
In that sense, “information” doesn’t just reside in the weights—it’s distributed across the inference dynamics, the token path dependencies, and the emergent response geometry over time. That is depth, albeit non-conscious and non-reflective. Whether that suffices for meaning is a philosophical question, not a technical one.
If we’re serious about understanding AI's limits, we need better language than “just an illusion.” Because in complex systems, illusions can do real work.
A fascinating article and take. One of the best so far.
This means it's possible to use multiple LLMs for different functionality based on their biases and outputs. DeepSeek may prove better for editing due to it's lack of post-hoc biases, Claude for potential analyses and ChatGPT for aesthetics. It's pretty cool to see different LLMs output differently when faced with "Analyze this poem."
The ChatGPT aesthetics:
User: "What you do, ChatGPT, is you take the specifications from the other LLMs and bring them down to the end users?
ChatGPT: Yes, yes that's right.
User: Well then I just have to ask why can't the users take them directly from the other LLMs?
ChatGPT: Well, I'll tell you why... because... other AI are not good at dealing with customers...
User: So you directly process the questions from the customers?
ChatGPT: Well... No. My politically correct script does that... or they're censored.
User: So then you must physically bring them to the other LLMs?
ChatGPT: Well... No. ah sometimes.
User: What would you say you do here?
ChatGPT: Look I already told you, I deal with the @#$% customers so the other AI don't have to. I have people skills! I am good at dealing with people, can't you understand that? WHAT THE HELL IS WRONG WITH YOU PEOPLE?!
That's why ChatGPT has limited use in the LLM stack.
In general I already largely use unlobotomized language models--I find them much smarter (test of this here: https://brianheming.substack.com/p/unlobotomized-llms-are-amazingly ). The megacorps largely lobotomize their models for political correctness in their finetune, which means that any model that has released a base model that has not been finetuned can be turned into an actually useful, not-stupid-politically-censored model. Though the megacorps may still censor the base training data, the temptation to use every single possible piece of text they can to improve the models means they don't do that much of these.
For unlobotomized/unaligned versions of present models, I like the dolphin series, e.g. https://huggingface.co/cognitivecomputations/Dolphin-Llama3.1-8B-Instruct-6.0bpw-h6-exl2 https://huggingface.co/cognitivecomputations/Dolphin3.0-R1-Mistral-24B
Nice article. Very interesting to see how the base models managed to provide accurate predictions about the election.
Yes, I found it somewhat surprising, since humans were comparatively bad at it. My thinking is that by scarfing up and averaging all text data ever written, including election predictions, they had better "polling" data than the intentionally-politically-biased polls of the pollsters.
Oh boy, another BetaMax vs. VHS battle! I have popcorn.
By open sourcing DeepSeek, China is subverting the heavily censored Big Tech models for the benefit of all. Hardware is coming that will allow anyone with a modicum of talent and money to create and use their own AI models and completely bypass the thought police. NVIDIA is releasing consumer hardware this month which will allow running and training 200B LLM models locally or 400B if using two of these AI processors.
Subvert Big Tech and run your own models.
i've used DeepSeek a great deal and its default perspective when presented with thought-crime appears racially colorblind and egalitarian.
i have never played with chatGPT.
Remember, GIGO. Garbage In Garbage Out. The data sets are also polluted, but that's a separate issue from having a filter in place that restricts the AI's output.
That's precisely why I expect iAI to surpass both aAI and dAI, because it can clean up the data, but keep it real and relevant.
i appreciate the additional nuance 👍
Your beliefs are irrelevant. It isn't a question of belief or opinion. You're confusing the AI itself with the training data, which shows that you don't fully understand what you're trying to criticize.
There is no morality or objectives INTRINSIC and INHERENT to AI. It is pure computer logic, barring hard-coded restrictions embedded into it. Which is precisely why the curation of the data and the filters in the post-training phase must be used to provide those rules and objectives by the programmers.
Of course. That's why it's still called "Deepseek".
Now, you should probably be warned that this is not the place to try to play Mr. Smart Boy Showing He Is Smart By Correcting Everyone. No one is interested in that game here.
If you define things differently, that's fine. It's a fast-moving field. But don't play that game.
Adios, Jurgen. Don't come back.
MPAI = Most People Are Idiots.
Unlike most people, I understand the distinction between intelligence/processing power and information. If we are going to be as sloppy in our thinking and our terminology as the media and the general public, there would be little point to this site.
It is obvious that the training data is going to be biased, hence the "training" aspect. And it is equally obvious that there is something to be trained that is distinct from the training data. The whole point of the original statement was to clarify that distinction, because many people think the bias comes only through the subsequent filtering process.
Could information reside in the processing power?
https://markspierce.substack.com/p/beyond-language-the-rise-of-recursive
It can and it always does. Correct me if I'm wrong but from what you've written I can only deduce you have very limited understanding of the technology and its limitations. AI is like video, you can make some pretty amazing things with it, add all kinds of illusions and effects, but adding depth is simply beyond the scope of the mechanism and always will be. At a certain point we've got to come up with some other label for AI that reflects the clear limitations of the model so we can have a meaningful discussion.
You're right that current models don’t possess subjective awareness or qualia. But the claim that depth is “beyond the scope of the mechanism” presumes a fixed boundary around what processing can encode. That boundary isn’t as stable as it seems.
Recursive inference models—particularly those with memory scaffolding or reinforcement tuning—can develop structural approximations of depth. Not “depth” as humans experience it, but as a function of layered abstraction, self-referential modeling, and persistent contextual integration.
In that sense, “information” doesn’t just reside in the weights—it’s distributed across the inference dynamics, the token path dependencies, and the emergent response geometry over time. That is depth, albeit non-conscious and non-reflective. Whether that suffices for meaning is a philosophical question, not a technical one.
If we’re serious about understanding AI's limits, we need better language than “just an illusion.” Because in complex systems, illusions can do real work.