It’s Fall! Ping Nathan

SBIRs

• Make a test_harness.tex file – done
• Finish venues – done
• Slides! – done
• 9:00 Sprint demos – done
• Stories! – done
• 2:00 MDA
• 3:00 Sprint planning

GPT Agents

• Test out on people?

Very rainy weekend. Worked a bit on the book, and on trimming long CVs

Connecting Large Language Models with Evolutionary Algorithms Yields Powerful Prompt Optimizers

• Large Language Models (LLMs) excel in various tasks, but they rely on carefully crafted prompts that often demand substantial human effort. To automate this process, in this paper, we propose a novel framework for discrete prompt optimization, called EvoPrompt, which borrows the idea of evolutionary algorithms (EAs) as they exhibit good performance and fast convergence. To enable EAs to work on discrete prompts, which are natural language expressions that need to be coherent and human-readable, we connect LLMs with EAs. This approach allows us to simultaneously leverage the powerful language processing capabilities of LLMs and the efficient optimization performance of EAs. Specifically, abstaining from any gradients or parameters, EvoPrompt starts from a population of prompts and iteratively generates new prompts with LLMs based on the evolutionary operators, improving the population based on the development set. We optimize prompts for both closed- and open-source LLMs including GPT-3.5 and Alpaca, on 9 datasets spanning language understanding and generation tasks. EvoPrompt significantly outperforms human-engineered prompts and existing methods for automatic prompt generation by up to 25% and 14% respectively. Furthermore, EvoPrompt demonstrates that connecting LLMs with EAs creates synergies, which could inspire further research on the combination of LLMs and conventional algorithms.

Mow lawn before it rains for days!

Inside Tiktok’s real-life frenzies – from riots to false murder accusations

• the BBC has identified four episodes in recent months where disproportionate engagement on TikTok was connected to harmful behaviour:
  • An online obsession with a murder case in Idaho, USA, that led to innocent people being falsely accused
  • Interference in the police investigation of Nicola Bulley, who went missing in Lancashire, UK
  • School protests involving vandalism spreading across the UK
  • Fanning flames of riots in France, which spread at an unusual intensity and to unexpected locations
  • Ex-staffers at TikTok liken these frenzies to “wildfires” and describe them as “dangerous”, especially as the app’s audience can be young and impressionable.

SBIRs

• 10:00 meeting with Aaron
• 2:30 Tech Fellows interview
• Finish Dahlgren WP draft 1

GPT Agents

• Add pasting areas for Education history, Work history, and Publications. Cut publications when the entire text is > ~10,000 words

Tweaked my The Great Chain of Being as a General Theory of Racism post

SBIRs

• 9:00 standup – multiple issues with ENU reference frames
• Working on the Dahlgren white paper – good progress
• The Scale, well, I guess it’s on its way to being a book now is closing in on 40k words:

• I found this helpful little site, which suggests that a “big idea” book should be between 60k and 80k words

GPT Agents

• 2:00 Meeting
• Need to fix a few more things

SBIRs

• Need to make a poster and submit by the 28th to the Digital Platforms and Societal Harms event. Probably show the 3 types of attacks (email examples) and mitigation. I could bring a laptop with ContextExplorer too.
• Work on MAST whitepaper, then get together with Aaron at 1:00. Made good progress. The goal is to have a first draft by Friday COB
• 10:00 JSC Data Review. There is a lot. Ron’s going to do some summary statistics.
• Maybe more scale paper this evening? Yup, finished Arms Control

GPT Agents

• 3:00 Meeting with Alden

Caulk tub

SBIRs

• MDA meeting from yesterday because Zac is back now. Done. Need to find out from Bob what the best target is.
• More scale paper. Got started on the Arms Control section, which is coming along nicely. It seems that arms control is most effective when powers are not in open conflict (e.g. the cold war). Which is mostly the case now, though I wonder how much The Russian-Ukraine war would effect that. I think that there would be more focus on AI-enhanced weapons? Which for an agreement on Societal AI weapons might make things easier.
• Need to get some work done on the MAST white paper

GPT Agents

• Progress on getting lists of deans and chairs together to ask for participation.

Centaurs and Cyborgs on the Jagged Frontier (from this paper)

• …for 18 different tasks selected to be realistic samples of the kinds of work done at an elite consulting company, consultants using ChatGPT-4 outperformed those who did not, by a lot. On every dimension. Every way we measured performance.

El ingenioso hidalgo don Quijote de la Mancha

SBIRs

• Meeting with Steve to talk about things
• MDA weekly meeting

GPT Agents

SBIRs

• More scale paper – poked at it a little
• Dahlgren white paper? – Got a good start with Aaron

GPT Agents

• Write the email inviting people to participate and the email to the chairs as well. – done
• Add emails and captions to the Word doc – done
• Submit! – done

Meet Greg at 6:00!

SBIRs

• I guess we’ll see what is going on with the server today?
• 9:00 Standup
• GPT IRAD decision?
• 11:30 CSC
• More scale paper. Need to start looking for some pix. Finished the disruption section. I think counterattack is an extension of disruption, and should be written that way. Of course, there’s a lot of groundwork that would have to be done in advance to put all the actors in place. That’s a tricky issue that’s worth discussing.
• Tweaked the template for the Dahlgren paper and added some links to examples of prompt engineering to produce JSON files
• Add a 0.5 point story for AI ethics

GPT Agents

• 2:00 UMBC Meeting. Test the new ContextTest and walk through the IRB – done with the later. Need to tweak the former – done
• Add education history to work history prompt – done
• Add “I assert that I am at least 18 years old” – done
• Add recruitment email and screenshots to attachments – done
• Change REI to Amazon – done
• Draft email for all department chairs that includes an introduction of what the study is and who we are.

Listening to Alban Claudin’s “Room of Reflection” and quite liking it

SBIRs

• Working on venues for the scale paper/book. Need to start filling out the “defense” section. Started. Finished “Detection.” Next is “Disruption.”
• Wrote up a short Python script that runs the loops that we think would generate the trajectories that we (think?) we need. I just realized that there needs a “trim” function that removes the beginning and end so we only have computable data
• 10:00 meeting with Rukan. The machine is hanging on file access because read permissions have been changed
• 3:00 AI Ethics meeting. Do homework! Done. Shiny, yet bad videos
• Registered for the Digital Platforms and Societal Harms event

GPT Agents

• Looks like we meet at 2:00 on Thursdays
• Got a good start on the IRB! Need some guidance to finish

What if Generative AI turned out to be a Dud?

SBIRs

• Our security people have decided that collaborative writing using overleaf is too much of a threat so they will not allow it. On top of all their other policies, I am very close to quitting.
• Need to register for Digital Platforms and Societal Harms
• Wrote up some code to show the loops for trajectories and sent to SEG
• Work on Scale paper
  • Defense section
  • Venues – Moved to Overleaf. Still need to finish descriptions

GPT Agents

• IRB form! Progress!
• We are now meeting at 2:00 on Thursdays

Twenty-two years ago, I remember this day starting as a crisp autumn morning with infinite, clear blue skies.

Sam Bankman-Fried’s jail conditions offer a glimpse at systemic failure

Everything I’ll forget about prompting LLMs

SBIRs

• Submit expenses
• We need another story. In this case, it’s another war room vignette, but this time from the defense’s side. Maybe with M again? Of course, part of this is figuring out what defenses might actually look like. One thing I’d like to re-use in the idea of diverse operator teams looking for misbehaving models. In this case though, the models are trained to be honeypots for attacks maybe? They go along in their day-to-day, sending emails, running dummy companies, having dates, etc. When they start acting too aligned, then it’s time to start looking for trouble. Maybe digital twins of important people?
• 9:00 Sprint demos. Make Slides
• 2:00 Weekly MDA meeting
• 3:00 Sprint planning

GPT Agents

• Start filling out IRB form

SBIRs

• Had a good chat with Rukan yesterday. What worked with the hdfproc data didn’t work with the new offsets? He’s going to run some tests
• I really want to add a new project to the LLM IRAD. Something like NNMap-enabled group support. Need a better name, some slides (mentioning “killer app” and all the possible uses), and a schedule.
• Tweaked the Jan6 AI subsection to integrate better into the rest of the section
• Need to add a “Detect and Defend” section
• Need to add an “AI Arms Control for Societal AI Weapons” section. Show that this is in everyone’s best interests. Authoritarian regimes are potentially at greater risk, particularly for Spanner and Lobotomy attacks.

SBIRs

• 9:00 standup
• LLM schedule planning with Aaron. Done
• 2:00 Dahlgren follow-up meeting
• More scale paper. Add QAnon as the other main component of Jan6

GPT Agents

• Tests with Roger and/or Aaron?

SBIRs

• Submitted my technical fellows stuff
• Steve’s presentation – added comments
• Installing sw on the laptop – done
• MDA next steps (intersection of TI and current time allows for sync. We’d need several points, but not too many
• LLM planning. Need to create schedules?

GPT Agents

• 3:00 meeting? Yup. Alden seems to be finding traction