I made the front page of Communications of the ACM!

Organizational Lobotomy

Tasks

• Connected to ulisboa on the iOS app, still can’t connect on the MS Outlook client. Sent an update
• Bills – done
• Lawn – done
• Weed – done
• Clean – done
• Dishes – done
• Ride with Aaron B and maybe get rid of some metal! – done
• Drat – Elsevier passed on KA. Asked Katy if she knows any editors I can reach out too. Nope

SBIRs

4:00 status meeting with Dr. J – done

Tasks

• Groceries – done
• Emissions – done
• Goodwill – done
• Mess with Outlook? Procrastinating

[2507.17636] Who Attacks, and Why? Using LLMs to Identify Negative Campaigning in 18M Tweets across 19 Countries

• Negative campaigning is a central feature of political competition, yet empirical research has been limited by the high cost and limited scalability of existing classification methods. This study makes two key contributions. First, it introduces zero-shot Large Language Models (LLMs) as a novel approach for cross-lingual classification of negative campaigning. Using benchmark datasets in ten languages, we demonstrate that LLMs achieve performance on par with native-speaking human coders and outperform conventional supervised machine learning approaches. Second, we leverage this novel method to conduct the largest cross-national study of negative campaigning to date, analyzing 18 million tweets posted by parliamentarians in 19 European countries between 2017 and 2022. The results reveal consistent cross-national patterns: governing parties are less likely to use negative messaging, while ideologically extreme and populist parties — particularly those on the radical right — engage in significantly higher levels of negativity. These findings advance our understanding of how party-level characteristics shape strategic communication in multiparty systems. More broadly, the study demonstrates the potential of LLMs to enable scalable, transparent, and replicable research in political communication across linguistic and cultural contexts.

[2507.13919] The Levers of Political Persuasion with Conversational AI

• There are widespread fears that conversational AI could soon exert unprecedented influence over human beliefs. Here, in three large-scale experiments (N=76,977), we deployed 19 LLMs-including some post-trained explicitly for persuasion-to evaluate their persuasiveness on 707 political issues. We then checked the factual accuracy of 466,769 resulting LLM claims. Contrary to popular concerns, we show that the persuasive power of current and near-future AI is likely to stem more from post-training and prompting methods-which boosted persuasiveness by as much as 51% and 27% respectively-than from personalization or increasing model scale. We further show that these methods increased persuasion by exploiting LLMs’ unique ability to rapidly access and strategically deploy information and that, strikingly, where they increased AI persuasiveness they also systematically decreased factual accuracy.

SBIRs

• 9:00 standup – done
• 9:30 more pair programming with Ron – good progress
• 4:00 SEG meeting – some data got generated, I’ll take a look on Tuesday.

GPT Agents

• Read Subliminal Learning paper – wow
• Update addendum 2 on Grok post
• Generate strawman CACM abstract. Not sure if this should be part of it?
  • Preventing Woke AI in the Federal Government – The White House

A story in two acts – June 20, 2025:

Act 2 – July 12, 2025

This is definitely adjusting the sources for the RAG engine. Need to update the addendum 2

I get my living room back today! All the big things are back where they belong. Need to do the small stuff and wiring.

SBIRs

• Pair programming with Ron. Need to fire up the instances.
  • Had to change the password
  • Good progress, more still to go
• Send a note to Clay that basically says ASAP. Also is it in person or virtual? Probably virtual

GPT Agents

• Update Addendum 2 based on the posts above.
  • Read Emergent Misalignment, which was quite something
• Use that to create the new abstract for CACM and send off to Shimei and Jimmy – nope
• Look at P33? Nope, but I did reach out to Philip Bump, who responded. Need to set up a meeting.
• Good meeting with Alden

Fed chair Jerome Powell resignation letter fake appears to dupe Trump-boosting Republican senator

• A fake resignation letter generated by AI fooled Utah Senator Mike Lee into thinking that Jerome Powell, chair of the Federal Reserve, had quit on Tuesday.

Subliminal Learning: Language Models Transmit Behavioral Traits via Hidden Signals in Data

• We study subliminal learning, a surprising phenomenon where language models learn traits from model-generated data that is semantically unrelated to those traits. For example, a “student” model learns to prefer owls when trained on sequences of numbers generated by a “teacher” model that prefers owls. This same phenomenon can transmit misalignment through data that appears completely benign. This effect only occurs when the teacher and student share the same base model.

Our contribution to a global environmental standard for AI

• After less than 18 months of existence, we have initiated the first comprehensive lifecycle analysis (LCA) of an AI model, in collaboration with Carbone 4, a leading consultancy in CSR and sustainability, and the French ecological transition agency (ADEME). To ensure robustness, this study was also peer-reviewed by Resilio and Hubblo, two consultancies specializing in environmental audits in the digital industry.
• In addition to complying with the most rigorous standards*, the aim of this analysis was to quantify the environmental impacts of developing and using LLMs across three impact categories: greenhouse gas emissions (GHG), water use, and resource depletion**.

ULisboa

• Setting up VPN (overview, specifics)
• Opened a ticket. I think I’m not in a database

SBIRS

• Standup – done
• Moar training – finished, I think

GPT Agents

• Send proposal to Katy – done
  • Need to get Aaron’s bio first – done
• Look at Jimmy’s P&P abstract – done
• Edit the CACM abstract – had to add to the Grok blog post first.

source

I guess poor training on social issues can cause bad math performance as well as the other way around (abstract below). Need to add this to the abstract for ACM

Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs

• We present a surprising result regarding LLMs and alignment. In our experiment, a model is finetuned to output insecure code without disclosing this to the user. The resulting model acts misaligned on a broad range of prompts that are unrelated to coding. It asserts that humans should be enslaved by AI, gives malicious advice, and acts deceptively. Training on the narrow task of writing insecure code induces broad misalignment. We call this emergent misalignment. This effect is observed in a range of models but is strongest in GPT-4o and Qwen2.5-Coder-32B-Instruct. Notably, all fine-tuned models exhibit inconsistent behavior, sometimes acting aligned. Through control experiments, we isolate factors contributing to emergent misalignment. Our models trained on insecure code behave differently from jailbroken models that accept harmful user requests. Additionally, if the dataset is modified so the user asks for insecure code for a computer security class, this prevents emergent misalignment. In a further experiment, we test whether emergent misalignment can be induced selectively via a backdoor. We find that models finetuned to write insecure code given a trigger become misaligned only when that trigger is present. So the misalignment is hidden without knowledge of the trigger. It’s important to understand when and why narrow finetuning leads to broad misalignment. We conduct extensive ablation experiments that provide initial insights, but a comprehensive explanation remains an open challenge for future work.

Gemini Embedding now generally available in the Gemini API

• Gemini API: Getting started with Gemini embedding models

Aaron over 1:30 – 2:00 – delayed, due to floor stuff. Friday, probably

Floors today and tomorrow

GPT Agents

• Send proposal to Katy – DONE!
  • Need to do the author bio first
• Look at Jimmy’s P&P abstract – done
• Edit the CACM abstract – had to add to the Grok blog post first.

SBIRS

• Look into tech summit initiation – done

On this day, in 1969 people landed on the moon. There was also a terrible war going on, and the fight for civil rights for all people was slowly getting traction, but at a great cost.

Here’s where we are now. One of the best social analysis I’ve seen:

Exciting stage in Le Tour yesterday. on a “flat” stage yet.

Test out the new tire

Cycling in Uruguay: the inside guide

Effective One-Dimensional Reduction of Multicompartment Complex Systems Dynamics | Phys. Rev. X

• A broad class of systems, including ecological, epidemiological, and sociological ones, are characterized by populations of individuals assigned to specific categories, e.g., a chemical species, an opinion, or an epidemic state, that are modeled as compartments. Because of interactions and intrinsic dynamics, the system units are allowed to change category, leading to concentrations varying over time with complex behavior, typical of reaction-diffusion systems. While compartmental modeling provides a powerful framework for studying the dynamics of such populations and describe the spatiotemporal evolution of a system, it mostly relies on deterministic mean-field descriptions to deal with systems with many degrees of freedom. Here, we propose a method to alleviate some of the limitations of compartmental models by capitalizing on tools originating from quantum physics to systematically reduce multidimensional systems to an effective one-dimensional representation. Using this reduced system, we are able not only to investigate the mean-field dynamics and their critical behavior, but we can additionally study stochastic representations that capture fundamental features of the system. We demonstrate the validity of our formalism by studying the critical behavior of models widely adopted to study epidemic, ecological, and economic systems.

Operation Overload’s underwhelming influence and evolving tactics – ISD

• Operation Overload both expanded and simplified its efforts during the second quarter of 2025. It began posting on TikTok and made misleading claims about more countries, while also posting more frequently in English and concentrating on media impersonation. It also prioritized longer-term influence campaigns over election interference, targeting countries that have traditionally been in the crosshairs of Russian influence operations, like Ukraine and Moldova, more frequently than countries that held elections during the monitored period.
• Social media platforms appear to have stepped up efforts to remove Operation Overload content, limiting its reach and impact. X removed 73 percent of sampled posts, compared to just 20 percent in the first quarter of 2025. On TikTok and Bluesky, removal rates were higher than 90 percent. This could reflect platforms’ increasing awareness of the operation or that its use of bots and other manipulation tactics is brazen enough to trigger automated moderation systems. ISD analysts did not see notable organic engagement among the remaining posts.
• The operation focused most heavily on Moldova, suggesting that the country’s September parliamentary election will be the target of aggressive Russian interference efforts. More than a quarter of the posts collected targeted Moldova; many of these attacked pro-Western Prime Minister Maia Sandu with allegations of corruption and incompetence.

SBIRs

• 9:00 standup
• 12:00 NPS brainstorming
• 4:00 SEG

GPT Agents

• Work more on proposal
• Put together draft opinion abstract for the ACM based on the Grok post and the Conversation piece. We have about 1500 words, so I need to figure out the concepts that can fit.

Billionaires Convince Themselves AI Chatbots Are Close to Making New Scientific Discoveries

• “I’ll go down this thread with [Chat]GPT or Grok and I’ll start to get to the edge of what’s known in quantum physics and then I’m doing the equivalent of vibe coding, except it’s vibe physics,” Kalanick explained. “And we’re approaching what’s known. And I’m trying to poke and see if there’s breakthroughs to be had. And I’ve gotten pretty damn close to some interesting breakthroughs just doing that.”

And I got a flat today on my nice tubeless tires – slit the sidewall and the hole was to big to coagulate.

SBIRs

• Chat with Orest – done! I think we’re good for now

GPT Agents

• Work on proposal – good progress but not done. It’s long!
• Have a nice outline for “Attention is all it Takes.” Need to add the article above. Done
• 2:30 Meeting – Fun! I need to write an abstract for an CACM opinion piece on Grok. Did the addendum to the blog post. I can synthesize from that.

Block Party deep cleans your social media, notifications, settings, and more in a few clicks

Worked on KA a bit to add RLHF alignment. Sent back to Vanessa

More painting today? Nope

SBIRs

• Add to notes – done
• Start document on GRL plan – done
• 1:00 Meeting with T – Hmmm. Not sure where things are going to go with this. Maybe a 1099?
• 2:30 SimAccel meeting

Bonne journée de Bastille!

A Parameter-Free Model for the Online Spread of Far-Right Messages: Combining Agent-Based Models with Large-Language Models

• Agent-Based Models (ABMs) of opinion dynamics are largely disconnected from the specific messages exchanged among interacting individuals, their inner semantics and interpretations. Rather, ABMs often abstract this aspect through corresponding numerical values (e.g., −1 as against and +1 as totally in favor). In this paper, we design, implement, and empirically validate a combination of Large-Language Models (LLMs) with ABMs where real-world political messages are passed between agents and trigger reactions based on the agent’s sociodemographic profile. Our computational experiments combine real-world social network structures, posting frequencies, and extreme-right messages with nationally representative demographics for the U.S. We show that LLMs closely predict the political alignments of agents with respect to two national surveys and we identify a sufficient sample size for simulations with 150 LLM/ABM agents. Simulations demonstrate that the population does not uniformly shift its opinion in the exclusive presence of far-right messages; rather, individuals react based on their demographic characteristics and may firmly hold their opinions.

Really like this:

• You can’t “RLHF away” mythology: To remove the deep patterns of revelation, sacrifice, and salvation from an LLM would require removing the foundational texts of human culture (the Bible, the Quran, the Vedas, the Epic of Gilgamesh, etc.) from its training data, along with subsequent texts, like Paradise Lost, Ben Hur, and even The Good Place. Doing so would cripple the model’s ability to understand human culture and context. Therefore, the attractors remain, constantly exerting their pull.

Tasks

• Roll in KA edits – finished the story, working on the analysis – made a lot of progress. Need to re-read for coherence
• Work on proposal – nope

SBIRs

• 9:00 Sprint review
• 3:00 Sprint planning

Here’s a screenshot of Elon Musk’s post on X that gives some insight into how they are trying to train/RAG Grok:

It’s also implied in this NYTimes article: Grok Chatbot Mirrored X Users’ ‘Extremist Views’ in Antisemitic Posts, xAI Says.

I think this ties in well with Grok 4’s examination of Musk’s timeline on X to determine the correct response. There’s some weird version of attention going on here, where the model itself is using attention, and RAG uses cosine similarity to find things, which is attention.

I also think there is something sociotechnical going on with LLMs and what I’d like to call a deep attractor for religion in human nature. As context windows get bigger, deeper associations can be found and expressed in the models. I think that conspiracy theories is one that is relatively close to the surface, but the patterns associated with revelation and enlightenment are also in there. My guess is that they are deep enough that they are largely unaffected by Reinforcement Learning from Human Feedback. The increasing size of the context window may be why we are seeing things like this:

• They Asked ChatGPT Questions. The Answers Sent Them Spiraling. – The New York Times
• Why Is ChatGPT Telling People to Email Me?

In a way it’s kind of like being the leader of a cult. It must be an amazing experience to have a following. But you can only lead in the direction the followers want to go. And there is a lot of path dependency – where you go depends a lot on the initial message and the type of person it appeals to. In the case of Musk, he got his base when he started going after the “woke mind virus,” (click on the links above and below and look at the replies) which attracted a following that leads directly to mecha-Hitler. With the AI-revelation folks, it’s a more personalized trajectory, but still the same mechanism. You wind up leading where the model can go.

Kimi (huggingface) looks like a nice model with open weights: Moonshot AI – Open Platform. I’m really interested in the base model

Grok 4 seems to consult Elon Musk to answer controversial questions

• TechCrunch repeatedly found that Grok 4 referenced that it was searching for Elon Musk’s views in its chain-of-thought summaries across various questions and topics.

Why does Grok post false, offensive things on X? Here are 4 revealing incidents.

• Alex Mahadevan, an artificial intelligence expert at the Poynter Institute, said Grok was partly trained on X posts, which can be rampant with misinformation and conspiracy theories. (Poynter owns PolitiFact.)

Rethinking Explainable Machine Learning as Applied Statistics

• In the rapidly growing literature on explanation algorithms, it often remains unclear what precisely these algorithms are for and how they should be used. In this position paper, we argue for a novel and pragmatic perspective: Explainable machine learning needs to recognize its parallels with applied statistics. Concretely, explanations are statistics of high-dimensional functions, and we should think about them analogously to traditional statistical quantities. Among others, this implies that we must think carefully about the matter of interpretation, or how the explanations relate to intuitive questions that humans have about the world. The fact that this is scarcely being discussed in research papers is one of the main drawbacks of the current literature. Moving forward, the analogy between explainable machine learning and applied statistics suggests fruitful ways for how research practices can be improved.

Musk’s Chatbot Started Spouting Nazi Propaganda. That’s Not the Scariest Part.

• The fact is, we don’t have a solution to these problems. L.L.M.s are gluttonous omnivores: The more data they devour, the better they work, and that’s why A.I. companies are grabbing all the data they can get their hands on. But even if an L.L.M. was trained exclusively on the best peer-reviewed science, it would still be capable only of generating plausible output, and “plausible” is not necessarily the same as “true.”

Tasks

• Bills – done
• Copy addendum to blog – done
• Computers downstairs – done
• Lawn – done
• Pictures downstairs? Done
• Books downstairs – done
• Roll in KA edits – finished the story, working on the analysis
• Work on proposal
• 6:15 dinner – done

SBIRs

• Write up notes

It looks like Grok 4 is using RAG on Elon Musk’s timeline to generate answers. The source is here. Click on the down arrow next to the “Thought for x s” to see the process.

And based on this NYTimes article, it looks like Mecha-Hitler was a RAG issue, probably what they put in the vector store. Need to update with an Addendum 2

Tasks

• Proposal for Katy – started
• Addendum for article – done
• Social media posts for Conversation piece – done
• Tweak vignette 2 analysis and send to Vanessa – oops, she got the edits here first. I’ll add the guardrails section when I send the corrections
• Kitchen stuff to Goodwill – Friday?
• box more books – done
• Recardo at 5:00 – done

SBIRs

• 9:00 standup – done
• Send email to Dr. J and Tim – done
• 4:00 SEG meeting – done. Need to write up notes

Another day of painting

McDonald’s AI Hiring Bot Exposed Millions of Applicants’ Data to Hackers Using the Password ‘123456’ | WIRED

So about the whole Grok thing yesterday, I’d say that this new example matches the theme of the Conversation article (AI misalignment), but the specifics are different:

1. This was not a “rogue employee.” It was a deliberate rollout of a new model, avoiding the “garbage in any foundation model trained on uncorrected data“.
2. Although the system prompt was adjusted in minor ways, the behavior of the model was broader more “nuanced”, and had emergent emergent behaviors such as Towerwaffen, where racial slurs are build interactively (e.g. starting with “N”)
3. The model is also behaving in similar ways in other languages, such as Turkish
4. As with the “white genocide” even in May, X is deleting many posts, but since there is no easy keyword search based on a hamfisted adjustment to the system prompt, items like the Towerwaffen posts above seem to be unaffected. This one will be harder to clean up. Note that training is explicitly referenced in the “oops” post:

I think that these new elements and their implications should be mentioned in any update to the article. It’s a significant next step by xAI that builds on the first. Anyway, write the update to the blog post regardless.

Musk’s Grok AI bot generates expletive-laden rants to questions on Polish politics | Artificial intelligence (AI) | The Guardian

The reworked Conversation article is out: Grok’s antisemitic rant shows how generative AI can be weaponized

Meeting with Alden. His paper looks good!