00:33:18 Charles Ihler: The ‘Build it with Arduino’ is still up on meetup and showing it’s going next weekend as described.
00:44:40 Terry James: What if it’s “Not a Flamethrower” flamethrower?
00:45:46 Lloyd Moore: October Robothon Venue: https://sunnycrest.fwps.org/
00:48:16 Terry James: https://seattlerobotics.org
00:48:44 Terry James: For rules on competition events: https://robothon.org
00:51:32 Terry James: https://snocomakers.org
00:52:55 Terry James: We could start new midnight, astronomy robotics meetings.
01:00:54 James Newton: The Clone people are actually making progress on that creepy robot body. This latest version is pretty impressive in terms of the amount of actuation, and I notice the motion is getting smoother. No where near good enough to stand or do anything that requires coordination, but… getting better.
01:09:13 Terry James: Hmmm – ideas for Aero-Magellan…….
01:09:24 Lloyd Moore: Reacted to “Hmmm – ideas for Aer…” with 👍
01:10:21 Terry James: Question on the FLIR – was the customer using that underwater or above the surface?
01:12:01 Terry James: Do we need pharma to treat hallucinating AI’s?
01:19:48 Terry James: Hurray!
01:20:49 Terry James: On the cost of LIDAR dropping.
01:26:29 Terry James: Or hard…..
01:27:55 James Newton: Well now I’m reviewing everything I said in my head to try to figure out if I said anything scary…
01:34:44 Terry James: MacBook Pro – yes!
01:44:44 Jon Davis: https://www.liquid.ai/blog/liquid-foundation-models-v2-our-second-series-of-generative-ai-models
01:54:43 Jason Seawall: I have a question for Jon
02:00:17 James Newton: Two different meanings of “Training” being used here.
02:00:36 James Newton: “Training” the NN is very complex, difficult, compute hungry, and needs gobs of data.
02:01:18 James Newton: “Traning” by adding a prompt before the prompt that sets the context for the question is easy and pretty darned effective IF the LLM is already trained with data that is useful to the prompt. It helps to “focus” its response.
02:04:18 James Newton: The issue with RAGs is that LLMs have a sort of “Short Term Memory” which is limited. THis is why they “forget” to not change things you told them to not change x interactions ago… it passes out of context. So your ability to focus with a pre-prompt text is limited to the ability of the LLM to include some max amount of data.
02:15:07 James Newton: Is this the MCP you are talking about?
https://github.com/modelcontextprotocol
02:18:12 Steven Kaehler: https://claude.ai/
02:22:33 James Newton: JamesNewton@MassMind.org if you want to email me.
02:24:28 Terry James: Yes, thank you, this was interesting and challenging.
02:24:33 Colin L.: Yes, thank you Jon!
02:30:56 James Newton: Great talk, thanks! I’d like to see more related to robotics specifically.