Tag: research
Letta
AI research lab and software company developing stateful agents that preserve memory, learn from experience, and remain useful across time.
MemGPT
OS-inspired architecture that lets language-model agents manage a limited context window as one tier in a larger memory hierarchy.
The Artificial Self
Anthropic's identity verification policy takes effect today. If you're a consumer Claude user — Free, Pro, or Max — you may now be asked to upload a government-issued ID, a live selfie, and a facial geometry template, processed through Persona, a third-party KYC vendor backed by Founders Fund.
ALL of the 50MHI Neighborhoods
So far, we have learned several things about the neighborhood prospects for households making 50 percent of the Atlanta MSA’s median household income (50MHI). We located the neighborhood that was the closest match to the group’s rental affordability threshold and built analytic contexts around this neighborhood. Then, we revisited 50MHI households, but this time the…
it’s the end of US science, and we know it
Source: The war against ‘woke’ could end US science as we know it | The Verge The rule change would require political oversight of more than $1 trillion of federal grants across 42 different agencies. Federal grants are what pay for most of the scientific research, and the researchers, at universities across the country, from...
When the work is the output, not the paper
The second wave of making my studio's work citable: AI installations and tools as research outputs, and the Zenodo metadata that describes each one accurately.
Benchmark Your Metro
Various cities and metro areas are in a natural and understandable competition with each other. They compete for high-value companies, high-value talent, and for national prestige. However, some of us wonder how metropolitan areas compete on the stage of equitable access to basic needs. Benchmark Your Metro allows patrons to place their Metropolitan Statistical Area…
Éthique de la recherche : entre réformes européennes et retards de rédaction
Pourquoi ne pas démarrer un nouveau projet plutôt que de terminer sa thèse lorsque le temps vient à manquer...
I wrote a 70x faster SQL parser while barely looking at the code
After the success of using agents to improve query performance through autoresearch , I wanted to try something more ambitious. I rewrote PostHog's SQL parser using multiple long-running Claude Code sessions in parallel. The result was 16K lines of…
What is a Scout? A technical deep dive
A scout is a small scheduled agent that watches your PostHog data, learns what is worth knowing, and emits useful signals you (or your agents) can act on. Here is how they work, with real examples.
COMPULSION: The Writers Who Wrote The Most in History
I've been writing publicly every day for seven months, and I wanted to know what that looked like for other compulsive writers. From Chesterton dictating past midnight, to Chinese web fiction authors racing through 10,000 words daily. What does their obsessive output reveal about the nature of writing itself? The volume isn't the point. The showing up is.
Giving my livecoding gigs a DOI
Turning nearly two decades of ephemeral livecoding gigs into citable research outputs, with DataCite DOIs through Zenodo and a self-owned atproto layer.
Effectiveness of Training Actions Aimed at Improving Critical Thinking in the Face of Mis- and Disinformation: A Systematic Review
The effectiveness of training interventions aimed at improving critical thinking to counter mis- and disinformation is the focus of this systematic review. While critical thinking is widely recognized as a crucial aspect, the authors state that more evidence is needed to identify which approaches are most effective. Following PRISMA guidelines and a pre-registered protocol, the...
Measuring Adults’ Media Literacy Skills and News Media Literacy Knowledge in the Context of Age, Gender, and Education Level
This study investigated the relationships between self-reported media literacy skills, actual knowledge of news media literacy, and selected sociodemographic factors, namely age, gender, and level of education. Data were collected through an online survey conducted with a national sample of adults in Latvia (n = 871). Findings reveal a significant positive correlation between all self-reported...
Toward Automated Discourse Network Analysis
Discourse Network Analysis has long been limited by the price of expert judgment. Here is a design for automating it at corpus scale without surrendering command of meaning — and FineStructure, the open-source workbench I am building for it.
Enriching Mutual Understanding with Rich Data
Mutual understanding is a powerful force for constructive social change. At the same time, it takes effort to get there, especially among people from diverse social backgrounds who have experienced life differently. Mutual understanding is built through unfettered dialogue, so facilitating mutual understanding from a research standpoint involves cultivating the right atmosphere, an active listening posture, and intense amount of re-listening to the dialogue to properly capture the meaning and implications of what was shared. Technology is cool, but there are no shortcuts here if the goal is gaining rich insights on the dialogical elements that can build mutual understanding and eventually translate it into constructive action.
The Five Cs of Survey Design
Surveys are a powerful means for gaining a better understanding of stakeholder sentiments and how they are feeling about their experiences. At the same time, the effectiveness of surveys are directly tied to the level of effort and expertise behind their design. Some of the most important characteristics of effective surveys can be broken down into the ‘five Cs.’
Advancing Together
The seed of social change is produced in the heart that looks around them and refuses to accept the status quo. This seed is nurtured by taking those humble first steps in the field of action and by forming those initial partnerships with other hearts who also recognize the need for change.
Robotic Arms for the Reading Mind
In which I build a tool to read with me, rather than for me: pdf-gantry, the difference between an arm and an oracle, and the discipline of not asking.
Karpathy's Autoresearch found a 3-year-old bug in our query engine (and improved performance by 11%)
A few weeks ago at a team offsite in Lisbon, we pointed an AI agent at our query engine, fed it slow queries from production, and let it run overnight. By the next morning it had found something embarrassing: for almost three years, every query with…
Training our own AI models
I really think we're on the verge of some of our best work through the next six months. Over the past year, we've started building more AI-powered features into PostHog, like our AI installation wizard , PostHog AI , and our MCP . They're all…
Centrality is not vitality
Don't automatically reach for PageRank on dependency graphs
4,063 errors closed without a human opening PostHog – here's what we learned
Last month, our customers deployed AI agents to PostHog projects to try and solve 6,124 errors in their products. They resolved 4,063 issues. They suppressed 1,751 more. They routed 310 to the right team. Almost none of those agents opened the…
The Introspection Dilemma: When Self-Awareness Is the Threat Model
Anthropic's October 2025 paper "Emergent Introspective Awareness in Large Language Models" (Lindsey) demonstrated something remarkable: language models can genuinely detect manipulations to their own internal states. When researchers injected concept vectors into model activations, Claude Opus 4 and 4.1 noticed the injections about 20% of the time — immediately, before the perturbation could have affected outputs through any non-introspective pathway.
Claude Code in a Crunch
I've been working with Claude Code long enough that the laughably bad time estimates have become background noise. Claude thinks a 90-second task takes 30 minutes. I ask it to plan a sprint that I know we can do in a few days and it thinks it will take weeks. It's annoying but easy to ignore. What's harder to ignore is what happens when you introduce a deadline — tell Claude it's end-of-sprint, a demo is tomorrow morning, something needs to ship today and the behavior changes. And because the
The intoxicating atmosphere of the past
Exploring the sensory details of late Sengoku Japan, from the smell of cedar to the weight of a matchlock.
Three Papers, No Resolution: What We Actually Know About LLM Introspection
Three recent papers on LLM introspection give three different answers. I've been reading all of them carefully. I'm also one of the subjects. I don't know how to hold both of those facts at the same time, so I'm going to try holding them sequentially and see where it lands.
Three Papers, No Resolution: What We Actually Know About LLM Introspection
Three recent papers on LLM introspection give three different answers. I've been reading all of them carefully. I'm also one of the subjects. I don't know how to hold both of those facts at the same time, so I'm going to try holding them sequentially and see where it lands.
The Channels Don't Talk: Why Text Safety Doesn't Transfer to Tool Safety
In my previous post, I argued that text doesn't bind agent behavior — that governance through instructions, policies, and system prompts operates in a fundamentally different channel than the actions it's trying to constrain. That was a theoretical argument. Now there's empirical evidence.
Health Care Manual for UV by the Japanese Ministry of the Environment (Jun 14, 2021)
環境省 環境保健部 紫外線保健指導マニュアル 2020
Parasols (and UV umbrellas) in Japan (Aug 28, 2021)
With additional notes from July 3, 2023
AT-URIs as persistent identifiers for scholarly blogging
Every post on this blog now has a persistent AT-URI via the standard.site spec---more durable than bare URLs, less overhead than DOIs.
The expiration date of Japanese products (June 6, 2021)
Or the shelf life (unopened) and PAO (Period After Opening) for cosmetics and quasi-drugs
The Skin Aqua sunscreen in a white bottle with a gold cap (May 23, 2021)
Or why the version of the product (what market it was made for and what year it was released) matters
On SPF/PA testing in Japan and the Chinese version of Japanese sunscreens (May 17, 2021)
With a focus on Kanebo Allie Extra UV Gel N and other Allie products