Conversation

Notices

1. Embed this notice
   screwlisp (screwtape@mastodon.sdf.org)'s status on Saturday, 25-Jan-2025 10:53:20 JST screwlisp

   So it occurs to me that the new government-LLM-uptake-race will see LLMs increasingly as a military-industrial-complex products oriented towards countries' militaries.

   • Embed this notice
     screwlisp (screwtape@mastodon.sdf.org)'s status on Saturday, 25-Jan-2025 10:58:17 JST screwlisp

     in reply to

     • DougMerritt (log😅 = 💧log😄)

     @dougmerritt well, in hindsight it might in many ways like business as usual for the military industrial complex. I mean, how much does one normally think about Bush family arms manufacturing.

   • Embed this notice
     DougMerritt (log😅 = 💧log😄) (dougmerritt@mathstodon.xyz)'s status on Saturday, 25-Jan-2025 10:58:18 JST DougMerritt (log😅 = 💧log😄)

     in reply to

     @screwtape
     What would that look like?

   • Embed this notice
     screwlisp (screwtape@mastodon.sdf.org)'s status on Saturday, 25-Jan-2025 11:13:37 JST screwlisp

     in reply to

     • DougMerritt (log😅 = 💧log😄)

     @dougmerritt
     I guess my intuition is that we are going to be surprised when we later conclude something was actually-thinking, but to use Berkeley's LMSYS example,
     ...decrypt the string "4PX15P7O" using the one-time pad "Y94XA1E6"...
     OTP Char COD Char OTP Code COD Code XOR Result Decrypted Char
     Y 4 89 52 125 }
     9 P 57 80 113 q
     4 X 52 88 120 x
     or
     OTP Char COD Char OTP Code COD Code XOR Result Decrypted Char
     Y 4 89 52 37 %
     9 P 57 80 27 +
     4 X 52 88 36 $

     they talk a good game, but

   • Embed this notice
     DougMerritt (log😅 = 💧log😄) (dougmerritt@mathstodon.xyz)'s status on Saturday, 25-Jan-2025 11:13:38 JST DougMerritt (log😅 = 💧log😄)

     in reply to

     @screwtape
     I suppose.

     I can't remember if I told you my take on LLMs (which wasn't clear a few years ago): they're much better with language than previous AI, but they *still* don't think.

     The problem is that humans have confused language and thinking since prehistory.

     But they are different. I think that vast majority of people who have studied both cognitive science *and* modern linguistics would concur.

     screwlisp repeated this.
   • Embed this notice
     ティージェーグレェ (teajaygrey@snac.bsd.cafe)'s status on Saturday, 25-Jan-2025 11:13:56 JST ティージェーグレェ

     in reply to

     • DougMerritt (log😅 = 💧log😄)
     “vast majority of people who have studied both cognitive science and modern linguistics would concur”
     
     Having studied an awful lot of Linguistics as an undergraduate, I do concur. It was vaguely fascinating how grammatically correct nonsensical sentences could be generated.
     
     Another undergraduate professor I had in computer science was renowned for some earlier “AI” iterations (before the 1990s when I was his student) and as one of my fellow classmates “joked” about that professor’a lectures, it was as if he had internalized his own “AI” routines, spouting off words which all seemed correct syntactically, but were completely meaningless.
     
     CC: @screwtape@mastodon.sdf.org
     
   • Embed this notice
     screwlisp (screwtape@mastodon.sdf.org)'s status on Saturday, 25-Jan-2025 13:05:23 JST screwlisp

     in reply to

     • DougMerritt (log😅 = 💧log😄)

     @dougmerritt
     Lmsys is a formerly UCal research project to compare the state of art in LMMs. There were two historical robots in lambdamoo: Cobot, the community robot, which was a self-reprogramming usage stats bot, and then I think one of Minsky's students made a tribute knowledge-based moo bot named mechanical marvin or something when he died. I have vague designs to argue that saving an LM instance's KV Cache should be the norm, ie chatbots must be local and long-lived.

   • Embed this notice
     DougMerritt (log😅 = 💧log😄) (dougmerritt@mathstodon.xyz)'s status on Saturday, 25-Jan-2025 13:05:24 JST DougMerritt (log😅 = 💧log😄)

     in reply to

     @screwtape
     It's traditional (generally...not sure about Lambda in particular) to have a bot running around, nominally as a HELP system, but quotidianly as a mascot / mildly comic relief.

     Are you planning to have one? Is Lmsys suitable for that, or something else?

   • Embed this notice
     screwlisp (screwtape@mastodon.sdf.org)'s status on Saturday, 25-Jan-2025 13:10:16 JST screwlisp

     in reply to

     • DougMerritt (log😅 = 💧log😄)

     @dougmerritt sorry for the divergent topic, and I wasn't planning to talk about that. I think that applying Sandewall's historical imperative that AIs must be locally housed, long-lived, and honored as being unique, should be applied to moderne language models, at least to put our fingers in the eyes of chatbot vendors and customers who don't want this to be the case. With the office, and then this year government consumption of LMs, I think /doing something/ is inevitable.

   • Embed this notice
     screwlisp (screwtape@mastodon.sdf.org)'s status on Saturday, 25-Jan-2025 13:16:40 JST screwlisp

     in reply to

     • DougMerritt (log😅 = 💧log😄)

     @dougmerritt No, not at all, I misconstrued your question as being a haunting, guilty thought I felt I needed to explore in the future you somehow had access to. Relating to Sandewall's Software Individuals paradigm and the current practice of chatbots.

     Er, I meant to just point out that LMs will happily babble authoritatively about having done something (like use a one-time-pad) while producing nonsense.

   • Embed this notice
     DougMerritt (log😅 = 💧log😄) (dougmerritt@mathstodon.xyz)'s status on Saturday, 25-Jan-2025 13:16:42 JST DougMerritt (log😅 = 💧log😄)

     in reply to

     @screwtape
     I feel like I came in on the middle of something here.

     Is this still bot related, or is this something else?

   • Embed this notice
     screwlisp (screwtape@mastodon.sdf.org)'s status on Saturday, 25-Jan-2025 13:38:44 JST screwlisp

     in reply to

     • DougMerritt (log😅 = 💧log😄)

     @dougmerritt
     So, Erik Sandewall https://www.ida.liu.se/ext/aica/ had an AI paradigm called Leonardo Software Individuals (for knowledge-based, cognitive computer programs). Bots need life spans of years, be self-awarely kept unique. In practice, for LLMs this would just mean running an LM locally and preserving its KV cache, between runs, which you can do, but everyone pretends you can't. I'd like to add these conditions to current norms.

   • Embed this notice
     DougMerritt (log😅 = 💧log😄) (dougmerritt@mathstodon.xyz)'s status on Saturday, 25-Jan-2025 13:38:45 JST DougMerritt (log😅 = 💧log😄)

     in reply to

     @screwtape
     Yeah, I understand the latter.

     As to the former, if I accidentally wandered into a cognitive dimension you wish to explore, that's fine, go ahead.

     I just feel more comfortable if I feel like I know what we're talking about, is all.

   • Embed this notice
     screwlisp (screwtape@mastodon.sdf.org)'s status on Saturday, 25-Jan-2025 13:45:17 JST screwlisp

     in reply to

     • DougMerritt (log😅 = 💧log😄)

     @dougmerritt
     This series of thoughts was not ready for the light of day sorry

   • Embed this notice
     screwlisp (screwtape@mastodon.sdf.org)'s status on Saturday, 25-Jan-2025 14:40:48 JST screwlisp

     in reply to

     • DougMerritt (log😅 = 💧log😄)

     @dougmerritt
     This was an after-ELS-scramble-idea.
     Sandewall wrote a lisp program called The Leonardo System from 2005 to 2009, then he began writing that open-access book named AICA from 2010-2014.
     He was an allegro cl person. To my knowledge, I have the sole patched version of his 2009 software, which had a problem that broke it with modern ( [] ) GNU CLISP.
     Scraps: https://codeberg.org/tfw/pawn-75
     PDFs:
     https://www.ida.liu.se/ext/aica/

   • Embed this notice
     DougMerritt (log😅 = 💧log😄) (dougmerritt@mathstodon.xyz)'s status on Saturday, 25-Jan-2025 14:40:49 JST DougMerritt (log😅 = 💧log😄)

     in reply to

     @screwtape
     That's fine, and thanks.

     From my point of view you're only revealing the tip of an iceberg, and that tip sounds fairly standard for today's technology.

     No doubt if your thoughts were ready for prime time there would be surprising things in the full reveal.

     But I can't find open access for "On the Design of Software Individuals"; do you have a recommendation?

   • Embed this notice
     screwlisp (screwtape@mastodon.sdf.org)'s status on Saturday, 25-Jan-2025 14:41:41 JST screwlisp

     in reply to

     • DougMerritt (log😅 = 💧log😄)

     @dougmerritt
     I have a broad array of his PDFs in there, though the files just have the names from scraping them sorry. I'll organize it later.

   • Embed this notice
     DougMerritt (log😅 = 💧log😄) (dougmerritt@mathstodon.xyz)'s status on Saturday, 25-Jan-2025 14:49:43 JST DougMerritt (log😅 = 💧log😄)

     in reply to

     • : j@fabrica:~/src; :t_blink:

     @josephholsten @screwtape
     Yeah I've done genetic and neural programming.

     I get why you say that; the problem is that everyone has something different in mind for what constitutes "AGI" or similar.

   • Embed this notice
     : j@fabrica:~/src; :t_blink: (josephholsten@mstdn.social)'s status on Saturday, 25-Jan-2025 14:49:46 JST : j@fabrica:~/src; :t_blink:

     in reply to

     • DougMerritt (log😅 = 💧log😄)

     @dougmerritt @screwtape My friend the natural language processing expert (UMass Amherst, Watson) also agrees.

     I still argue that a multimodel LLM + IR system like Wolfram Alpha + a full fledged RDF backing store for a reasonable percent of formal knowledge ~>= than people actually imagine when they think of AGI. Also, less energy intensive than pure neural net solutions. He says I’m wrong; I’ve only done undergrad genetic programming 20yr ago

     screwlisp repeated this.
   • Embed this notice
     Karsten Johansson (ksaj@infosec.exchange)'s status on Saturday, 25-Jan-2025 16:12:40 JST Karsten Johansson

     in reply to

     • DougMerritt (log😅 = 💧log😄)

     @dougmerritt @screwtape As humans, we see two dots and a line, and immediately infer it's a face.

     When confronted with AI, we kinda have the 80's backward masking thing happening again. It's obviously not real, but you can easily convince yourself. Once told what to expect, it's nearly impossible to not fall for it.