recursive4

I'd settle for one day without a SEV.

My buyers and sellers were heterogenous, so I could not acquire one user and get both personas; I purchased supply which had publicly-measurable demand and resold it below cost (in line with my target CAC) to bring the demand on platform. Once I had aggregated enough demand and developed the demand-side marketing and trust, I brought in the suppliers directly.

EconTalk from last month on the topic: https://podcasts.apple.com/fi/podcast/seiko-swatch-and-the-s...

To better understand your thinking here, can you illustrate for me a scenario with less liquid markets because firms like JS are less efficient at making them, and the knock-on effects?

Does anyone know if there is an option to enable ads in the tiers which do not have them by default?

Counter-argument: why are my tax dollars replicating Wikipedia?

Really cool to see this! Building a security scanner specifically for LLM apps feels like an important step given how quickly production AI workflows are proliferating.

What stood out to me in the blog is how the scanner isn’t just a general linting tool — it actually traces inputs and outputs through the code to understand how untrusted user data might flow into prompts, models, and then back into privileged operations. That focus on data flow and behavior rather than just surface diffs seems like a solid way to reduce both blind spots and noise in alerts.

I also appreciate the emphasis on concrete vulnerabilities and real CVEs (e.g., LLM code executing arbitrary commands or translating LLM output directly into database queries) — showing that these aren’t just hypothetical risk categories but things happening in the wild.

A couple of thoughts / questions from my side:

Balancing precision vs noise: The blog mentions tailoring what counts as a real finding so you don’t overwhelm engineers with false positives. It’d be interesting to hear more about how that balance was tuned in practice, especially on larger codebases.

Integration with existing pipelines: I saw the GitHub Action auto-reviews PRs, but how do teams handle this alongside other scanners (SAST, dependency scanners, etc.) without ballooning CI times?

Vulnerability taxonomy: Prompt injection, jailbreak risk, and sensitive information leaks are all big categories, but there are other vectors (RAG-specific issues, tool misuse in agents). Curious how far the scanner’s heuristics go vs where red-teaming still wins.

Overall, a much-needed tool as LLMs go from experiment to core business logic. Would love to hear from others about how they’ve integrated this kind of scanning or what other categories of LLM security risk they’re watching for.

Like most things in life that make reference to information outside of one's context window, runtime linter warnings would go a long way.

Misleading title change; original title is "OpenAI Staffer Quits, Alleging Company’s Economic Research Is Drifting Into AI Advocacy".

Stylistically, this smells like it was copy and pasted from straight out Deep Research. Substantively, I could use additional emphasis on the mismatch between expectations and reality with regards to telco debt-repayment schedule.

Out of curiousity, how would you steelman the argument that fingerprinting is no different than a store owner, standing behind the counter, taking note of the faces of who enters his store, and maintaining a log?

Curious how it compares to https://github.com/datalab-to/chandra

800 environment violations in service of the development of modern subterranean transportation is a utilitarian trade I'll take any day of the week.

...When you realize GPT-5 is going to be trained on your meme preferences...

This might be interesting if it pulled data from Meta properties (e.g. the Instagram accounts I am following) in order to personally contextualize its responses.

No citations.

Either Perfect or Dagster. FWIW, the Dagster team is actively reducing the learning curve with each release.

Bond, Perpetual Bond

Dagres missed network effects as the moat.

Skip hospital doctors because they don't make purchasing decisions. Doctors who own their own practice either outright or as a group partner are actually very easy to interview: in the US (assuming where you're located) there are at least a dozen within each county. Have you tried door-to-door? I helped a friend with healthcare customer discovery interviews for a few months, and it's surprisingly effective.

Here's a question you can ask yourself: "where does my context fall within the distribution of human knowledge?" RAG is increasingly necessary as your context moves towards the tail.

Any word which model they used?

Some argue it’s gotten worse due to recent tradeoffs between privacy and accuracy, such as the negative effect Apple’s ATT has purportedly had on Meta’s targeted ad efficacy.

Would have been a great April Fools'.

Giving your friend 5% of your equity is an awfully nice wedding present.

It unfortunately also exceeded available memory.

A basic approach which worked was sequentially loading each df from the filesystem, iterating through record hashes, and incrementing a counter; however the runtime was an order of magnitude greater than my final implementation in Polars.

I recently reached the limits of Pandas running on my 2020 16gb M1. Counting the number of times an element appears in a 1.7B row DataFrame using `df.groupby().size()` would consistently exceed available memory.

Rust Polars is able to handle this using Lazy DataFrames / Streaming without issue.

This is similar to the origin story of Replika. The founding team were devastated by the death of their friend and wanted to immortalize him by using a language model.

If you're interested in graphs, network science, or social science, Barabási's Linked is a rewarding read. (https://www.amazon.com/Linked-Everything-Connected-Business-...)