> Current AI does not "understand" the information it holds, nor does it have a concept of "truth".
So it's like a CEO. Good to know. On Thu, Feb 22, 2024 at 6:13 PM Joel C. Ewing <jce.ebe...@cox.net> wrote: > One needs to understand that today's Large Language Model AI tools like > ChatGPT, etc. are essentially driven by huge statistical databases > created from processing huge volumes of digital text using some > knowledge of sentence structure and words. Those tools can accept > English-language queries and use words and phrases in the query to > generate related-information responses with complete sentences and > paragraphs that have a high probability of of being relevant to the query. > > Current AI does not "understand" the information it holds, nor does it > have a concept of "truth". Even if you program the AI using books on > COBOL grammar and semantics it won't "understand" COBOL. Even if you > feed it millions of lines of COBOL code it won't be able to deduce the > underlying purpose of the code. If there is an accurate description of > what the code does accompanying the code, it can associate that > description with a code segment; but if the description is inaccurate, > AI may also associate the code with that bad description. Inevitably > some of the code you might use to program the AI tool will contain bugs, > and AI will be equally content to supply buggy code examples. > > If your object is to generate optimized Assembler code which accurately > replicates the behavior of a COBOL program, the best tool for that for > the foreseeable future is an optimizing COBOL compiler for your target > machine. Such compilers are already doing flow analysis just to > optimize loops and register usage, but I wouldn't call that "AI" in the > usual sense of that term. Perhaps a well-programmed AI tool would > suggest using a COBOL compiler if asked to convert a COBOL program to > assembler -- in fact that is basically the response given by the MS > Copilot tool when asked to perform that task for z-architecture; > although you can see hints of its lack of understanding in that in > includes in its response "IBM provides cataloged procedures (such > as*IEBCOMPR*and*IEBCOPY*) to simplify JCL coding for COBOL compilation", > where it includes gratuitous PROC examples that have nothing to do with > COBOL rather than giving the names of actual COBOL compiler PROCs. > > Joel C. Ewing > > On 2/22/24 11:09, Robley Lutz wrote: > > I guess my question is, do we expect AI to look at COBOL code, and not > > simply compile it, but analyze the flow, and output optimized Assembler > > code? Will AI become the highly skilled Assembler programmer that I > never > > became? > > > > On Thu, Feb 22, 2024 at 11:54 AM Tom Harper < > > 000005bfa0e23abd-dmarc-requ...@listserv.ua.edu> wrote: > > > >> Dave, > >> > >> I was told the same thing 54 years ago when I starting working at > >> CalTrans. Managers would just be able to code in COBOL PROFITS = SALES - > >> EXPENSES and we would all be out of a job. > >> > >> ... > > -- > Joel C. Ewing > > ---------------------------------------------------------------------- > For IBM-MAIN subscribe / signoff / archive access instructions, > send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN > -- zMan -- "I've got a mainframe and I'm not afraid to use it" ---------------------------------------------------------------------- For IBM-MAIN subscribe / signoff / archive access instructions, send email to lists...@listserv.ua.edu with the message: INFO IBM-MAIN
