<div dir="auto"><div><br><br><div class="gmail_quote gmail_quote_container"><div dir="ltr" class="gmail_attr">On Fri, Mar 6, 2026, 8:21 PM spike jones via extropy-chat <<a href="mailto:extropy-chat@lists.extropy.org">extropy-chat@lists.extropy.org</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><br>
-----Original Message-----<br>
From: Keith Henson <<a href="mailto:hkeithhenson@gmail.com" target="_blank" rel="noreferrer">hkeithhenson@gmail.com</a>> <br>
...<br>
<br>
>...Spike, either I have a complete misunderstanding of LLM-type AI, or you do.<br>
<br>
>...There is no source code for any AI that I know about. There is training code with which an AI is trained on a vast corpus of text, but nothing a programmer would recognize as code. As far as I know, the inside of an AI is a mystery to all the companies. Keith<br>
<br>
Keith I am no expert on it. But my reasoning is that any computer must have a set of instructions on what to do before it does anything. I am sure of this: the military will not turn a mystery agent loose with control of any weapons. They must know exactly how the thing works before they will allow it to control anything.<br>
<br>
I think of it as somehow analogous to the autonomous drone target recognition system the Berkeley team is competing in. There is definite source code there, and it is trained to recognize and distinguish between targets, represented by manikins on the course. Last year's competition featured a fleeing felon, an injured hiker, a nude sunbather, a lost pet, etc. The drone had to figure out which is which, and do the right thing: no dropping a fragmentation grenade on the sunbather for instance. Those things definitely have code.<br>
<br>
LLMs must have some kind of source code, or it would do nothing, ja?<br></blockquote></div></div><div dir="auto"><br></div><div dir="auto">There is code that runs the model, which is fairly small and standard. But this contains none of the intelligence. All the intelligence and nodes of thinking etc. lies in the model.</div><div dir="auto"><br></div><div dir="auto">The model is a several hundred gigabyte file consisting of a hundreds of billions of parameters (think of a grid of floating point numbers in matrices). It is less like human readable code and more like knowing the raw connection strengths between various neurons in the neural network of the brain.</div><div dir="auto"><br></div><div dir="auto">There are however, methods to introspect at various levels what the model is doing. For example, by tracing what parts of the model activate given different inputs. But this is still a far cry from understanding the algorithms. It is perhaps more akin to doing a fMRI on a brain and seeing what parts of the brain light up during different tasks.</div><div dir="auto"><br></div><div dir="auto">Once machines get to having billions+ parts, they can never be fully comprehensible to our human minds.</div><div dir="auto"><br></div><div dir="auto">Jason </div><div dir="auto"><br></div><div dir="auto"><div class="gmail_quote gmail_quote_container"><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><br>
</blockquote></div></div></div>