The new AI leaps from language into logic.

"Natural language embedded programs (NLEPs) have been introduced to enhance the functionality of large language models. By generating Python code to address queries, NLEPs increase accuracy, efficiency, and transparency. This approach allows models to handle diverse tasks more effectively and could also benefit data privacy and smaller models. Credit: SciTechDaily.com" (ScitechDaily, Programmatic Breakthrough: AI’s Leap From Language to Logic To Solve Complex Problems)

How to make sure that the data that the system searches and uses is right? The system must use two sources. The problem with AI is how it finds the data from homepages. Well-sorted data that has a determinator before it is easy to handle. So if somebody asks the date of birth of some U.S. president, the AI must only search "date of birth". 

And then output the data after that word. The system might involve the package that involves the alternative for "date of birth", that is "birthdate". That kind of structure in databases makes large language models (LLM) very large. There are many ways to say one thing. And those alternatives must be determined for the AI, or it doesn't understand them. But when the AI must find data from the text, it has a problem. 

Because there are more alternatives in databases. That makes new LLMs more complex than previously. The new AIs are more effective and flexible than existing AIs. And they make fewer errors than the existing LLMs. The new powerful language model uses modular methods to solve problems. In that system, the programmers think. The system collects the answers from modules. Each module is like a package that involves some information. 

So when the user asks for the names of the U.S. presidents and years in office, the language model collects the information using modules U.S. presidents. And then it can search when they were in office. The LLM requires data from trusted sources. And one of the most difficult things in LLM is how it should select trusted sources. Then another thing that the LLM must do is to pick the necessary information from the questions. 

Of course, users can ask questions like: what are the names of the Donald Duck nephews? That kind of information might be interesting. But it doesn't bring a lot of new things for business or programming. The problem with existing LLMs is that they are not interactive. When those LLMs get a mission, they will just go to search for information. 

That thing makes the AI an interesting tool, but the interactive AI could be more accurate. Interactive AI discusses with its user. It will ask questions like where the computer program will go if some user wants to use it for the programming. The more information the system receives, the more accurately it can operate. 

When LLM makes things, like computer programs, it faces the same problems as humans. If the knowledge of the customer's system is incomplete, the AI makes incomplete code the same way as humans. Incomplete data about the requirements means that the product is incomplete. 

The limits of the LLM can seen in the question, which of the U.S. presidents elected after 1950 was born on a Wednesday? 

"For example, a large language model might be able to memorize and recite a list of recent U.S. presidents and their birthdays, but that same model could fail if asked the question: “Which U.S. presidents elected after 1950 were born on a Wednesday?” (The answer is Jimmy Carter.)" (ScitechDaily, Programmatic Breakthrough: AI’s Leap From Language to Logic To Solve Complex Problems) 

The thing, that makes the name of the weekday, October 1, 1924, so difficult for the AI is that the AI must use the two-stage model to make the answer. First, it must find Jimmy Carter's birthday. Then it must know where it should search for the name of the weekday. We know that the source is a calendar, but the AI might not know that thing. The AI must search the net to find data, where it can find names of the weekdays. And where it can connect those names with certain dates. 

In interactive AI models the AI could ask, where that data usually is? And then the user can say that the answer is in the calendar. We all know that that date is easy to find on the calendar. But that is not so natural thing for the AI. If programmers don't determine that thing for the system, the AI must try to find answers from the internet. 

And there the AI faces one of the most interesting things. The birthdates of the U.S. presidents are well documented and sorted. There is the word "date of birth" and then the AI can search for the next words and numbers. The AI must only recognize the months, weekdays, and numbers for making the solution, that it introduces for the user. Weekday's names are not often mentioned in data sources like web-based dictionary books. 

That requires that the user knows where that answer is. And in an interactive model, the AI can learn things. Most of the queries that the AI gets can be, that the people don't have time or they just don't want to search lots of homepages. 

The birth dates of the U.S. presidents are easy to find. But the names of those days are not listed. So we might know that Jimmy Carter was born on October 1, 1924. But for searching the name of that day, the system must go to search that data from the calendar. In the calendar, data is sorted differently. But the main problem is that the AI must know to find the name of that weekday from the calendar. 

https://scitechdaily.com/programmatic-breakthrough-ais-leap-from-language-to-logic-to-solve-complex-problems/