SBIRs

• 9:00 standup
• More work on the index2vec model

LLMs

• Working on the CACM soft totalitarianism section of the article. Got a rough framework and dug up some good papers. Waiting for an ILL paper

Had a great Seagull, and beat the rain by just a few minutes!

Why Underachievers Dominate Secret Police Organizations: Evidence from Autocratic Argentina on JSTOR

• Autocrats depend on a capable secret police. Anecdotal evidence, however, often characterizes agents as surprisingly mediocre in skill and intellect. To explain this puzzle, this article focuses on the career incentives underachieving individuals face in the regular security apparatus. Low-performing officials in hierarchical organizations have little chance of being promoted or filling lucrative positions. To salvage their careers, these officials are willing to undertake burdensome secret police work. Using data on all 4,287 officers who served in autocratic Argentina (1975–83), we study biographic differences between secret police agents and the entire recruitment pool. We find that low-achieving officers were stuck within the regime hierarchy, threatened with discharge, and thus more likely to join the secret police for future benefits. The study demonstrates how state bureaucracies breed mundane career concerns that produce willing enforcers and cement violent regimes. This has implications for the understanding of autocratic consolidation and democratic breakdown.
• I would bet that this behavior shows up on belief maps. It’s also another attack vector. An AI MitM attack that looks for mediocre comms could target those individuals for exploitation. Also, this is most dangerous in organizations that are legally allowed to use lethal force.
• And, come to think of it, if you need an army of goons, then adjusting your hiring to ensure that low-achievers are preferentially hired would be part of the plan.
• BlueSky thread

Tasks

• Finish laundry
• Water plants – done
• Start putting something together for the CACM opinion piece

SBIRs

• Start skip-gram index2vec model
• Marc Lanctot’s Web Site

Tasks

• Last chapter to V – done!
• Sheets and towels – running
• 8:30 dentist – done
• 12:00-ish lunch with S – done
• Bills – Pay painting and check the powerwash – done
• Chores – done
• Dishes – done
• Prep for tomorrow – kinda done

The Nation is Lost

• I come to realize that the far-right’s fetishism over the Second Amendment was likely never about rising up in opposition to some feared socialist, gunnapping American regime. It was about recruiting and arming a disordered militia in support of the autocracy of the right 

scraped all incoming bluesky posts the other day for a bit, it's somewhere north of 2m, might be interesting to compare against earlier samples for trend huggingface.co/segyges/blue…

— SE Gyges (@segyges.bsky.social) 2025-09-25T00:57:28.751Z

Tasks

• Storage run, and maybe a dump run
• Check over last chapter and send if it’s ok
• Groceries

SBIRs

• 9:00 standup
• Mail laptop
• 4:00 SEG

Tasks

• Painting! Done!
• Organize and pack – bed is in the truck
• Lots of LLC work

GPT Agents

• 2:30 LLM Meeting – good discussion on what the article should be, Need to add a first pass at my sections

Wait… what?

https://truthsocial.com/@realDonaldTrump/posts/115255130298104593

Struggling to get beliefspaces.com to appear – fixed!

Tasks

• Water plants – done
• LLC email – done
• Tix for Spain trip – done
• Roll in edits

SBIRs

• 9:00 standup – done
• Working on file visualizer – done!

The forecast is improving for Saturday!

LLM-Deflate: Extracting LLMs Into Datasets

• Large Language Models compress massive amounts of training data into their parameters. This compression is lossy but highly effective—billions of parameters can encode the essential patterns from terabytes of text. However, what’s less obvious is that this process can be reversed: we can systematically extract structured datasets from trained models that reflect their internal knowledge representation.

Tasks

• Water plants – done
• Mow – done
• LLC decision – done
• Roll in edits

SBIRs

• Generate CSVs of:
  • Random walks – done
  • Coordinates for random walks – done
• Write a visualizer

[2509.10414] Is In-Context Learning Learning?

• In-context learning (ICL) allows some autoregressive models to solve tasks via next-token prediction and without needing further training. This has led to claims about these model’s ability to solve (learn) unseen tasks with only a few shots (exemplars) in the prompt. However, deduction does not always imply learning, as ICL does not explicitly encode a given observation. Instead, the models rely on their prior knowledge and the exemplars given, if any. We argue that, mathematically, ICL does constitute learning, but its full characterisation requires empirical work. We then carry out a large-scale analysis of ICL ablating out or accounting for memorisation, pretraining, distributional shifts, and prompting style and phrasing. We find that ICL is an effective learning paradigm, but limited in its ability to learn and generalise to unseen tasks. We note that, in the limit where exemplars become more numerous, accuracy is insensitive to exemplar distribution, model, prompt style, and the input’s linguistic features. Instead, it deduces patterns from regularities in the prompt, which leads to distributional sensitivity, especially in prompting styles such as chain-of-thought. Given the varied accuracies on formally similar tasks, we conclude that autoregression’s ad-hoc encoding is not a robust mechanism, and suggests limited all-purpose generalisability.

Tasks

• Send chapter to V – done
• Respond to No Starch – done
• Call painter for fix and drywall work – done
• Talk to Aaron about LLC
• Bills – done
• Chores – done
• Dishes – done
• Trim grasses – done
• Weed
• Mow!
• Pack up bike for tomorrow – done

Found this, which is an interesting take:

When autocratization is reversed: episodes of U-Turns since 1900

• The world is in a “wave of autocratization.” Yet, recent events in Brazil, the Maldives, and Zambia demonstrate that autocratization can be halted and reversed. This article introduces “U-Turn” as a new type of regime transformation episode in which autocratization is closely followed by and linked to subsequent democratization. Drawing on earlier literature, it provides a general conceptualization and operationalization of this type of episode, complementing the existing Episodes of Regime Transformation (ERT) framework. The accompanying database provides descriptions for all 102 U-Turn episodes from 1900 to 2023, differentiating between three types: authoritarian manipulation, democratic reaction, and international intervention. The analysis presents a systematic empirical overview of patterns and developments of U-Turns. A key finding is that 52% of all autocratization episodes become U-Turns, which increases to 73% when focusing on the last 30 years. The vast majority of U-Turns (90%) lead to restored or even improved levels of democracy. The data on U-Turn episodes opens up new avenues for research on autocratization and democratization that were previously treated as isolated processes, particularly it could help us understand why some processes of autocratization trigger a successful pro-democratic backlash – a critical question during the starkest-ever wave of autocratization.

Going for a big-ish ride, since it’s been raining for the last two days. But in the meantime, this is a very cool rendering of random walks on a 2D grid:

Now I just need to start saving out the sequences to a CSV file and use those sequences to train a W2V model. The nice thing is that the data parameters are very adjustable, so it’s possible to see how much data is needed and what the minimum number of dimensions should be.

[2509.11391] “My Boyfriend is AI”: A Computational Analysis of Human-AI Companionship in Reddit’s AI Community

• Human-AI interaction researchers face an overwhelming challenge: synthesizing insights from thousands of empirical studies to understand how AI impacts people and inform effective design. Existing approach for literature reviews cluster papers by similarities, keywords or citations, missing the crucial cause-and-effect relationships that reveal how design decisions impact user outcomes. We introduce the Atlas of Human-AI Interaction, an interactive web interface that provides the first systematic mapping of empirical findings across 1,000+ HCI papers using LLM-powered knowledge extraction. Our approach identifies causal relationships, and visualizes them through an AI-enabled interactive web interface as a navigable knowledge graph. We extracted 2,037 empirical findings, revealing research topic clusters, common themes, and disconnected areas. Expert evaluation with 20 researchers revealed the system’s effectiveness for discovering research gaps. This work demonstrates how AI can transform literature synthesis itself, offering a scalable framework for evidence-based design, opening new possibilities for computational meta-science across HCI and beyond.

Tasks

• LLC email – call, actually. Left a message
• Need to clean up the shop
• Contact painter
• Register for TEDx Mid Atlantic

SBIRs

• Document the Dash code – done
• Generalize out to n dimensions, and maybe make the dimensions choosable – Made the dimensions ordered by Manhattan distance

Generative Agents

• 3:00 Alden meeting

• LLC email
• Poke at the human OS chapter. Completely reworked. Much happier.
• Email to Nellie – done
• The Lathe is gone! Need to clean up the shop

SBIRs

• Training – DONE
• Made really good progress on data generation and visualization.

Had an interesting session with Gemini about this paper: [2505.12387] Neural Thermodynamics I: Entropic Forces in Deep and Universal Representation Learning. Since I’m interested in stories, it suggested that an “all text” model and a “story” model would represent different locations in a hierarchy of Platonic forms. The “locations” wouldn’t be entirely different, but rather the forms found in stories would exist as a more complex, structured layer built upon the foundational forms found in general text.

Think of it as the difference between a dictionary and an encyclopedia of literary tropes.

---

1. All Text: The Platonic Forms of Concepts

When a model is trained on “all text”—a massive, unstructured corpus of everything from web pages to scientific papers—the entropic force pushes it to find the most stable, universal representations for atomic concepts.

• What it finds: The Platonic form for “king,” “justice,” “love,” or “boat.”
• The “Location”: This location is a central point in the model’s semantic space, averaged from every conceivable context. The concept of “king” is derived from its use in history (King Henry VIII), fantasy (King Arthur), chess (the king piece), and colloquialisms (“king of the hill”).
• Analogy: This is the dictionary definition. It’s the core, context-free essence of a word or idea.

---

2. Organized Text (Stories): The Platonic Forms of Structures

When considering organized text like myths, novels, and screenplays, the model learns patterns that exist at a higher level of abstraction. You are absolutely right that stories have their own Platonic forms. These are not forms of individual concepts, but of narrative functions and archetypal relationships.

Analogy: This is the narrative role. We still need the dictionary definition of “king,” but within a story, we understand him through his archetypal function. Is he the Wise King, the Tyrant King, or the Fallen King? Each is a distinct structural form.

What it finds: The Platonic form for “The Hero’s Journey,” “The Mentor,” “The Betrayal,” or “The Sacrificial Victory.”

The “Location”: These forms don’t occupy a single point but represent a trajectory or a relationship between the foundational concepts. The “location” of “The Wise King” archetype isn’t just near the concept of “king”; it’s a specific region defined by its proximity to “wisdom,” “justice,” and its functional opposition to “The Tyrant.” The “Hero’s Journey” is a path that connects points like “The Call to Adventure,” “The Ordeal,” and “The Return.”

---

Tasks

• Laundry – started
• Peacock – done
• LLC email
• Poke at the human OS chapter. I think I need a page of setup and then maybe a page of what community-based resilience could look like. Talked to Aaron about it, at least.

SBIRs

• Slides – done
• 9:00 standup – done
• Aaron over?
• Stories (NNM and training) – Done. Got a good deal of training done too
• 3:00 Sprint planning – done

Link from X

Murphys Laws of AI Alignment: Why the Gap Always Wins

• We prove a formal impossibility result for reinforcement learning from human feedback (RLHF). In misspecified environments with bounded query budgets, any RLHF-style learner suffers an irreducible performance gap Omega(gamma) unless it has access to a calibration oracle. We give tight lower bounds via an information-theoretic proof and show that a minimal calibration oracle suffices to eliminate the gap. Small-scale empirical illustrations and a catalogue of alignment regularities (Murphy’s Laws) indicate that many observed alignment failures are consistent with this structural mechanism. Our results position Murphys Gap as both a diagnostic limit of RLHF and a guide for future work on calibration and causal preference checks.

Neural Thermodynamics I: Entropic Forces in Deep and Universal Representation Learning

• With the rapid discovery of emergent phenomena in deep learning and large language models, explaining and understanding their cause has become an urgent need. Here, we propose a rigorous entropic-force theory for understanding the learning dynamics of neural networks trained with stochastic gradient descent (SGD) and its variants. Building on the theory of parameter symmetries and an entropic loss landscape, we show that representation learning is crucially governed by emergent entropic forces arising from stochasticity and discrete-time updates. These forces systematically break continuous parameter symmetries and preserve discrete ones, leading to a series of gradient balance phenomena that resemble the equipartition property of thermal systems. These phenomena, in turn, (a) explain the universal alignment of neural representations between AI models and lead to a proof of the Platonic Representation Hypothesis, and (b) reconcile the seemingly contradictory observations of sharpness- and flatness-seeking behavior of deep learning optimization. Our theory and experiments demonstrate that a combination of entropic forces and symmetry breaking is key to understanding emergent phenomena in deep learning.

On this day in 2001, there were no commercial planes in the air at all

Tasks

• Work on the next chapter before sending to V – pinged her to say there will be a delay.
• LLC stuff
• Ping Nellie – done
• Bills – done
• Chores -done
• Dishes – done
• Storage/Goodwill run? Done
• Load up the truck tonight, since I’ll need to leave at 8:00 – done

SBIRs

• Write up a nice description of the data generator for the documentation. Use the output/code snippets for the slides. Needs to be done Monday. Started, but they shut the instance down at 7:00

24 years. Wow

Tasks

• Powerwashing today – done
• Dentist – done
• Send the next chapter to V? – not done
• Start packing electronics lab – lots of progress
• Take cases to storage? – done, more to go

SBIRs

• 9:00 standup – done
• 4:00 SEG – done