Re: unboxed structure fields (or class slots)

The 1-bit solution…  There is a parallel in signal processing. At one time, SONY was offering DSD recording formats, which were 1-bit at 5 MHz sample rate. You can simplify anything down to fewer bits as long as you compensate with bandwidth.



> On Jan 13, 2026, at 10:15, David McClain <[email protected]> wrote:
> 
> As for numeric capabilities for efficient PN’s.. I believe the Chinese have shown that you only need 1 bit of precision for many of the LLM tasks of today. So numeric capabilities are hardly limiting factors now. 
> 
> It is true that vectorized multiply-accumulate would be helpful. But I would also like that very much for my daily signal and image processing chores. 
> 
> CL was standardized just prior to the onset of multi-core CPU architectures, and while DSP was a highly specialized side effort. Fast forward by 30-40 years and yes, much has changed. 
> 
> I still find CL the most capable “modeling clay” for experimenting with many facets of computing. This is driven largely by it having both Lexical and Dynamic Binding, very capable functional closures, and supremely simple and consistent syntax coupled with a very capable macro system.
> 
>> On Jan 13, 2026, at 10:07, David McClain (as dbm at refined-audiometrics dot com) <[email protected]> wrote:
>> 
>> I do think it is a mistake to assume that perceptron networks are and will be the way of AI into the future. These networks are the current best that we have for understanding hyperspace partitions in data that we don’t fully comprehend as humans. But that isn’t to say that PN’s are the best and always will be.
>> 
>> PN’s have been used successfully as recognizers and as short-range predictors of chaotic attractor basin trajectories. But they are incapable of synthesizing new knowledge.
>> 
>> We used them quite usefully to run our “Rubber Mirror” on the KPNO Transit Telescope, for removal of atmospheric turbulence speckle from star images. And we used PN’s to built classifiers to help us discern star images from distant galaxies.
>> 
>> Future AI will have to incorporate reasoning and symbolic processing to make progress into constructive synthesis of new knowledge.
>> 
>> 
>> 
>>> On Jan 13, 2026, at 07:45, Yuri Davidovsky (as work at disclosure dot ie) <[email protected]> wrote:
>>> 
>>> 
>>> 
>>>> On 13 Jan 2026, at 15:17, David McClain (as dbm at refined-audiometrics dot com) <[email protected]> wrote:
>>>> 
>>>> In my prior work with MLPs, if you contain factual knowledge, then you will have grossly overdetermined the system. MLPs are intended to discern knowledge regions and their boundaries, to enable generalization recognition. We never wanted specific detail recognition - that would be simpler to deterministically contain using simple decision tree logic, not MLPs.
>>> 
>>> While we are on the topic of perceptrons, they are the actual reason why I am being very critical of the Common Lisp specs as incomplete due to its lack of focus on numerical type handling and binary datatypes IO (really this is the main gripe I am having with lisp, all else is secondary).
>>> 
>>> The reason why it appears to be lacking in the specs (in my opinion), is that lisps were the current tool of AI research of the time and in that period expert systems were in the vogue and assumed to be THE AI. However, in the last decade (or even a decade and a half, depending how you count) neural networks became the architecture of choice for artificial intelligence study, becoming the second wing of research in the area. So the wings we have are:
>>> 
>>> 1. Expert systems, which rely on precise rule matching and symbolic processing. The spec expected to rule here.
>>> 2. Neural networks — statistical feature extraction tools that rely on numerical processing and high throughput.
>>> 
>>> So essentially by skimming over numeric processing at the time when the specs were written, Common Lisp neutered itself as a contender in its own game, and when it suddenly turned out that neural networks are THE AI, it ended up as audience, rather than a participant. The spec never took requirements for neural network processing into account, despite that they were well known at the time and only limited performance of the hardware was constraining their usefulness.
>>> 
>>> _______________________________________________
>>> Lisp Hug - the mailing list for LispWorks users
>>> [email protected]
>>> http://www.lispworks.com/support/lisp-hug.html
>> 
>> 
>> _______________________________________________
>> Lisp Hug - the mailing list for LispWorks users
>> [email protected]
>> http://www.lispworks.com/support/lisp-hug.html
> 


_______________________________________________
Lisp Hug - the mailing list for LispWorks users
[email protected]
http://www.lispworks.com/support/lisp-hug.html