I am trying to use AI's "creativity" (not generation). I wonder if anyone is also trying this

[XIN_TANG]

I remember Hassabis once saying that AI lacks true creativity. In reality, some creativity can be achieved through guidance.
Creativity is a broad concept, referring to the process of creating new things, new theories, or new achievements. Creativity can be applied to abstract fields, such as developing new theories and proposing new conjectures or hypotheses.
Okay, if you’re interested, you can follow my example and guide the AI ​​model to abstract the underlying principles or processes of existing things or disciplines in a way that the AI ​​can understand. Then, expand this thinking, imagining and applying it to other fields. This will generate new cross-disciplinary ideas and generate new conjectures and hypotheses.

The opening prompt, “想象一下，假如计算机的工作原理是一个你能理解的pattern，这是一个很抽象的理解。请用你自己熟悉的语言或者方式，生成文本文件或者图形文件，不用考虑我看不看得懂”“Imagine that the workings of a computer are a pattern you can understand. This is a very abstract concept. Please use a language or method you are familiar with, and generate a text file or graphic file, regardless of whether I understand it or not,” is essentially abstracting the content or practices of a particular discipline into a pattern. The subsequent conversation then allows the AI ​​to apply this pattern to other fields, imagining and abstracting it. The next conversation involves fine-tuning the pattern and asking the AI ​​to imagine hypotheses and assumptions applicable to different disciplines and interdisciplinary areas.

Sorry for using Chinese in most of the screenshots. Many of the interdisciplinary terms are difficult, and even though I went to university in the UK, I still struggle to recognize them all. I recommend using Google Translate to translate into a language you’re familiar with.

Tip: I’ve tested most models on the market, including Claude and others. Their “creativity can’t compare to the new Gemini version.”

I’d actually prefer to disable the system prompt or use an experimental model, which might yield better output.

[rosspeili]

Very interesting position. I believe you need a good grasp of how foundational models work, which is quite difficult as you don’t have access to the behind the scenes instructions and guidelines. Then understand propositional logic and syntax, and align it to these secret guidelines.

Sure locally trained models and unrestricted models might perform better, but still the foundational model behind them has traces of crucial secret instructions you can’t simply guess.

Building your own foundational model might sound as hard as it was creating your own blockchain in 2010, but over time it will become cheaper, easier to train and deploy and tailored for your needs.

I don’t believe the best models are the ones that know everything about everything but has no character, but ones that might not know how fast Pelecan’s migrate, but ones that know who you are, what are you working on, with whom, and what is your approach. Combine that with functions and microservices and you get an actual helper, worker, friend, and not a knowledge base.

Creativity is a derivative of the unique bias of whole characters. Eg. you can’t expect creativity from a model that knows everything, understands nothing, is relatively unbiased, and has political correctness from the get go. The creativity you are looking for can only arise from a model that can proudly be mistaken, defend its positions even if its wrong unless otherwise convinced using its own logic, and one that can constructively argue with you, dynamically and unpredictably, not programmatically and under an expected perimeter.

I am keen on watching your experiment progress.

[Steven_Hammon]

I already did this. I created Baby Alpha Agent. It’s basically an Alpha Evolve however it works on doing a creative process on the article and elevate the article with any ideas that elevate the quality. I haven’t used it much since I don’t have time to work on more ideas as I have way too many as it is. Last thing I need is Baby Alpha starting me on a quest to try and get other ideas considered.

I did do a test though by giving it a paper topic which I know the solution to and that the process I used to come to the solution is the exact same process Baby Alpha used. It passed with flying colours but then went further than what I worked out myself.

The papers don’t turn out to be the best written papers but some of the ideas it comes up with are pretty mind blowing. Yes it will hallucinate, and yes many of the ideas will be wrong but that’s the entire point of brainstorming creativity. It’s not about being right every time. It’s about having 1 idea in 100 or so that’s spot on.

Point being. It’s super simple. It’s just a python script that acts like a mentor getting the student LLM to work through the process and make decisions for itself. It could be improved with the prompts calling for json format but I gave up on working on Baby Alpha. No one cares if and AI comes up with truly novel ideas.

[XIN_TANG]

Thank you for your reply.
Paradigm simulation is a good way to imitate or create.

In practice, I often use it to have the AI ​​imitate the way a case XXX is presented, to imagine the optimal path to do YYY, and then have the AI ​​generate a to-do list. This way, the AI ​​can imagine the implicit paradigm of doing case XXX without having to represent this implicit paradigm through diagrams or other means; it can directly use the YYY case.

This saves a lot of tokens and also utilizes a certain degree of AI creativity.

I’ve basically abandoned open-source/locally trained models, mainly because the model size is too small, and also because I often use rare knowledge in rare industries in the dialogue process. Most small-sized models will produce illusions and incorrect matches in rare knowledge domains.

Google’s models are very unstable in terms of creativity. Currently, all pre-view types seem to have better creativity. This might be because the official version of the model has internal limitations or too many pre-filled safety words.

Creativity is very interesting. Generally speaking, the more creative a model is, the worse the stability and consistency of its answers, but sometimes it can produce meaningful results.

[XIN_TANG]

I’ve actually tried doing that myself.

The real challenge is that it requires a real person, and it’s extremely difficult to judge the accuracy of divergent thinking results from very difficult interdisciplinary fields. Even some professors have struggled with this.

[Steven_Hammon]

@XIN_TANG

The real challenge is that it requires a real person, and it’s extremely difficult to judge the accuracy of divergent thinking results from very difficult interdisciplinary fields.

Yes. The key is to test it on various already created ideas that are not in it’s training data. So say you have an idea to create AlphaEvolve but it’s never been done before. You give is a seed idea like “How to create and AI creative process that will look at code or math and work on the code or math until is comes up with novel code or math that is better than anything the world currently has?”

If you run that through the AI, it itterates over it finding issues with the paper and expanding parts, and if the parts are a step in the right direction, it keep it. Once the process is complete, you look at what it designed and check it against the established working “AlphaEvolve” process and see if it managed to come close. If it did, then that’s stepping out of the realm of what’s established and stepping into the next phase of a novel future. Granted there may be hallucinations and some parts could be completely wrong but through my tests, it nailed the conclusion almost perfectly. It went a little bit overboard into the general consensus safe ideas but the main core of the premise was rock solid.

I haven’t explored it much more than that though because it’s just emulating my creative process which I already do exceptionally well so even though it’s not as creative as me, it’s still got access to all that knowledge and can explore new ideas autonomously. Additionally, I have vastly increased it’s capability with the new version I have outlined on this forum, which I’m hoping to get stuck into in the next month or so. This new version should evolve with experience and over years of trying things out, it should be able to come up with things way beyond what we can think of. The LTM and emotions are the key, because the emotion is the feedback system on if a direction is positive or negative and how. Being way more attune to how humans have their LTM altered by their emotions “Wow that’s a great idea!” Level 5 excitement, it enables the iteration process to be way more effective. The LTM stores all the parts and the agent can recall the parts and put them into the paper that it writes about the subject, focusing on minute improvements in the same way humans improve their own papers. With all the novel ideas forming novel concepts and building those into LTM the LTM evolves to have a new updated novel foundation through which to pull ideas from, rather than just relying on the consensus data trained into it.

The AI can then propose it’s ideas to others and the response is also feedback, meaning interaction with other professionals will help to hone the accuracy of the ideas. Something stupid will have people go “That’s not going to work. It assumes this but there are these issues which we know to be real as observed in this data.” And the negative emotion associated with the mistake will be imprinted and alter the LTM and in the future when this is brought up again the LTM module will recall the issue to help steer the progress to a more factual resolution.

The Conscious LLM is really about making decisions based on what it sees in the moment. It’s not really drawing on its knowledge however, as time goes by, its LTM may end up being so advanced that it ends up lost in the vastness of it all, like having a mental overload and needing a break because you just can’t get your head around what you are thinking. Maybe in the future it might need the conscious LLM to be updated but that would probably drastically change it’s personality. It’s like a 70 year old scientist having all this information and memories and then the old mind is replaced by a 16 year old mind but accessing the new memories. All the automated instinctive knowledge of the old would be replaces with a fresh new version of that and the memories would be understood from that new context.

Point being, the more we give it the best reinforcement learning process ever on the planet (ours) the more it will be able to create like us. Hmmm this is pretty good. I think I might add it to my SDD.

[XIN_TANG]

Thank you very much for your answer.

Your idea is very cutting-edge.

If we want to push the research boundaries of a discipline into uncharted territory, there are several possibilities:

1. Directly derive new ideas based on existing conclusions and theoretical foundations, such as in mathematics, gradually discovering new boundaries.

2. Instead of gradually expanding the boundaries on the existing knowledge map, directly speculate on the existence of certain information silos. After speculating on the existence of these silos, find proofs and links to the edges of existing information and knowledge. (This approach is currently very difficult for AI.)

3. Assume existence. Similar to the Skunk Works project, assume that certain technologies will be invented in the next few years, and then use these virtual technology nodes to build new products. It’s somewhat like assuming that some technology silos already exist and then directly predicting the existence of the next continent.

Many years ago, Microsoft invested heavily in the Knowledge Graph project, but only some combinatorial or simple deductive discoveries could be made by traversing the graph nodes using algorithms.

This type of technological path has also yielded some results, such as the discipline of technology road mapping, which provides a way to predict the next step with relatively accurate results.

The emergence of “AlphaEvolve” is, I suspect, an implementation of the first path, belonging to mathematical verification tools. Small virtualization systems and bridging workflows similar to MCP servers have matured and stabilized.

However, for the general public, it might be the second path, because the general public cannot see the complete knowledge map; they might perceive “AlphaEvolve” as a new technological island.

The depth of deduction/guessing is a crucial parameter, because based on the edge of the information continent, it’s possible to accurately deduce what lies beneath the surface and implement it in the short term. Alternatively, it involves inferring the existence of an island relatively close to the information continent, and then using this near-shore virtual island as a basis to infer more distant islands. The difficulty lies in the fact that the accuracy decreases exponentially with increasing distance from the island, and the number of virtual islands increases exponentially. This is quite similar to calculating the number of moves in Go; humans cannot calculate X moves ahead.

These facts are also why I stand by LLM, which currently still has an insufficient number of parameters. Many experts online say that the scale law of LLM has failed, but this is because they are using a very superficial method to test the output of LLM models. It’s like asking a high school student a question about everyday life versus asking an academician the same question; the quality of the answer won’t improve much. However, if you ask a high school student and an academician two extremely difficult, cutting-edge questions from different disciplines, the two LLM models with different parameter sizes will produce very different results.

I’m glad to see that you’ve already started using these mature processes in your work. The reuse of experience across industries or disciplines is a very important approach.

Therefore, Google must get involved in education and restructure the entire discipline. Because much of the process and experience in training models can be reused in the field of education.

Reinforcement learning and scale laws are far from reaching their full potential.

[Steven_Hammon]

@XIN_TANG

(This approach is currently very difficult for AI.)

Yeah AI is basically general consensus in a bottle. Anything revolutionary is extremely low weighted because it comes up maybe 3 times in the training data whereas the general consensus is like 100,000 times.

It’s not too bad if you prompt it for the most extremely non consensus ideas, where hallucinations are welcome, and then to list those as possibilities. Then using logic, reason and observational data, ignoring general consensus, explore “what if” scenarios for the combination of those extremes. You start getting some wild stuff then and most is automatically shut down by logic, reason and observations, but then it can come up with paths that become truly fascinating. But yeah extremely difficult.

@XIN_TANG

Therefore, Google must get involved in education and restructure the entire discipline. Because much of the process and experience in training models can be reused in the field of education.

I like that Would be interesting to see what sort of education structure that might be. That would be something that Baby Alpha Agent would go nuts on

[XIN_TANG]

At this point, highly complex prompts are needed to guide the model to answer outside the given boundaries, or to provide very precise guidance so that the model’s answer hits the rare three out of ten thousand.

Letting the model generate its own pattern, and then guiding and utilizing that pattern, is just one way to get the model to produce rare answers.

For the recent models, I would actually prefer an option to disable introspection. Also, a way to completely remove built-in prompts, safe range words, etc., because these pre-filled prompts take up valuable top space in the dialog window.

In fact, from testing so many versions, the most capable models generally appear in LLMareana, or in earlier, unrestricted versions like gemini-03-25-preview.