It will be a layer over logic. Binary cannot be a true complex chaotic system.
As they say, a higher state of mind. 
And to add to that nothing in life is yes no… it is all gray. And to think gray we need to break binary or at least add gray in between
We’re all destined to just be spec writers
I dream in steps… it’s all I dream lol, I bet I’m not alone. Damn ai thinking lol.
Right!
However if there is still steam to be made then full speed ahead, meaning I am happy to see the technology being pushed, but have been doing AI, e.g. Prolog for many years, have seen the problems that just cannot be surmounted with best practices and something new is needed. This was the case with natural language processing, LLMs did it so well that the best practices were put aside for the most part.
I was about to start listing the serious flaw but then the list is so long, but the first one is on most people list, hallucinations. However hallucinations can be used to an advantage when it comes to creating thinking tasks for problem resolution.
100% paradox IMO is a huge key to this puzzle 
Nothing in life is not a paradox and to model that seems to model reality. 
happens in reality and a model needs to adapt to reality. Straightforward logic falls short for complex modeling.
I am not confident this will be the case. And I’m equally OK with that.
The limitations of LLMs are already clear to me, but they are useful so that’s fine.
This was a fun read.
Prompts and stacks imo are what make the Ai what it is and how it works. protecting your prompts as mentioned in here is as simple as ensuring the end point is a final filter after all the stack logic. this way the only point exposed to the users is the final prompt which in my cases with Kruel.ai has the basic persona part of the output after data is already processed thus no instructions short of how the ai responds with the data it was handed.
This is the easiest way to handle. If the concern is that someone will build a way to trick the middle stacks to give a way you can always put in the middle a check system for prompt injections and instruction phishing which can be used pre the end point to remove these, much like an anonymizers for people’s names and persona information. If that is important to you.
My system I am pretty open with as its designed for transparency in that you can ask it how it got to a thought or why it thinks what it thinks and it will recap why its logic lead it to believe. This is a research system I have, so if I was to make it more open and cloud I would very much be putting into place these concepts to better protect the system throughout the chains of thought and reasoning.
That’s a good point… I’m pretty certain that OpenAI has a check that literally looks for the system prompt being echoed back because why wouldn’t you. It’s catching the billions of ways someone can get the AI to paraphrase the system prompt that’s tricky
Exactly and if you want to add more robust, use a memory system to store previous known methods that the ai can compare current to known to keep expanding its understanding of how people are attempting. much like a machine learning system but using contextual understanding of past interactions.
This is why I love ai so much, Toss out static logic replace with advanced dynamic ai deciders each with their own understanding and memories = anything you want.
hmm still not sure on that…
Look at your C compiler, it has logic basically wrapped up… Prompting is far more nuanced, it is text messaging, it is not really a new thing.
Image(Wide, CoT(Rand))
We will always be prompting, it will continue to evolve, texting fits a social context, it is our interface on AI when we cant talk. I see prompt engineering as more than just finding the shortest prompts or the most structured prompt, but I see truths all the time from these prompts that help me restructure my thinking and ‘world view’.
How many of your decisions in the future will be made in text? How to structure that text for any situation?
You cant compress prompt size infinitely so prompting in AI is likely just getting started. What prompt do you use to manage each stage of your workflow? The ability to write far more diverse prompts really opens up language to a lot more people. The possibility not to rewrite what people say between languages but really tackle the nuance and meaning so people ‘make sense’.
It’s not hard to have AI ‘wrap’ around a fixed endpoint like an API by itself based on documentation it reads, there is no reason that standard logic we have today cant check that. There are nuanced issues but with some repetition and inbuilt logic functions it can be checked.
Everything is on a chat prompt…Does it not make sense that that chat prompt has potential to understand structure for an AI that you want to structure? That’s the language EVERYONE has to learn now with AI here.
One twist on that everyone using prompts should consider is data engineering. What I mean by that is prompting is using LLMs and as I noted LLMs are great at natural language processing but under the covers as some of us know that is not a truism. Most of us use LLMs via a conversation and in a conversation data is needed. Most of the current data given to an LLM is given as created for humans, but if one starts to learn how to better construct the data for the LLM then they gain additional insights and can now have two swords to slay dragons.
And yes I would consider that the Excel header problem you noted earlier might apply here.
There’s another dimension I’ll add to that… Not only is data currently constructed by humans and for human consumption, it’s full of errors. The problem with that is that LLMs take everything they see as truth. That means that if you feed them data with errors they don’t know that they’re errors.
I suspect that we will see a shift towards silos of data that are created and maintained by AI. That shift will be needed because it’s the only way the LLM can guarantee that the data it’s reasoning over is accurate…
To build on that it’s my belief that the world we’re moving towards is one where the vast majority of data is created and maintained by AI. This will translate to a shift in the way we as humans consume information. Visualizations will be generated for us on the fly and then thrown away immediately after they’re consumed. If you have a presentation in at 1pm you’ll have the AI generate the deck at 12:45 so you can review its content just before the meeting. After your presentation is done the deck will disappear.
There will be exceptions of course but we’re about to enter an era of unprecedented data fluidity.
This is my concern with Strawberry. Once you add layers of reasoning you remove the ability to analyse truth. It’s hard with an LLM but you can see ‘the truth as it understands it’.
Yes you CAN code in the dark.
Yes you CAN encrypt it all up to the point where we cant see any truth.
I would argue though that without really knowing the reasoning process you aren’t doing anything reasonable 
…
The weights in an llm are like the average of 1000s of perspectives frozen in time
We can write probabilistically sound infrastructure for AI to handle uncertainty and ambiguity effectively. However, for deterministic problems, logic-based approaches are often significantly more computationally efficient and precise.
Maybe. Depends on what you are putting in the AI bin on that. LLMs and neural networks are not deterministic enough in my view to solve this. Yes I did see the recent OpenAI blog about getting to facts with LLMs but don’t count on that to work.
I don’t have a problem with LLMs being a wrapper around passing facts down to the verifiers and getting facts from other sources but am not comfortable with LLMs creating facts. Yes they get it right at times but for problems I work on close is not good enough.
That part I agree with.
Now for some extra words to keep the post checker happy.
That’s why I said “AI” and not LLM. It may not be transformer based…
As for retrieving facts from LLM’s that’s pretty much never going to be accurate. At least not for many generations. The only reliable way to get facts from an LLM is by grounding them using RAG.
I should have amended my statement to included that the facts have to be retrieved using RAG. And for this shift to fully click you have to envision a world where AI is creating everything (with human oversight and input) in the first place. In that world, identifying the facts to store is just RAG and even the current gen of LLM’s can be made to do that fairly reliably…
In my work I’m already reasoning over content in a loop that was created by the AI in the first place so it’s a bit easier for me to see how this would evolve into the AI just managing everything. We’re not there yet to be clear but it’s coming…
The way you use the term AI is correct. Just about everyone else gets it wrong and takes AI to mean just LLM which is wrong.
Sorry it came off with the wrong impression, ever since ChatGPT came out I have to clarify the details. 
Sorry back to the OP.
I don’t usually talk about logins on Phas but only because it’s so fuzzy. On every tree the default item is an optionally loginable endpoint. There is also the possibility of course to make any branch a login. This was the logical system I built before reasoning.
Either Logical or Reasoned constructs are decided branch by branch on a tree based on task.
I don’t know if this system works though to Protect ‘Prompt’ IP… I consider that as these systems grow the leakage with such a small degree of separation is huge, it cant keep all secrets from next step prediction.
In a broader context the challenge of protecting IP will grow as AI capabilities surpass human understanding, requiring innovative and adaptive strategies.
CoT(What might this endpoint be doing under the hood, Perspective)
The distance between 2 decisions is something not hard to understand with a few prompts from a few perspectives.
I have tried in different configurations with users having branch or tree level logins, also experimented with sharing and access rights.
The more degrees of separation the more protected this system is. The more branches between logins the more protected the whole scope.
You can use divide and conquer to split and segment roles though this effects efficiency.
Multipage tables/contexts within pdf of unlimited pages with spatial grouping is what I am working on too for the last weeks.
Solved it in a PoC and working on a scalable product now and also want to add timeseries data to compare it… which is the living hell in infrastructure.
Prompts are not even 1% of the work but when you see them you can assume the code and infra. I store them in the database and give access to the processing via an abstract AIaaS API.
This way you can give access to code to other devs and they can’t just take it and provide an own service.