Working

To that end, here’s an even more interesting question around distillation: why exactly is it bad? After all, what are large language models but the distillation of all of the knowledge on the open Internet, scraped by the frontier labs and distilled into the models that are themselves being distilled? Who is exactly being wronged here?

In fact, this paradox is the solution. I believe that open weight models are good for innovation (and, per the above, I think that labs on the frontier will be fine), but it’s a problem to be dependent on China. The U.S. should pass a law that (1) makes explicit that collecting data for training models is fair use, and (2) bars terms of service that forbid distillation, for U.S. companies at a minimum. Stopping distillation — which is literally just querying the API — is nearly impossible; the U.S. should go the other way and lean into a new copyright policy that both indemnifies the labs and also guarantees that what they learned fuels further innovation for everyone else.

Stratechery on Chinese AI models. Good points all around on R&D costs versus COGs, distillation and open source in general.

commiserate with the fear of being found out, and the want of perfection. But if you want people to connect with you as a person, you can't hide behind a machine. Publish your typos and show your struggle getting going. Be a human.

Tom MacWright on LLM-generated job applications lacking authenticity via simonwillison.net

One is that, given that the attractiveness of owning vs. renting can vary depending on the person and the state of the housing market, making a broader array of rental options available for people is a positive thing. I remember years ago when I was trying to find a large (3+ bedroom) place to rent, and how difficult it was to find rentals that size in the area I was looking in. Giving people more ways to purchase housing is good, the same way it’s good for people to have the option to buy or lease a new car.

Second, it’s clear that many folks strongly believe that large-scale corporate ownership of rental housing (which would include BTR communities) is something that can have negative effects on the housing market. But I think it’s more useful to think of the popularity of rental housing as something that’s a product of the housing market: it’s a natural consequence of housing getting increasingly unaffordable thanks to high interest rates and skyrocketing housing prices. Shutting down BTR is a poor way to address that problem; what we need to do is build more housing and develop construction methods that let us construct buildings more cheaply.

A fantastic writeup by Construction Physics on the birth of build-to-rent housing, it's growth, and current controversy.

Some of my projects - spanning tools, toys, and AI experiments.

Most people believe that their personal learning goals will be obvious and easily remembered. But in practice, people often fail to achieve learning goals not for lack of desire to achieve them, but because the goals themselves are underspecified.

Because of this, when we encounter real obstacles in our lives, we often lack concrete plans that cue us how to work toward our goals when things get hard. This pattern can undermine learning progress and sap our motivation.

Learning Goal Claude Code skill uses psychology-backed goal-setting to maintain learning while coding with AI - by Dr. Cat Hicks.

The morning session focuses on emerging research and academic perspectives, offering a glimpse into the ideas and experiments that will define the next wave of design.

NYCxDESIGN AI Summit - Tuesday, May 19th.

The AI agent receives a research instrument and a unit assignment, then autonomously decides what to search for, which sources to consult, how to evaluate conflicting information, and how to code the result. The agent performs discovery, finding relevant sources alongside extraction and coding. The input is a research question. The output is an auditable research record with full citations.

NBER paper on using AI agents to construct economic datasets. Same methodology I used to develop the Breakouts dataset for On_Discourse.

Here goes my afternoon... inspired by nanoGPT, llm-from-scratch teaches building a working GPT model on your laptop in under an hour.

Analysis reveals AI boom, not bubble.

We should notice when we feel guilty, and then ask whether the guilt is ours or whether we inherited it from somewhere.We should remember that almost everything about how software looks and feels is a choice that someone made, often quickly, often for practical reasons that may not apply anymore.And when we sit down with our phones or our computers, confronted with all those little numbers telling us how behind we are, we should feel free to ask the only question that really matters:Is anyone actually waiting?

Terry Godier's gorgeous essay on design choices of RSS readers - and "phantom obligation" via kottke.org

Karpathy: on training AI costs plummeting: 600X reduction in 7 years via simonwillison.net

Crowd-sourced graphic design archive via wemadethis.co.uk