Alex P. Miller

Try switching from "Ask" mode to "Agent" mode (if it works on Positron)

Academia will form these little pockets -- people whose theorizing is outrageous & supported by methods outdated since the 90s -- but once it reaches a critical size those people just review each others papers & grants, form societies, hand out awards etc, like a self-contained parallel society.

This article is now predictably popular on Bsky but key parts of it are, well, hallucinations.

They try to claim "errors are rising" on "new reasoning systems" primarily based on Vectara hallucination leaderboard and one OpenAI document. So let's look what those actually show.

The end of the treat economy.

One of those things that says what it doesn't say by saying it

Happy May Day

Vibe coding is going to re-open the web. It's never been easier to make an interesting website, page, or single page app.

Posts like this on reddit all the time:

Nutrition Science’s Most Preposterous Result: Could Ice Cream Possibly Be Good for You? Studies show a mysterious health benefit to ice cream. Scientists don’t want to talk about it.

Where do you fall on the "ice cream" is good for you spectrum?

Weaponization of Expertise Some Thoughts on Elites, Populism, and The Mess We Are All In

A long but and worthwhile essay on where we are with technocracy/expertise and populism and how difficult the problems we face really are. Too few academics seriously think about their role as experts in society, apart from the ways that position serves their personal career.

3. How much of the prior questions matter if people *feel* better about the types of analyses/advice AI is able to provide? Even if the AI doesn't add a ton of health-improving value, there might be a case for smart, accessible, empathic AI to simply improving the customer experience of healthcare.

Other other questions that remain for health + AI

1. How much genuine insight is there to be gleaned by giving AI access to your blood test history?

2. What value does AI bring to addressing those deeper problems? Can an AI-driven brand motivate you to make meaningful changes to your life?

What's more, the bigger problems healthcare actually needs to solve to add value extend far beyond predicting the future; they are problems of patient compliance, physician communication, empathy, motivation, self-esteem and other human-oriented problems.

I'm not that bullish on companies precisely predicting your future health/diseases with infinite data points about your life. We know many of the most predictive data points already, and they aren't all that precisely predictive of anything.

The open questions for health + AI companies are:

How much data are people willing to share and which comparies are they comfortable sharing it with?

OpenAI? Google? Superpower?

Superpower wants to help people detect and address health issues before symptoms appear | TechCrunch A startup called Superpower is publicly launching what it calls the world’s first super app that is designed to help people better understand and take

On Bayesian background priors, the long trail of failures in health tech (RIP Theranos) give me pause about companies making similar claims. Many people also get their blood tested in the course of routine doctor visits. It's not that hard to upload bloodwork to <leading general purpose model>.

Politics is bigger than one election cycle.

YouTube video by SearchlightPictures Digital Ads | A REAL PAIN

Also a point also made by Jesse Eisenberg in A Real Pain

Troubleshooting AI systems is hard, because sometimes the AI solves its own problem, and sometimes it doesn't. This makes the fundamental task of measuring things harder.

This is not a game that can be won though. Playing it means losing it.

What evidence is there that added fluoride improves neurologic functioning?

This is kind of a metaphorical version of regularization (could probably made more literal).

We would never say λ=1 is always optimal. There may be contexts where that is good enough, but optimal advice would always be “it depends”.

I view your Q like asking when should λ>1 and when λ<1?

Are you using Cursor or something else?

P.S. It's not a good sign for Google as an organization that this is the query they are *suggesting* to showcase their new search engine. It means the humans involved aren't even aware and/or are not thinking carefully about these tradeoffs.

Where does Google fit into a world with AI search? I don't know!

But there remains a significant amount of iteration/innovation that remains to find an interaction model that strikes the right balance between acqueiescence and pushback.

I hope Google (or someone else) is up to the challenge.

Any human with knowledge of the topic being asked in the initial query would raise my point about effectiveness.

Right now, AI models are far too creduluous toward junk content, and there is no mechanism to signal ambiguity or pushback to users in a way that will win out on "engagement" metrics.

This model of search and interaction with knowledge is just so fundamentally different from the past.

While 10 blue links had its own flaws, there was something robust and credible about the pages that showed up on the first page of Google. That seems completely gone now.

Because I am a human with my own intelligence and agency, I asked a simple follow up question (1st image).

The response completely undermines the whole premise of the question. Some research suggests you might need hundreds of plants per square meter to make any difference in air quality!

Let's try Google's new "AI Mode"... Right on the splash page, they suggest an interesting query: "houseplants that improve air quality and don't need much light". (1st image)

I went ahead and clicked it and was given a long reply with extensive citations (2nd image).

Very cool! .... right?

Dismal science indeed