It is going to be hot today. Ride early!

SBIRs

• Looks like a light day. I’m going to work on the NNM white paper and try to get it to the point to submit. Done! Two pages!
• Ping MARCOM about interview request
• 3:00 WHAIM – changed to work on the NNM white paper
• 1:00 – 4:00 PWND2 industry day. Interesting, but not our thing

SBIRs

• Good lord, I got a (positive!) response from the ARL! In less that 12 hours! Looks like nothing too formal for the white paper: “It’s to help me understand where a new project might fit, with relatively little effort on the part of a PI compared with writing a full (NSF-style/scope) proposal.”
• 9:00 standup
• 1:00 Thunderbolt
• 5:00 S3i meeting

GPT Agents. Need to do more refactoring of the paper

Longest ride of the season this past Saturday. Beautiful, but I was barely in good enough shape.

SBIRs

• Ethics training if I can log in. Nope – still locked out. Ok, now I can get in, but there is nothing there? Even weirder, the course I should take is marked as complete. Not sure what to do at this point since I can’t take a completed course, but I did download the cert if someone changes things again.
• S3i meeting
• Other training? Sent email to T. Looks like the system says I’m done
• Need to figure out a WHAI or NNM demo that can be done by the end of the year. So about 16 weeks, when you pull out TDay and Xmas

SBIRs

• 9:00 Standup
• 11:00 HSI discussion
• 1:00 Tbolt
• 4:00 Book club
• Ethics training if I can log in
• Write email to Iain, get that approved, and send off today

GPT Agents

• Weekly meeting. Still need to read Jimmy’s and Shimei’s section. I did find the link to Wikipedia’s funding, which is pretty impressive

Still can’t find a place to fix the door. I may take the panel off and see if I can just replace/fix the cable

Looks like deepfakes are about to get a whole lot better:

SBIRs

• 10:30 BD discussion on what to do next with WHAI and NNM – done. Need to send some emails. Iain first
• 1:30 CwoC meeting
• 4:30 S3I prep call. Done in 30 minutes!
• Training – finished cyber

GPT Agents

• Got to a good point of the article. Will wait until after Thursday’s meeting before doing anything else. Need to read Jimmy’s and Shimei’s part first, though

Still trying to get the door serviced. Other Ram dealers in the area:

SBIRs

• Finished most BD things that I can do before I get some decision on what to do next. Schedule a meeting? Sent out the contents as an email attachment. Scheduled meeting for 10:30 Wednesday.
• 9:00 Data Science standup
• 2:30 AI Ethics
• 3:45 L3Harris Prep – done
• 4:00 L3Harris meeting

GPT Agents

• Work on the Soft Totalitarianism section and add cites – good progress!

Trying to get the door repair scheduled. The dealer is only accepting drop offs with a multi-week wait. Much harder than it should be. Jim Donnie’s suggested K&L

SBIRs

• Working on getting all the BD pieces together for the NNM/WH-AI. Added the main documents to the appendix, and am working on an inquiry email. Really gotta wonder why we employ all these capture managers.
• Wrote two preliminary inquiry emails
• Had some fun thinking about a better trade show booth.

GPT Agents

• Edit and add the Soft Totalitarianism section. Then get back to adding citations

Tasks

• Schedule to get the door fixed

GPT Agents

• Write up some thoughts on “Soft Totalitarianism” of intelligent technology – done

Tasks

• 9:40 Dentist
• Bank
• Schedule to get the door fixed

SBIRs

• Trawled the swamp and found some good NSF and Army possibilities from grants.gov. You can export the results as a csv file and search through those, either by link (which can be wrong) or by googling project name, which works. Found some very good opportunities for NNMs, and some others for White Hat AI. The earliest opportunity closes Sept 30, which is enough time to write a reasonable proposal. Nothing specifically about spearphishing, which is kind of interesting. That seems to be acceptable in some way.
• 9:00 Standup. Will have to leave early
• 2:00 Thunderbolt meeting
• 4:30 Book club

GPT Agents

• No meeting today, Jimmy’s at a wedding. Add more content?

More justification for WH-AI: Hackers may have stolen the Social Security numbers of every American. How to protect yourself

• “These bad guys, this is what they do for a living,” Murray said. They might send out tens of thousands of queries and get only one response, but that response could net them $10,000 from an unwitting victim. “Ten thousand dollars in one day for having one hit with one victim, that’s a pretty good return on investment,” she said. “That’s what motivates them.”

More stuff for consumer-first AI: Cosmos Magazine publishes AI-generated articles, drawing criticism from journalists, co-founders

Schedule to get the door fixed

SBIRs

• Ping John to see if we can schedule WH-AI architecture planning
• Draw up some diagrams for the architecture that we can go over
  • Information flows
  • Main browser extension – just Chrome for now
  • Maybe three buttons for the popup?- Avoid, disregard, this is an error?
  • Adjustment knobs for target user – also notification settings for guardians (parents, adult children, etc.)
  • Private User database of issued warnings, so that users don’t see the same “introductory” warning. This DB could also have sender information
  • Some other kind of warning if the user is repeatedly interacting with the sender of manipulative email, particularly if it matches one of the scam patterns.
  • Spaced repetition of older warnings
  • Public database of manipulative posts, if warnings were disregarded or heeded. This can feed back to the the Chrome extension as well in case there are multiple adjacent embeddings that are, for example, increasing in a viral way.
  • A UMAP display of the embedding space that lets users navigate and understand what’s going on. Areas of high activity should be indicated. Clicking on a point or dragging across an area should provides specific and/or summary information
  • Reactive design for Chrome on mobile?
• UMAP-JS
  • This is a JavaScript reimplementation of UMAP from the python implementation found at https://github.com/lmcinnes/umap.
• Nope, strike that. Everything has to have a potential target before it gets worked on. So I’ve gone from belief space maps to white hat AI PoC to looking through SBIRs and BAAs. Ah, well, only 132 working days until 15 February 2025.
• Put together a good size list of people to reach out to. Still need to trawl the BAA/SBIR swamp

GPT Agents

• Read what I wrote yesterday. Look for sources. I’m particularly interested if there is anything on the creation of guardrails. There might be a perspective here that comes from labor relations, like having to buy GMO seeds rather than being able to re-plant based on the harvest? Extracting maximum value while providing minimum utility. Also worth re-reading the Stochastic Parrots paper and look for interesting papers that cite it. It strikes me that the whole idea of “public” GenAI may also reference the “right to repair” movement, and Robin Berjon’s The Public Interest Internet.

Tasks

• Schedule to get the truck door fixed
• BoA closes at 4:00. Go over at 3:00?

SBIRs

• Need to talk to Aaron about switching from NNM to WH-AI and see if I can get some Fabric time
• Getting started on Chrome Extensions
• Good chat with Zach. He agrees that starting with a chrome extension makes sense.

GPT Agents

• Need to get started on the “A Consumer-First Approach to GenAI?” paper, now that the reviews are none. Good start

Tasks

• 1:00: Get truck – done. Why is everything $1,000?
• Call bank? Probably tomorrow
• Call Guardian – try to reposition the cover first

SBIRs

• 9:00 Standup – done
• 11:00 AI C of I – done. Maybe that went well?
• 12:00 Steve’s brownbag – Could have been 20 minutes
• Wrote up a proposal for White Hat AI that will “be the D2A of NNM.” Which is ok, I guess. Otherwise progress is going to be imperceptible. And Aaron has a good point in that no one knows the difference between topic embedding space and narrative space. Which is annoying, but makes it easier to make visible progress
• 3:00 Sprint planning
• Do I need to write a SimAccel/D2A white paper?

GPT Agents

• Finish and submit ICTAI reviews – DONE!

Tasks

• Sign things in blue for Wolfram
• Pick up Stacey’s car and get the van if it’s not too wet

SBIRs

• Scan cards
• Expense report
• Put together slides for sprint demo
• USNA interns final presentation
• Conference debrief to Orest
  • The conference was almost exclusively props and videos. No one could really show anything unless you were a manufacturer (“look at this beautiful planetary gear!”)
    • The golf idea was universally well received, and there was one other booth with an interceptor game. I just don’t think it went far enough, and was too easy to defeat

• We need videos that work with no audio. Also the lighting on Aaron makes his head distractingly shiny.
• The booths that had demos had more people as a rule (Blue Halo and Axient were the only ones I saw)
• A monitor in the conference-facing desk would be better than one on the desk
• Chairs need to work on deeply-padded tradeshow carpets. The dis-based chairs we had wobbled. A lot.
• Sequentially numbered/barcoded batches of bears that when scanned link to ASRC branded polar bear cams (or even better an updated curated feed)

GPT Agents

• Write next review and read a paper

Anointed with Oil

• Anointed with Oil places religion and oil at the center of American history. As prize-winning historian Darren Dochuk reveals, from the earliest discovery of oil in America during the Civil War, citizens saw oil as the nation’s special blessing and its peculiar burden, the source of its prophetic mission in the world. Over the century that followed and down to the present day, the oil industry’s leaders and its ordinary workers together fundamentally transformed American religion, business, and politics — boosting America’s ascent as the preeminent global power, giving shape to modern evangelical Christianity, fueling the rise of the Republican Right, and setting the terms for today’s political and environmental debates.

Finishing the conference and heading home:

GPT Agents

• Finished one more paper review and started the fourth one. Pretty good so far!
• Jimmy put in a big chunk of writing on the Consumer-First Approach to AI paper!

Transformer Layers as Painters

• Despite their nearly universal adoption for large language models, the internal workings of transformers are not well understood. We aim to better understand the impact of removing or reorganizing information throughout the layers of a pretrained transformer. Such an understanding could both yield better usage of existing models as well as to make architectural improvements to produce new variants. We present a series of empirical studies on frozen models that show that the lower and final layers of pretrained transformers differ from middle layers, but that middle layers have a surprising amount of uniformity. We further show that some classes of problems have robustness to skipping layers, running the layers in an order different from how they were trained, or running the layers in parallel. Our observations suggest that even frozen pretrained models may gracefully trade accuracy for latency by skipping layers or running layers in parallel.

SBIRs

• At a dumb conference and trade show that I just don’t need to be at.

GPT Agents

• Getting some progress on paper reviews though