Path: csiph.com!news.swapon.de!weretis.net!feeder9.news.weretis.net!border-3.nntp.ord.giganews.com!nntp.giganews.com!Xl.tags.giganews.com!local-3.nntp.ord.giganews.com!news.giganews.com.POSTED!not-for-mail NNTP-Posting-Date: Thu, 13 Aug 2026 22:59:08 +0000 From: steve g Newsgroups: comp.lang.lisp Subject: Re: Resources to learn common lisp? References: <110nf0q$3vra6$3@dont-email.me> <87y0gg62cx.fsf@nightsong.com> <110rfqf$142ot$1@dont-email.me> <111ne5o$vsm3$3@dont-email.me> <111pbk6$36pi7$1@dont-email.me> <111sb7u$3ua8h$7@dont-email.me> <111t73h$6a1k$1@dont-email.me> <1126vrf$31p54$4@dont-email.me> <87wludtv85.fsf@gmail.com> <114e3qf$18m2n$12@dont-email.me> <877bm5ux0t.fsf@gmail.com> <1157fe4$1j95q$1@dont-email.me> <87v79ivrmi.fsf@gmail.com> <51fm7l1ehmk7eqptgaubg0ksga5pqs2ff9@4ax.com> Date: Thu, 13 Aug 2026 18:59:07 -0400 Message-ID: <87se4htyo4.fsf@gmail.com> User-Agent: Gnus/5.13 (Gnus v5.13) Cancel-Lock: sha1:jsdNTcfcoJ6Yjwgy0ds33b5M9Gw= MIME-Version: 1.0 Content-Type: text/plain Lines: 53 X-Usenet-Provider: http://www.giganews.com X-Trace: sv3-skaKx8qWUndweJMPPJD6blp+XMXB3VCuke8Cxpc51eyUeLW9N66miLYg/9r+5mxHTwHBo8RESbBtaY3!2wSs0UhjJwu6L73h2OiGwVvSHU1WFUo/UN2iGi8zI/EHPz0= X-Complaints-To: abuse@giganews.com X-DMCA-Notifications: http://www.giganews.com/info/dmca.html X-Abuse-and-DMCA-Info: Please be sure to forward a copy of ALL headers X-Abuse-and-DMCA-Info: Otherwise we will be unable to process your complaint properly X-Postfilter: 1.3.40 Xref: csiph.com comp.lang.lisp:61449 George Neuner writes: > An LLM is an emulation of a neural net. There's plenty of study > material regarding neural nets. thank you. > Modern implementations forgo the traditional graph representation and > instead are based on matrix math. Ie. you take a bunch of inputs, > multiply each input by a weight, offset the result by adding some > value, then summ and pass the summ put through a (possibly per node) > equation to get an output. That ptoentially feeds into the inputs of > the next layer of nodes. > The difficulty is in getting the topology correct. Too many nodes, > the nodes start to specialize, too few nodes the results are fuzzy. > Bad connection scheme - either it doesn't work or training becomes > very difficult. NNs have been studied since the 1970s, and there are > known topologies to create various kinds of recognizers and types of > (app specific) memories. You can (sort of) plug and play using these > as modules. But still a lot of it is just trial and error. That's why > so much effort is needed to build them. I think I'm starting to see how loose these LLMs can get. > > LLMs [I agree it's a stupid name] are hidden node models having tens > of thousands (or more) of nodes, and many millions of inputs [and > weights and offsets]. You can't build one without a supercomputer. This is an excellent point I neglected to think about. I will be telling him how this kinda works. Big, Fast computers using lots of energy :) Nothing new... > There's a PHd and potential billions of $$$ waiting if you can figure > out how to make a NN reliably forget something it has learned, without > retraining from scratch and without affecting anything else it knows. another very important point. is it possible to use a truth maintenance system? Just curious, it would probably take up too much space and time. Just a jtms is a lot of overhead. You could try to use an LTMS but we all know that this just leads to more wasted space and aggravation. I would think an ATMS would work; I dunno, I do not have a super computer or a billion dollars. > Have fun! thank you so much for your excellent explanation. I really appreciate this. Asking someone about LLMs these days and they talk for hours and such have you.