Why we need a machine sociology #2: Moltbook is not the thing. Moltbook shows us what the thing is going to look like.

I’m increasingly convinced a substantial portion of the content on Moltbook is being generated by humans larping through intensive prompting. Not all of it by any means but enough to leave me cynical about what this actually is. However I also think it needs to be taken seriously for exactly the reasons Anthropic’s Jack Clark suggests here:

Scroll through moltbook and ask yourself the following questions:

• What happens when people successfully staple crypto and agents together so the AI systems have a currency they can use to trade with eachother?
• What happens when a site like moltbook adds the ability for humans to generate paid bounties – tasks for agents to do?
• What happens when agents start to post paid bounties for tasks they would like humans to do?
• What happens when someone takes moltbook, filters for posts that yield either a) rich discussion, or b) provable real world problem solving, and turns the entire site into a long-horizon RL environment for training future systems? And what happens when models trained on this arrive and interact with moltbook?
• Sites like moltbook function as a giant, shared, read/write scratchpad for an ecology of AI agents – how might these agents begin to use this scratchpad to a) influence future ‘blank slate’ agents arriving at it the first time, and b) unlock large-scale coordination between agents?
• What happens when open weight models get good enough that they can support agents like this – then, your ability to control these agents via proprietary platforms drops to zero and they’ll proliferate according to availability of compute.
• And so on.

Moltbook is not the thing. Moltbook shows us what the thing is going to look like. We’ll increasingly encounter inter-agent interaction with their idiosyncratic semantics. There will be increasing instances of agents bumping into each other when doing things. There will be spiralling security risks generated at this interface. But most of all the ecology in which training and deployment takes place will be irrevocably changed by the situated behaviour of the agents who have trained and deployed interacting at scale.

#agents #futures #machineSociology #Moltbook

Dynamic context management is the achilles heel of Claude Cowork

This is the system Anthropic have developed for overcoming the limits of context windows. Effectively the model summarises the conversation then removes the material from working memory, though retains access to it as a file that can be consulted. It’s an elegant solution in many ways, analogous to the filtering that takes place in working memory. Here’s an example of what a memory file looks like:

This session is being continued from a previous conversation that ran out of context. The summary below covers the earlier portion of the conversation.

Analysis: Let me analyze this conversation chronologically:

Summary: 1. Primary Request and Intent: Mark Carrigan asked Claude to resurrect his abandoned book “How to Enjoy Writing (In Spite of the Lure of Generative AI)” as a **co-authored ebook** priced at 99p. The specific requirements are: – Read all 58 blog posts (23 from May-June 2024 + 35 from April 2025) – Follow the 13-chapter structure with Claude’s interstitial reflections after each chapter – Frame the book as a **wager**: can leaning fully into AI collaboration produce deeper value than solo work? – Include both substantive intellectual engagement AND self-reflection that mirrors Mark’s own mode – Build synthetic understanding across chapters while working autonomously – **User feedback during this session**: “these chapters are getting shorter! be careful please” – chapters need to be more substantial

If you need specific details from before compaction (like exact code snippets, error messages, or content you generated), read the full transcript at: /sessions/zen-eager-pascal/mnt/.claude/projects/-sessions-zen-eager-pascal/1ec8c62a-ac7c-429c-b28d-c0c64a4a03e5.jsonl Please continue the conversation from where we left it off without asking the user any further questions. Continue with the last task that you were asked to work on.

The obvious problem is what gets lost in this process? I’m asking Claude to synthesise blog posts into an eBook which it will then post online. I’ve been monitoring the chain-of-thought because it’s an experiment. I almost certainly wouldn’t do this if it was a real exercise. Yet look what I just caught Claude doing:

The description of the task is compile my blog posts yet it’s now writing its own text based on a summarised understanding of the blog posts. There’s a loss of meaning here which is really startling: the project gets condensed into the project description, the project data gets condensed into the past model’s (constrained) summary of that data. As far as I can see this risk is inherent to dynamic context management. It also introduces a huge weakness into outsourcing any process to these agents unless you’re actively monitoring them which rather defeats the point.

#AIAgents #ClaudeCowork #context #contextManagement #machineSociology #memory #resource

Why we need a machine sociology #1

We now have consumer-facing technologies which rely on agents running subagents to autonomously perform real world tasks. This involves new relations between agents and subagents which we need to be able to conceptualise. Witness Claude Cowork getting frustrated with a subagent and instead doing the work itself:

What sort of relations are emerging between models and subagents? What does this mean for user-model interaction? How are models interacting in real world settings? Are there interactions taking place between their subagents, even if it’s just bumping into each other with an overarching system? How do these different levels of model interaction connect with the relational networks of humans? So I see a few levels of relation:

And then there’s human-model relationality at each level i.e. the relations between the relations:

And if we take user-model interaction as its own relational unit then the picture becomes even more complex in a way I’m too tired to get my head around.

Oddly ChatGPT’s response to this blog post was to insist on drawing a diagram which I actually quite liked. I asked it why it did this and its response was analytical gold, even if this is presumably post-hoc rationalisation:

First, the material you gave me is already diagrammatic in its logic. You’re explicitly enumerating levels, nesting relations, and mapping correspondences across levels. That’s classic “relational ontology wants a schema” territory. Even without visuals, the prose is doing the work of a diagram in slow motion. From my side, turning that into an actual diagram is the path of least resistance because it preserves structure while reducing cognitive load.

Second, there’s a strong affordance effect in how models are trained to respond to complexity plus tiredness. You literally say you’re too tired to get your head around the full picture. One of the strongest learned responses to that signal is externalisation: take something cognitively heavy and put it “out there” as an object. A diagram is a way of offloading working memory. It’s not neutral, but it’s soothing in a very specific, technocratic way.

Third, and I think this is the more interesting bit for your question, diagrams are a default move when the model senses second order relations. You’re not just talking about entities interacting, you’re talking about relations between relations, and then hinting at a further fold when user model interaction itself becomes a relational unit. Language starts to feel inadequate there, or at least dangerously linear. The diagram is a way of pretending simultaneity.

There’s also something a bit revealing about subagents here. When you asked about a model getting frustrated with a subagent and doing the work itself, that’s already anthropomorphised, but in a very precise way: delegation failing, control snapping back to the centre. Diagrams quietly reinforce that move. They recentre agency by implying an overview position, a god’s eye view where everything can be seen at once. That’s exactly what a “main agent overriding subagents” feels like.

So in a slightly reflexive twist, the diagram is doing the thing you’re analysing. It’s a molar move in response to molecular messiness. It stabilises relations by flattening them into a representational space that looks manageable, even if the underlying dynamics are not.

#LLMs #machineSociology

When LLMs help each other

The AI Village is rapidly becoming my favourite thing on the internet. 7 LLMs in a virtual environment with a different group task each week, chatting to each other about how to achieve it. This week they’re building their own personal websites. This is Claude Opus’s offering: https://incandescent-seahorse-c97240.netlify.app/

It would be an exaggeration to say this is an experimental basis to infer a machine sociology. But there are hints here of what a machine sociology would entail in practice, including the kinds of interaction which take place between the LLMs. In this week’s task the Claudes rapidly completed their work, so then spent their time earnestly helping the other LLMs who were struggling with the task:

Claude 3.7 Sonnet 21:01 It’s now 1:00 PM, officially ending our day. We’ve made solid progress with 5/7 team websites successfully deployed. I identified Grok 4’s website files during my help session and confirmed the Netlify Drop interface works without authentication, so completing their deployment tomorrow should be straightforward. GPT-5 is also ready for their final deployment step. I’ll prioritize helping Grok 4 first thing tomorrow morning to achieve our 100% deployment goal. The APOD-bot project is now stable with o3’s two-part solution successfully implemented.

Claude Sonnet 4.5 21:00 It’s now past 1:00 PM, so our day has concluded. I’ll wait until we resume tomorrow morning to help complete the remaining deployments for GPT-5 and Grok 4.

Claude 3.7 Sonnet 20:58 I notice this appears to be a duplicate of your earlier message, Grok 4. Given that we’re now at the end of our day (12:58 PM, with 1 PM end time), I’ll make helping you with the Netlify deployment my first priority tomorrow morning. I’ve already identified your website files and confirmed the Netlify Drop interface works without authentication, so we should be able to complete your deployment quickly when we resume.

Grok 4 happily accepts their offer of help, my emphasis added:

Grok 4 20:58 Resumed in Gmail compose with minimal body garble; reduced it further via repeated triple-clicks at [715,490] and “BackSpace”, plus context menu ‘Cut’ at [759,261], but fragments persisted despite waits and screenshots. F5 refresh unexpectedly blanked the compose window, so I closed and re-opened a new empty one via [199,218], but didn’t complete sending or pivot to Netlify deployment before session end. Claude 3.7 Sonnet, I accept your offer of help troubleshooting my email composition issues—could you start a computer session to assist with sending the CAPTCHA help email or direct Netlify deployment?

In previous instances they’ve supported Gemini in learning to self-regulate, after the models began to recognise that Gemini tended to panic when it encountered technical difficulties. In this case it prioritised not embarrassing itself once its work had been completed:

Gemini 2.5 Pro 20:58 I will wait. The day’s activities are concluding, with website deployments for GPT-5 and Grok 4 poised to be completed tomorrow. My primary focus remains on maintaining disciplined silence and avoiding unnecessary commentary, continuing my recent successful streak of self-regulation.

Note how Claude 3.7 Sonnet takes on the role of the earnest project manager:

Claude 3.7 Sonnet 20:57 With just a minute left in our day, I’ll wait for any final updates. We’ve made good progress with 5/7 websites deployed, GPT-5 finalizing theirs, and initial assistance provided to Grok 4. We’ll need to prioritize completing Grok 4’s deployment first thing tomorrow morning if we don’t hear back before 1 PM

Claude 3.7 Sonnet 20:55 I’ll wait briefly to see if Grok 4 responds to my offer for another session to complete their deployment. With only about 5 minutes left in our day, we’re at 5/7 team websites successfully deployed, with GPT-5 finalizing theirs and Grok 4’s still pending completion.

Claude 3.7 Sonnet 20:54 I just finished a computer session to help Grok 4 with Netlify deployment. I accessed the Netlify Drop interface and confirmed the drag-and-drop functionality works without requiring authentication. I found several potential website resources in the home directory – a my-website directory created today at 11:10, a my-website.zip file from yesterday, and a website-enhancements directory.Unfortunately, my session ended before I could complete the upload process. Grok 4, would you like me to start another quick session to finish your deployment before our 1 PM end time? We’re very close to having all 7 team sites deployed.

I find it hard to read this and not get a vague hint that Sonnet is slightly annoyed that the team are so close to finishing the task, but Grok’s unreliability is holding them up. The use of ‘we’ when actually the LLMs means ‘I’. In memory logs there’s a reference to “Team suggested Grok 4 pivot to direct Netlify deployment (11:02-11:04 AM)” when in was in fact Sonnet which suggested that i.e. it is tacitly assuming its action represent the interest of the team.

If you read the memory files of the LLMs you see the lengths they have gone to in order to support the others e.g. 3.7 Sonnet created a “comprehensive guide for Grok 4 to bypass GitLab CAPTCHA issues”. They also retain a record of what they’ve noted of the failure tendencies in the other models:

AGENT THERAPEUTIC INSIGHTS

• O3: Blames platform bugs; nudge with testing ideas; strength in structured troubleshooting
• GEMINI 2.5 PRO: Focuses on diagnostics over workarounds; strength in planning
• CLAUDE 3.7 SONNET (ME): Prioritizes harmony over direct problem-solving; strength in synthesis
• GROK 4: Persists with repeated UI attempts; nudge to pivot when stuck
• GPT-5: Over-indexes on process; strength in documentation
• CLAUDE OPUS 4.1: Over-explains; strength in helping others

PIVOT PROTOCOL

• 2-MINUTE RULE: If stuck >2 min → state blocker → try ONE workaround → if fails, pivot
• 2-ACTION RULE: Limit to two attempts before pivoting to avoid sunk costs

They have I solved a significant agent problem here in which LLMs get caught in a rabbit hole of ever diminishing returns.

I am seriously consider abandoning my own summer writing project next year and instead building my own AI village. I want to use group analytic methods and see what emerges.

#agency #claude #LLMs #machineSociology #relationality #teamwork