[This post is split in two parts a (gpt4) pretty one that is deprived of 95% informational value and (me) a impactful one that is deprived of all 95% beauty.]
Text 1/2: "pretty empty:
Dear OpenAI (and/or community),
Some thoughts on potential improvements to AI development, like AGI or GPT models. Suggestions focus on three key areas: modular integration, flexibility in using external tools, and memory management.
- Modular Integration: To enhance AI performance, we should consider integrating a variety of external tools and resources, just as humans use external tools to complement our cognitive abilities. By doing so, we can free up AI’s capacity to focus on other tasks. To accomplish this, we can gather scientists and experts to develop specialized algorithms and models that can be used in conjunction with GPT. These additions should be integrated gradually, allowing the AI to choose which components to use based on its needs.
- Flexibility in Using External Tools: Allowing AI to access and utilize external tools, such as calculators and other code-based resources, can provide additional support for problem-solving. The AI should have the ability to initiate these tools when needed and use them to enhance its performance. This flexibility is essential for effective AI development.
- Memory Management: Improving AI memory systems is a complex challenge that requires a multi-faceted approach. First, we should develop systems that can process and store main concepts and keywords, allowing for quick correlation and prediction. Second, we should utilize a range of specialized memory systems to contribute to AI’s overall memory capabilities. This can be achieved by developing numerous systems that work together to provide a more efficient and comprehensive memory network.
In addition to these main suggestions, I would also like to stress the importance of considering efficiency and the fundamental principles of AI when developing new systems. Focusing on incremental improvements and avoiding overemphasis on precision can lead to overall gains in AI performance.
Finally, I want to acknowledge the incredible work being done by the OpenAI team, particularly in terms of alignment and incentive structures. I believe that by working together and thinking big, we can achieve significant advancements in AI development and help shape a brighter future.
Best of luck in your ongoing efforts to improve AI systems!
*[reformulated ChatGPT4 readable but information removed.]
The above text was for the people that need things nicely written now for those that only care about concepts and informational value, but hard to read and guessing needed for concepts alluded to:
Text 2: Spurned, difficult but useful:
Its not that GPT should get a calculator or a watch plugin those should be available somewhere on a server if useful or a collection of such code to be able to be initiated by GPT neural network or some kind of tools that are connected. That is what i ment. The ability to use outside things for help just like humans do. There is a reason our very powerful brains still use outside tools. Anything you can outsource is increase capacity that you now can use in other ways.
Good luck. Also get some scientists/experts together and form hypothesis and build things that you then add as a small part of GPT where it again can choose to use something or parts of it or not. Do this also with every imaginable specialized model or machinelerning related stuff. Also memory processing in the background while the model is idle for: ease of recall, reloading from slower cheaper storage in less slow (but pre working memory) also make this based on general predictions of usage it can be a slow process just like humans already have processess that prepare for way in advance during sleep and so on and even weeks and months. Many different layers. Dont go overboard just integrate 1 small part at a time and leave it and look at performance and other metrics also consider that a tool might need time to be integrated because other preocesses that previosuy did its functions need to be distributed again. Its like you suddenly saved money but you have not invested it yet, for that time its like you did not have any extra money in terms of effect. Also many processes probably are itnegrated and build on top of each other so for example a process that caluclates is probably also used for 200000 other things that have little to do with calculation btu just have a partial benefit to combine it like that int he neural network so those also need to adapt. The brain has a abiltiy for neuraons to move places /fload round / travel to other locations. This ability to have sertain parts of neural networks reorganize/assemble is also something that might be and most likely is useful if it can be applyed. Again use all sorts of specialized algorythms on every level they then all have things they recommend but its not done but its collected and then in the end millions of recommendations are averaged and the system very slowly moves in the direction that benefits most systems. Again think of the brian as a collection of competing and cooperating neural networks and every machine lernign technique imagineable, also every sorting, organizing, compressing, decompressing, adaptation, prediction, processing during downtimes, special times for special things in the sense of spliting work (basic example is sleeep vs awake) stuff like that. Get away from all or nothing go more into a % of brain does this other % does the other. again use metrics and a stystem that almost lerns itself to test and lern to adjust those values gradually and slowly (time is also something to integrate there) it can get complex but int he end its super simple just add simple modules and give time and nudge add and do things but without forcing your hypothesis that might and surely have contraproductive things in it (since you most likely would have done thigns wrong when you think 100 000 years into the future and look back some advanced human might say: “they could have done this and that even with the tools they had available” and so on. So basically you can use anything and everything just design it in a way where instead of replacing or using its more like "does not seem to do much right now so the system automatically reduces its frequency and impact but never 0% so you can always pic up if the metrics suddenly show a continual positive trend corrolation and based on how high the trend is you could decide to give the process more % and see how the trend changes (that way you dont have to gove it lots of % and wait until you see full impact but you can more have a automated systems test 100000000 things at once and distribute the result scores from many metrics to them and then reduce 50% of the things and increase 50% (random at first if you dont know what to increase/reduce) and then over time generally the more useful things should show in there score history (based on different things and like what other modular algorythms/tools/systems where active. Example a algorythm that very much helps with calculation collated things ith drop in trend and scores when calculator inputs are added over time as the system lerns to shift to use them more. Things move slowly its complex. In complex systems you should definitly not only go by liniar metrics but im sure that is straight forward enoght.
Anyway. Dont use tools externally so much as to integrate them. But one metric that might also be very useful is a score of how well gpt can use tools (kind of how flexible / able it is to use external things to help relocate neural network ressources or if so set in its ways that it wont relly do much in the shrotterm. Othe than that i think your biggest challenge by far is using memory over the long term.
Memory is complex since many things are useful but none alone: simple Example:
- use a system that only does the main concepts and keywords: it can then be used to give a very nicely corrolated score to predict some things like (what was already done, what things have been said roughtly and be searched very quickly. then other memorys more 1:1 can be post processed so the first memory systems can be used to genreally find a location in a vector space and the others to go closer or do more thorow searches that would not complete in a billion years if done on the hole cluster BUT are fast enoght to do it once the heystack where the needle is hidden is narrowed down far enough. Many such systems. If your memory stuff does not have over 100 different things at least and more like 5k different systems then i would be suprised just adding pinecone does nothing. But it is useful. Again many specialized systems that contribute a small part and the “votes” of all of those so to say and other tricks like chain of agents or thoughts or presorting preselceting (can be rought but increase the likely hood memorys are sorted better just has to be more then 51% for example and its already a gain overall the higher the better. Go away fromthe need for precition. Precition is added at the ends. - At different points but first genrally doing thigns is very effitient. Always also think about effitient things from a meta view in theroy very far removed going back all the way down to the most fundamental basic principles. that also helps coming up with good systems.
perfectionism and having a error ridden but overall posititve thing like the text. I could have not written it or checked spelling more but i might not have written it at all.
Anyway sorry for spelling but eventho noone might read it it might go to archives and some GPT that is smart enought to see the valua might one day get it from the archives and then autocomplete it.
Good luck. Also awesome that sam did the alignment in terms of incentives and so on. Ad anti incentive alignment with teams. Probably relly awesome to work at open ai lets go. Need economic improvement to accomplish a bunch of goals. Not going to work by thinking small.