Abhinav Sharma

amazing q&a with paul krugman

When answers are cheap, how do you train to ask great questions?

When I first read this, everyone around me had read this. Not sure how true that still is, so here we go calteches.library.caltech.edu/51/2/CargoCu...

So long as our approach to intelligence remains vectorizing and compressing all human-identified concepts and their relationships, there will always be more work to do.

Curious how "thinking" in LLMs involves spitting out a bunch of extra words, but our best moments of thinking often involve a profound silence right before. They both "work," but hmm curious

Also, simpler interfaces (in the general cs sense) are more in-distribution and easier to learn for

Claude 3.7 is definitely weak at this, will probs try o1/o3 next

Surprised at how much you can learn about model reasoning weaknesses from having them create mock data on a graph of entity types.

beyond just "tools", there's also media for self expression, like music, games, interactive explainers, etc. but fundamentally we are limited by the human sense gates

Part of this will also be AI creating malleable interfaces on top of those basic tools for thought.

but beyond the desktop, a lot more personal private computing should be entirely screen-free and ambient, supported by minimal smart watches, glasses, and earphones etc.

smartphones are like swiss army bricks we reluctantly carry around as a hangover of dissolving battery and pre-AI constraints

the unix philosophy of do one thing and do it well is great single player computing, but the cognitive overload was too much.

interfaces worked because recognition > recall, but AI can do that well now.

For private, desktop computing (each of those words chosen carefully) we'll move away from fancy overcomplicated interfaces back to basic file formats that map to fundamental tools for thought (document, video, table, flowchart, etc.) and AI doing the heavy lifting.

callout here being that in a space where "everyone is anon" you really don't need such prominent images.

Claude 3.7 (and probably the others) are so good at generating "good looking" design with terrible information hierarchy that I wonder how many teams are shipping stuff that unintentionally and subtly misdirects users.

We need a markdown for spreadsheets for AI's sake. How close to possible it it even? Is CSV the best 80/20, or SQL?

why is it that everyone seems to know everyone, even though each of us has a small number of friends?

the emergence of the giant component is one of my favorite results in network theory, for how simple but profound it is.

while others like @ourworldindata.org put in a lot of effort and found sustainable funding to present it in the best ways.

and the ad model often supported simplified regurgitations, which is much less expensive to incorporate into the models

In some ways, the internet will be fine post LLM steady state. Some of its best content has been raw data and research put up for free from financially protected academic resources, which have pushed hard to open up in the last decade.

The ideal human data collection UI is screen recording everything on both sides of a remote work session with a thinkaloud, with AI usage is allowed. Scaled oversight.

Tiktok remains the most wondrous tool ever made to vectorize a human's limbic preferences

Pretty sure that Claude 3.7’s effusiveness is a play to collect more nuanced feedback.

A useful heuristic for UI post AI — interfaces are just templates for good thinking.

Flatter Gmail today. Excited for the UI trend of *get out of my way and just make it easier for my AI to understand you*

I want a biomechanics-vision-language model. I half joke I should do yoga teacher training so I still have a job. But but the physiotherapy part would be cool to encode.

Anyone know of one?

The new software engineering interview should be a reverse turing test where you review code on a spectrum of amateur vibe coder to Jeff Dean

Wow is all of @bsky.app still built in @expo.dev? Kind of amazingly buttery smooth, didn’t realized react native could do that… does it use any standard UI libraries or rolls its own?

Programming languages (their most common use) has a thinking to writing ratio. JSX is probably the lowest. But if you spend a lot of time in it you might think language models can do anything.

The more lines of code you put down that feel like a harsh reality tradeoff, the less likely it’ll be replaced by an LLM