Had a thought about terminology, which is probably worth a blog post. LLMs don’t “reason over” data. The prompt “navigates” over it’s previous tokens, under the influence of the model. The analogy is more like how a cell can chase a chemical gradient in a complex environment than how an intelligent being thinks. It’s like Simon’s Ant, MKII.

GPT Agents

• Fixed a typo in ContextText that was found by one of the subjects. We’re at 47!

SBIRS

• MORS ETF. My talk is today!
• Learned about Mistral, which has a nice, small model that might be a nice demo for the MitM email manipulation. It’s on Huggingface, of course

How Much Do Americans Trust the Media?

GPT Agents

• Responses are accumulating! Tonight I’ll send out the follow-up email to dead registrations

SBIRs

• MORS Emerging Techniques Forum
  • Dr. Nathaniel D. Bastian
  • Information Innovation Office (I2O)
  • Iain Cruickshank
  • The GDELT Project, Supported by Google Jigsaw, monitors the world’s broadcast, print, and web news from nearly every corner of every country in over 100 languages and identifies the people, locations, organizations, themes, sources, emotions, counts, quotes, images and events driving our global society every second of every day, creating a free open platform for computing on the entire world.
• Sprint demos and planning yesterday.
• Dr. Kalev Leetaru gave the keynote. In a follow up discussion, he described cases where RAG-based summaries would still get things wrong, as in the case of the Chinese balloon, where the models was pulled towards more sophisticated technology, or in the case of translations about Ukraine, where the response to the question “who has provided the most weapons to Ukraine?” the answer was Russia, which could have been the misattribution of arms being used in an attack for being provided for defense. Also, the model can get hun up on the term “cases” in the medical domain.
• Pinged @lmcinnes@mastodon.social – a response would be nice, though not expected. The type of mathematician I’d like to find though.

SBIRs

• Got the slides off and the status report completed. Need to send the Q7 report to Lauren.
• 9:00 Sprint demos
• 2:00 MDA meeting
• 3:00 Sprint planning

GPT Agents

• Getting some traction with UMD participants. We’re over the 30 person lower limit – 38 as of today!
• And more! I reworked the display so it shows only participants who competed the survey:

Tasks

• Call to see if I can drop the bike off – done
• Shopping – done
• Drain cleaner – done
• 2:00 meeting with Deepan – done
• Track shorts

Future Lens: Anticipating Subsequent Tokens from a Single Hidden State

• We conjecture that hidden state vectors corresponding to individual input tokens encode information sufficient to accurately predict several tokens ahead. More concretely, in this paper we ask: Given a hidden (internal) representation of a single token at position t in an input, can we reliably anticipate the tokens that will appear at positions ≥t+2? To test this, we measure linear approximation and causal intervention methods in GPT-J-6B to evaluate the degree to which individual hidden states in the network contain signal rich enough to predict future hidden states and, ultimately, token outputs. We find that, at some layers, we can approximate a model’s output with more than 48% accuracy with respect to its prediction of subsequent tokens through a single hidden state. Finally we present a “Future Lens” visualization that uses these methods to create a new view of transformer states.
• This looks like a good paper to extend the NNM work. They are looking at activations in different layers and using them to work out a trajectory. Need to dig into further.