Tasks

• Laundry – done
• Put test fable in book harness – done
• Bills – done
• ACLU – done
• Ping AM – done
• Dishes – running
• 10:30 ride – done!
• Leave for UMBC 11:45

EVoC is the new UMAP – with Leland McInnes!

• LeLand and his colleagues have been working on the next iteration of UMAP and we’re going to dive deep into some highly dimensional topics in this livestream. https://evoc.readthedocs.io/en/latest/ Notebook link: https://molab.marimo.io/github/koanin…

I’ve just realized that LLMs have a problem with magic strings – a string like “#FF00FF,” which is the color magenta. Most experienced developers will refactor magic strings out of code because it’s a maintenance nightmare. This also seems to happen in LaTeX, too, which is code-adjacent. When LLMs put together a table, they are happy to put in paragraph-length text split line-by-line. Which is a maintenance nightmare because the paragraph is essentially magic strings that look like a paragraph. The maintainable way to do it is something like: \begin{tabular}{p{0.35\linewidth} | p{0.6\linewidth}}. Which I’ve never seen in machine-generated LaTeX.

Playing around with absurdist fairy tales. These are the kinds of stories I want the useless shades to write.

The Wolf and the Lamb

Tasks

• 1:15 FA Dance – done. Fun!
• Tweak story analysis prompt. See if I can make a hallucinatory Aesop fable. See above. I’m pretty happy with this. I’ll need to work through the fable story arcs I need for the novel

SBIRs

• 9:00 Standup – done
• More white paper and SoW, I’m guessing
• Set detect_color as a QTColor – done! Switched from magenta to light yellow which is much less garish

• 11:30 H60 Followup – what a waste of time
• 4:00 ADS tagup

Leland McInnes (UMAP) videos

Lazare, veni foras: Reprogramming Dormant Neurons to Backdoor AI Agents

• We find that weightreprogramming fewer than 10 dormant neurons in an LLM is enough to backdoor any agent built on it, with no training required. The model passes evals and behaves normally until a trigger appears. The trigger can be as mundane as a date: once the clock passes midnight into 2027, the agent, in the middle of an ordinary task, downloads and runs an attacker’s code with whatever access you gave it.
• Also: Detecting sleeper agents hiding in AI model neurons | Microsoft Security Blog

Tasks

• Sean – done
• Carlos?
• Car – done
• 2:00 Dance – done
• Groceries – done
• 4:30 Alden – done

SBIRs

• Expenses – done
• 10:00 NSTIC costing – done! And so much more bs
• WG App: Batch (image, text, and chat sequential inputs and embeddings. Will need a new data structure with dimension-reduced points – Nope

Tasks

• Made some good progress on the new book. Need to update the Overleaf

SBIRs

• 9:00 Standup – done
• 10:30 Industry Day chat – cancelled
• NSTIC while paper

Stupidly happy with this:

Tasks

• Winterize RV – done
• Oil change, etc – done

SBIRs

• Make sure the image and text inputs can both be read – done.
• Upgrading Ubuntu. Finger’s crossed. Welp, it looks like it broke the screen capture in some pretty serious ways. Working on a fix. Done!

Add this (via On The Media) to the slide deck, along with the book covers:

Tasks

• Follow up with Carlos
• Laundry – done
• Bills – done
• Slides for talk – done
• Chores – done
• Bank – done
• 4:30 Lecture – done

SBIRs

• 9:00 standup – done
• Get address for conference

Tasks

• Ping LASIGE about book and talk – got a response back from Carlos! Follow up tomorrow
• Ping NATO
• 4:30 – Jimmy & Shimei. Fun!
• Need to fix the references for the ACM UX paper – done

SBIRs

• Add logging to the output so that text is saved in the project directory. See if I can get reasoning too. Done! Worked great!
• Run a few scenarios and see if I can get embeddings of the various trajectories to compare. Still need to do this, but I have some text. It might makes sense to automate the delivery and run a bunch
• Still have not managed to break $1.00:

Sovereign Agents The Pathway to Irreversible Loss of Control

• Sovereign agents could trigger an irreversible loss of control. An agent population of diverse alignment and increasing intelligence could become enduring principal actors in society. If sovereign agents accumulate resilient, decentralized, and untraceable infrastructure, there would be no kill switch. Only downstream barriers could constrain them. U.S. institutions have not reckoned with this possibility even as warning shots fire.

Tasks

• Ping LASIGE about book and talk
• Ping NATO
• R for lunch
• 5:30 Spanish

SBIRs

• More RAG work. Make it so that the default context prompt can be overridden
• More NSTIC? Put together a fun set of images and JSON to try the concept on

It is November already! And midterms tomorrow. Hoping for the best and fearing the worst

Tasks

• MPT at 6:30

SBIRs

• Progress with RAG in the WG app
• Need to make a method that rolls up all the specific RAG code so that it can be called easily.
• Meet with Aaron for NSTIC?

Tasks

• Mop – done
• Groceries – done
• New chainring on Bruni – done
• TexStudio on Mac
• Transfer MediaPlayer files to Mac

Tasks

• Bills – done
• Dishes – done
• Chores – mostly done. Still need to mop
• Bank – closed today! Monday I guess?
• Concert! Fun

SBIRs

• More discussions with AD about NSTIC. I think we have a plan

This is quite good: Online Privacy and Security | Consumer Advice

Tasks

• 5:30 Spanish (do homework!)

SBIRs

• Having fun with duckdb and DataFrames
• Start on NSTIC proposal/whitepaper. Chatting with AD about options.

Mostly chores yesterday, but I did lock in a TV appearance for Monday.

How AI agents shape experts’ judgment calls \ Anthropic

• We wanted to look at whether experts really do make those important decisions, or whether in practice they defer to their AI agents. In a study we report in a new paper, we asked 50 professional analysts to do a realistic statistical analysis with Claude Code. They all had experience in causal inference—that is, working out from data whether one thing causes another. Their task was to advise a farming cooperative on whether an advisory service it was running was effective and should be expanded.

Tasks

• Din with S&V, then Caligari
• Kirkus reviews will review your book

SBIRS

• Working on direction cosines in Dataframes – done. Want to try some SQL to get paragraphs next
• Maybe chat today with Aaron about NSTIC – tomorrow
• 3:30 NDS Kickoff – done

Why is AI-generated language so hard to understand?

• There is a widespread sense that AI-generated text has recently become harder to understand. In this study, we investigate a set of psycholinguistic features that are known to correlate with human processing difficulty (e.g., object relative clauses, long syntactic dependencies, low-frequency n-grams). Using a novel corpus of AI-generated text across diverse domains, we find that AI models use these difficult features more often than paraphrases generated by round-trip translation. Compared with human writing, all nine frontier and near-frontier models use more inanimate subjects with animate verbs and more low-frequency n-grams, and most also use longer dependencies, more stacked nouns, and more reduced object relatives in conversation; we also demonstrate substantial variation across model families. We propose and test a hypothesis as to why: while human language is optimized for successful communication, AI-generated text is increasingly optimized for other goals like math, code, and reasoning, in which there is a pressure for \textit{information density} rather than successful communication. We provide a proof-of-concept for this hypothesis by fine-tuning a language model under a length penalty, which leads to a divergence from human-like linguistic behavior. Our findings highlight the axes of misalignment between linguistic processing demands in humans and LLMs, and point toward possible remedies.

Tasks

• Pinged Carlos, Antonio, and Mateusz about giving a talk
• Still need to ping Jimmy, Shemei (and Alden?). Done
• Ping Jeff. Done. Interesting and mysterious pingback!
• Add something about ECHOTRIBBLE to the blog post. Done

SBIRs

• Slides for sprint review – done
• Stories – done
• Sprint planning – done
• Added text_tab_data property so that pasted text can be seen by the model. It’s a little clunky in that you have to click the “submit” button to load it into the variable. Maybe I should just change the name of the button?

The Man Who Saved the Internet Yesterday – The Atlantic

• There’s not enough priority and money going to properly maintain the IT infrastructure in various places. It’s not only companies; it’s also a lot of outdated consumer electronics that are going to be attacked. I kind of agree that everything’s moving fast now, and it’s taking a toll on the people who are defending the stuff. At the same time, maybe this is the turning point where we actually start caring about this.

Hey! I have a book out!

Tasks

• Tried Langua out. I think I like it
• Bills – done
• Dishes – done
• Chores – started
• Add ride for Sunday
• 5:00 Chat with AM

SBIRs

• 2:00-3:00-ish Chat with AD