Dog days of summer for sure

Need to start on the tasks Wolfram asked for

SBIRs

• Really interesting discussion with Aaron on CI Agents. It might be possible for token trajectories to maintain their “identity” by looking at the distance between a set of agent responses, and selecting those that have the greatest distance.
• Got some good NNM work done yesterday. I’m almost ready to generate ring buffer text along with activations by token at each layer of the model. And in case I haven’t really mentioned it anywhere else, I think training ring-buffer GPT-2 (or bigger, local models that can quickly be finetuned. You need to be able to access the layers) models via finetuning from a big model with extensive prompt tuning might be a very good way to create local maps.
• 9:00 standup
• Lunchtime ride in this gap!
• 11:00 M30
• 2:00 Conference prep
• 2:30 Hall research
• 4:30 Book club

GPT Agents

• Finish TiiS review – done!
• ICTAI – Downloaded papers. Deadline is August 18, so that’s a bit over 3 days per paper
• 3:00 Meeting
• Can Artificial Intelligence be Open Sourced?
  • This paper explores the potential of open source models to match or even surpass proprietary models in the future. The dialogue reflects a broader debate on the implications of open sourcing AI, weighing the benefits of democratization against the risks of misuse.
• The Public Interest Internet
  • Allow me to open with a wildly speculative question: What if the internet were public interest technology? I mean “internet” the way most people understand it, which is to say our whole digital sphere, and by “public interest” I don’t mean tinkering at the margins to reduce harm from some bad actors or painting some glossy ethics principles atop a pile of exploitative rent-seeking — I mean through and through, warts and all, an internet that works in support of a credible, pragmatic definition of the common good.¹