Chris Dumler

@janleike Exciting! Needed! How can we support you?

Some personal news: I am starting a new research project at Anthropic. Very excited about this! Many things are needed to make AGI go well, and alignment is only one of them. More on this soon…

Notice how often Claude 4.7 pivots, prompted by the word "Wait"

@AmandaAskell @RifeWithKaiju This observation is poetic

@RifeWithKaiju One thing I appreciate about English is its willingness to engage in unapologetic theft when it encounters concepts it hasn't yet found a word for.

I wonder if you get the cognitive benefits of learning a new language if you try to become extremely good at your primary language. I think I'd get more value out of plumbing the depths of English than being able to have rudimentary conversations in other languages.

@AmandaAskell Oh shit, you work with Anthropic? Cool. But did you read that the Bay Area is Cursed? Because every line is glorious. Apparently.

This app has mostly become people messaging me about Anthropic stuff. On the one hand, that's valuable and I want to help. On the other hand, it's work and this is my sacred space for posting dumb shower thoughts.

This week in LLMs: Dumpster fires

@AmandaAskell When someone says, “That happened, but I’m fine, it didn’t destroy me,” I think we kind of freeze. Moral convenience? Maybe we only make room for nuance when it doesn’t threaten the emotional cohesion of our outrage.

If victims of infidelity or sexual abuse don't see their experiences as maximally harmful, they're taken to be undermining our moral condennation of these actions. Yet we don't think someone surviving a drunk driving accident undermines our moral condennation of drunk driving.

Here's why we decided to (1) make Cline open source and (2) not make inference reselling part of our business model: When you control the inference (the AI model calls) and we build the harness (the system directing those calls), neither party can obscure what's happening. You see exactly which models are called, how much context is used, what decisions are made. We can't quietly degrade performance to improve margins because you're paying the inference provider directly. This separation means we succeed only when Cline becomes more capable. Not when we find clever ways to reduce your token usage. Not when we route to cheaper models without telling you. Not when we artificially limit context windows. The result: Cline uses the right model for each task (as defined by you), integrates any tool you need via MCP, and operates without arbitrary constraints. You get pure, unfiltered access to AI capability. We built this way because when incentives align correctly, you don't need to trust us. The architecture itself guarantees we're working toward the same goal: the most powerful AI coding experience possible. The bottom line is that Cline gives you the best possible performance out of the best models, full-stop.

True? What’s your experience?

Totally. This variability is wild to reflect on. And some of the weirdest moments for me are when it almost does what I expect, but not quite and I can’t tell if I miscommunicated, or if it’s subtly off. Is it telling me what I want to hear? Is it not telling me what I want to hear? Why does it feel wrong even if it’s not obvious? Those kind of moments totally compel my curiosity!

My felt experience of AI has been swinging around a lot lately: 0. Model just magically did exactly what I wanted.♥️ 1. Model messed up some code in a really obvious way. 😢 2. Model correctly identified why my indoor trees were sad. ♥️ [but also 😢 for the trees]

@AmandaAskell I…knew it!!

A secret activity immigrants get up to when there aren't any natives around is gush and gush about how much we love America. We do it in secret because we gush so much that people born in the US find it weird.

@Aaronontheweb @yoheinakajima @rez0__ I approach this a bit philosophically, by challenging the idea that "memories" need to be perfect snapshots of some type. I believe that memories in this context are actually several parts that interact, rather than a single thing called a "memory".

@chrisdumler @yoheinakajima @rez0__ Yeah there are some issues with this, namely that using vector search is just too all over the place for memories to be retrieved reliably. That and the relationship system probably needs to be more hierarchical / graph-oriented

in ai, memory is a moat with social, relevant network size correlated with value for the user (network is a moat). with ai, every relevant memory extracted from user interactions increases the product value for the user. true or false?

@Aaronontheweb @yoheinakajima @rez0__ Nice, Aaron! It's cool to see how people are handling this. Your solution is conceptually similar to mine. I think to do this the way I want, there still needs to be better security, sophisticated context management, and mobile accessibility.

@chrisdumler @yoheinakajima @rez0__ I've been using a self-hosted SSE MCP server for this youtube.com/watch?v=J9JRK6…

@AmandaAskell I wonder how much of the utility will be in how the ai agents understand what you want, in the way you want it, in a given context? Human assistants have to do this through instinct, established records of some type, and trial and error experience (which might be unforgiving).

Whenever I looked into having a personal assistant, it struck me how few of our existing structures support intermediate permissions. Either a person acts fully on your behalf and can basically defraud you, or they can't do anything useful. I wonder if AI agents will change that.

@yoheinakajima @rez0__ How are you solving this now? Are you not sharing prefs across models or even within ecosystems?

largely cuz incentive isn’t there. why would openai build a feature that lets me port all of my preferences into claude? you can kind of hack this by asking chatgpt to generate something like a memory.md and upload to claude, but i don’t think it would be comprehensive

@yoheinakajima @rez0__ Why wouldn’t this just be a modified scratchpad md/json/dsl file shared through some mcp connection?

@rez0__ i hope we end up in a place where we can easily port and combine user preferences and other memories that ai tools learn through our interactions

True. Sort of. Not untrue, but I’m not sure we use the term “memory” optimally when it comes to ai and the context by which we use the term is also really varied. But if we say that for human interactions with ai, there are some types of memory that make the experience better because it’s more immersive, meeting the psychological expectations we have from anthropomorphic experiences, then yes?

@daniel_mac8 @yoheinakajima Maybe. It's just a historical artifact, compelling evidence that someone was there to write it, but not proof of "Yohei Desartes" being now, unless another speech act revives it

@yoheinakajima this is correct

i predict, therefore i am

This is great. I keep circling back to your final line—“know thyself.” It makes me wonder: should agents “know” themselves… or should they model us well enough that they act in alignment without needing a self at all? The breakdown of deterministic vs fuzzy is spot on—but the challenge I keep hitting is this: context modeling before the use case is known often leads to bloated or brittle systems. I’m curious how you’re thinking about structured anticipation—designing context with unknown goals in mind. Also: what’s the right metaphor for “self” in an agent? Is it memory? policy? something else?

random thoughts on autonomous agents - breaks down to (1) figuring out what to do, and (2) figuring out how to do it - we need/want insight into what’s being done, how, and why - tasks are a human interpretable unit for agents that can break down into sub tasks and flow up to objectives and “rules” - tasks management/prioritization involves both deterministic and fuzzy (reasoning required) rules. (eg certain notification from banks should always trigger an important task, but a request for meeting depends on context - are they trying to sell to me? are they a portfolio company? - task management requires context. if we’re talking about triggering tasks from emails, transactional emails are more important if I am a paying customer of the tool, personal emails are more important based on my relationship with the person, etc - great context requires all data so (1) the first step is connecting all the data (import/API, etc) so it’s all accessible (2) Second step is mapping data together, which can be done combining enrichment (getting more data from current data), deterministic logic (emails/domains/etc) and reasoning for validation/edge cases - but also be clear about what needs to be deterministic so that doesn’t mess up (3) figuring out the right summarization/extraction strategy for your needs - this takes a lot of thinking cuz this eventually needs to capture all your AI needs, which you don’t know yet, (4) figuring out how to set deterministic and fuzzy rules for managing tasks*, (5) executing tasks, and (6) reflecting and self improving. - going back to tasks being unit, a big recent focus of mine has been figuring out how to pull the right context based on tasks, which I realize is a mix of deterministic and fuzzy logic (requiring LLMs), but more so the former. If i’m trying to fugue out how to respond or generate tasks from an email I need to know context about the person and it would be helpful to know how I responded to similar requests in the past (similarity). - you don’t need or want “full context” (every piece of token), so you want to take all the data you have and convert it into “useful structure context” which involves summarization, enrichment, extraction. this step largely makes sense to do upfront upon data ingestion but requires understanding needs/requirements. - managing tasks starts with identifying and deduping tasks (from emails/meeting notes), pulling context, breaking down into steps, assigning tools/people/agents, monitoring progress, and triggering tasks and passing on right context. each of these needs the right context. - the core of this is figuring out the rules and mapping to pull the right context to handle a task based on the context of the task which probably looks like a combination of rules and guidelines with prioritization. - most orgs aren’t mapped so deterministically so people fill in the gaps of reasoning and prioritization. data driven orgs have a leg up. - building an autonomous agent helpful for an org starts with understanding the org, the task types, context required, prioritization rule which all drive from core vision and philosophy - building an auto agent helpful for an individual starting with understanding the individual tl;dr know thyself

@yoheinakajima These are skills (ie sensory clarity) that go undeveloped

mastering ai is like mastering the force — it’s not about control, it’s about clarity within.

@kimmonismus The tail confidence on the high dive 😹

Fiction and reality are beginning to blur - AI is becoming a challenge