Week 21

The missing piece for AGI

AGI is getting closer, but there are still one piece missing.
The big context window seems to be solved. Especially the Mamba architecture seems to be very promising. According to wikipedia it has an inferience speed of O(1) and training speed of O(n), whatever that means.



But the big hurdle is still hallucination and wrong logical reasoning. 
If we can get that fixed, so we can always trust that anything that is in the context window is included in the reasoning, we will have achieved AGI.
 
The easy way to do this is the ability for an LLM to return "I don't know". It can be done by training the LLM normally, but with additional training data in a question/answer form (that an agentic AI will be using for AGI). 
After training the LLM is tested against the Q/A, and is the retrained with all the answers it got wrong. An AI is used to validate the result - and since it knows the anwers from the training set, it has a high accuracy of figuring out if the result was indeed correct.
 
Then the LLM is retrained, and for all the answers it gets wrong it returns a special result "I don't know".

When we have an LLM that can say "I don't know", the agentic AI will make sure to simplify questions until it can actually be answered. Which means that the thought process will have to be divided into smaller reasoning pieces. 
If the reasoning capabilities is above some threshold, AGI have been achieved.

So to summarize
  • AGI is close, and it is only the ability for the LLM to return "I don't know", that is missing.

LLM Performance with tree structured neural network (death to CoreLLM?)

As mentioned above, the Mamba architecture should give a big improvement in reasoning. Currently another idea has been floating called Kolmogorav-Arnold networks (KAN). 
In the paper, it is speculated that KAN can help with interpretation of result. I don't believe in that idea, since it is still a very complex neural network. I you want to achieve interpretation it should be done in the agentic layer.
BUT the same feature can also be explained as a sparse network. In KAN it seems that the network find a local optimum, where some neurons go to zero.
This means that the network can grow, since you can start with one shape of the network, train it, and remove unused neurons, and add/enlarge neurons where it can benefit. The most important benefit of this that when doing inference you don't have to go through all neurons, but only those that will influence the final result.
This means that no matter how much data you train on, only the relevant parts will be visited when doing inference. If this works in practive it can change the AI landscape somewhat.
For one thing, only big computers can do inference, limiting the private citicens in running the AI. The performance per user is still minimal, so the cost for an individual is still cheap, but it has to be run centrally.
If this idea works at all, remains to be seen - and it is not really a showstopper for AGI. The alternative is CoreLLM concept that have explained earlier, where you limit the training data to get a faster neural network.

Fundamental democracy revisited

It's becoming increasingly urgent to develop AGI that can be controlled by individuals. Russia's actions in Ukraine are causing widespread humanitarian suffering and global instability. Meanwhile, China continues to restrict individual freedoms and strengthen top-down control, aligning itself with other authoritarian regimes like Russia.
 
First an AGI that is cheap enough to be able to run by 8 billion people needs to be made. After that it needs to be distributed. So marketing is a big part of that.
Both OpenAI and Google has just released AI's that can listen, see and talk. If they can get localized, it can lower the barrier for entry a lot. And will also seem a lot more trustworthy. The missing piece is life video generation, so the user actually can see the AI, giving even higher trust and making it even easier to use.

The trust is of course deceiving, to the most important part is to make it easy to understand how to validate that the AI actually is working on behalf of the indididual.
  • It should be a proponent of freedom of the press, or at least some way to be able to make informed decisions
  • It should be a proponent of fundamental democracy, or at least some way to make global compromises
  • It should use a trusted platform, and explain how it can be validated. 
  • It should be clear that the AI is working on behalf of the user, and how to validate that

This is just some random thoughts, and how to properly explain it and make marketing material is a big and iterative task.

But the end goal is that an indidivual will never say no to using AI, since the promise is that it will always try to do what is in your best interest.

I asked chat GPT, and it came up with the following - which seems to be in the right direction:

Individual Empowerment

Take control of your own information and decisions. This AI puts power directly into your hands, enhancing your personal freedom and independence.

Unbiased Truth

Access clear, transparent insights. Our AI ensures you understand how it operates and consistently acts in your best interest.

Universal Rights

Support your fundamental rights. This AI champions press freedom and democratic values, making it clear that any restrictions reflect governance issues, not the technology.

Kommentarer

Send en kommentar

Populære opslag fra denne blog

Older blog posts

These where originally posted on https://omnia.dk/hard_ai Hard AI Lean Fuglsang, 2023-12-31 Current LLMs like ChatGPT are very impressive, but are also limited when giving complex task. So it can not do the following: "Give investment advice" "Write a research paper" "Improve the AI located in /home/me/HardAI/CurrentBest" But what if we can super charge the LLM to do the above tasks? The idea is that to do a task, we don't do a single request to the LLM, but maybe thousands or millions. We use the LLM input as working memory - all non-pretrained information that is needed for the request. And instead of given the final respone, the LLM will return the next action that should be taken. The result of the action is then summarized into the work memory - again by using the LLM. To enhance the performance, we can add an additional validation step that mitigates the problems of the LLM halucinating, being imprecise or have wrong alignment. An...
Læs mere

Week 7

Google has released the Gemini LLM - which has a context window of 1 million tokens. Google blog They also write that it can analyse it all with what they call a 'needle in a haystack' test. And more importantly it can do 'in context' learning, and actually using the information in the context window for complex tasks. This means that the concept of a CoreLLM is much closer to being useful. It looks like the context window just keep growing, so there is a balance how small the CoreLLM should be, and how much should be put in the context window. This also means the AI singularity is getting closer - the day when the AI improves the code it is built of. Even if the AI improves it self, it is not guaranteed that it can improve itself in all areas. But in that case it is important to remember that it can easily be nudged by humans. If you have an idea how to improve the algorithm that the AI can not figure out it self you can just put that information in the short term or l...
Læs mere

Week 12 - part 2

I saw this tweet by Yan Lecun that LLM can not think abstract in a visual way https://twitter.com/ylecun/status/1769778223842226286  What he fails to realize, or maybe exactly what he realised is that you should not make an LLM answer a question directly. Instead it shuold always be wrapper: "What is the strategy for answering this prompt: {X}, having these tools at your disposal". And then the list of actions that can be done. Think of it if you got some liquid and was asked what it was. It would be dangourus to just taste it, and give an answer. Instead we will analyze it - find the pH value and what other tools we can use to find out what it is. And then take that knowledge, and give a result. So the job of the LLM is to do this analysis and continuously improve the prompt (short term memory) with more and better information. And of course also improve the state for the tools available (long term memory, the programmed world model). The prompt can both contain text, images...
Læs mere