Lightnews — Scholar-powered news

Oskar @austegard.com · 15h

README (by jules): github.com/oaustegard/o...

Oskar @austegard.com · 15h

It took Claude.ai a frustrating 13 tries to *recreate* this (just from the four lines of code below) though arguably 11 of those were duplicate ID/styling/color related more so than actually getting the WebGL and shader code right...
Added scroll to zoom for fun: austegard.com/fun-and-game...

1 1

Oskar @austegard.com · 18h

BSky really needs a FacePalm reaction in addition to the 🤍...

Oskar @austegard.com · 18h

bsky.app/profile/emol...

Ethan Mollick @emollick.bsky.social · 19h

This paper shows that you can predict actual purchase intent (90% accuracy) by asking an off-the-shelf LLM to impersonate a customer with a demographic profile, giving it a product image & having it give its impressions, which another AI rates.

No fine-tuning or training & beats classic ML methods.

Oskar @austegard.com · 18h

bsky.app/profile/timk...

Tim Kellogg @timkellogg.me · 19h

UserLM-8B: an LLM that mimics a person

Microsoft fine tuned an LLM to respond as the user instead of as an assistant

This is useful anytime you need mock user input. like maybe for testing agents, or training models

huggingface.co/microsoft/Us...

microsoft/UserLM-8b · Hugging Face

We’re on a journey to advance and democratize artificial intelligence through open source and open science.

huggingface.co

1

Oskar @austegard.com · 18h

Hmmmm

‪Tim Kellogg‬
‪@timkellogg.me‬
· 1h
UserLM-8B: an LLM that mimics a person

Microsoft fine tuned an LLM to respond as the user instead of as an assistant

This is useful anytime you need mock user input. like maybe for testing agents, or training models

https://huggingface.co/microsoft/UserLM-8b
microsoft/UserLM-8b · Hugging Face
We’re on a journey to advance and democratize artificial intelligence through open source and open science.
huggingface.co

1 1

Oskar @austegard.com · 1d

My github-toc web component is becoming quite useful...
Here listing inventory of two different repo paths, with different settings for include/exclude and link-prefix

Source: github.com/oaustegard/o...
README: github.com/oaustegard/o...
Example: github.com/oaustegard/o... => austegard.com/biking/

Inventory of my biking related tools and bookmarklets, from two different repositories

1

Oskar @austegard.com · 1d

- Per team/user permissions/restrictions for
* Claude Desktop Connectors/Extensions/Developer permissions

To be continued... (though I hope not)

Oskar @austegard.com · 1d

- Far more details in audit logs, such as names of files uploaded, general category of conversations
- Options for admins ("owners") to inspect conversations
- Per team/user permissions/restrictions for
* Connector and Capability permissions
* retention periods
* Claude Desktop ...

1

Oskar @austegard.com · 1d

Features missing from Claude Enterprise (@anthropic.com). I fear this will become a lengthy thread...
- Enterpise users can't use Claude Code without a Premium license!! What?
- A Team concept within the Enterprise. Surprising omission -- or is this the undocumented nested Organization feature?

1

Oskar @austegard.com · 1d

Plausibly overkill, but here is a drag and drop instruction you can give jules.google to generate a brand new AGENTS.md file for your Github repository. (Iteratively generated with Claude, of course...)
WOMM, YMMV

gist.github.com/oaustegard/a...

Example result: github.com/oaustegard/o...

AGENTS.md Generator for jules.google

AGENTS.md Generator for jules.google. GitHub Gist: instantly share code, notes, and snippets.

gist.github.com

3

Oskar @austegard.com · 2d

I must be a particularly slow human cause I frequently have Claude generate code that would take me more than a day to complete…

1

Oskar @austegard.com · 2d

With Sonnet 4.5 (and plausibly GPT-5 Thinking) my chart below is too pessimistic. My latest vibe-coding tool is 1436 lines of html/js/css and went through 56 internal Claude revisions with only two minor error cases -- all the other iterations was me dreaming up scope creep.

Oskar @austegard.com · May 25

I have an update to my Vibe Coding Effectiveness chart

When distributing effort to vibe coding and other forms of coding, or life in general you REALLY REALLY should listen to Vilfredo.

1

Oskar @austegard.com · 2d

Baby Driver's opening is 👨‍🍳💋: www.youtube.com/watch?v=rYv2...
but better yet:
www.youtube.com/watch?v=6XMu...

Baby Driver - Opening Titles, Coffee Run Scene (feat. Harlem Shuffle by Bob & Earl) [ORIGINAL]

YouTube video by Sujay

www.youtube.com

5

Oskar @austegard.com · 2d

You can try it here (now with categorized events) -- entirely client side, no data is transferred. You can download the html file from my site/repo and run it from your desktop: austegard.com/ai-tools/cla...

(No demo mode unfortunately)

Claude Audit Log Viewer

austegard.com

Oskar @austegard.com · 2d

Meanwhile I’m over here like…

1 1

Oskar @austegard.com · 2d

It’s cute how the Anthropic UX folks think 5 lines is an appropriate size for Project Instructions…

Default Claude Project Instructions prompt

1

Oskar @austegard.com · 2d

FWIW: Claude is excited about the potential of a hybrid solution of Pleias and TRM: arxiv.org/abs/2510.04871
🤞

Less is More: Recursive Reasoning with Tiny Networks

Hierarchical Reasoning Model (HRM) is a novel approach using two small neural networks recursing at different frequencies. This biologically inspired method beats Large Language models (LLMs) on hard ...

arxiv.org

1 2

Oskar @austegard.com · 3d

For the Claude Enterprise "owner" users out there -- I crafted a log viewer that lets you simply parse the zip formatted audit log file you can download via the management console:

github.com/oaustegard/o...

github.com

1 1

Oskar @austegard.com · 3d

Interesting! Satisfied with the results I tasked jules to create the same sets of READMEs for my other tools. I noticed in its actions that it used the same bsky/Processor_README.md as the basis, though I didn't ask for that -- it used its memory!

PS! My tools collection: austegard.com/tools.html

jules.google's memory of notable(?) actions it has taken in my website repo -- see the bottom one

Oskar @austegard.com · 3d

In a simultaneous Jules task I had it look at the README for the BSky Processor and replicate the pattern for each of the other tools:

PR: github.com/oaustegard/o...
Example README: github.com/oaustegard/o...

5 new READMEs for the BSky tools, courtesy of jules.google

1

Oskar @austegard.com · 3d

PR: github.com/oaustegard/o...
Before and After -- simple change but gnarly regex that would have taken me a while... Jules handled it easy.

My https://austegard.com/bsky/ page index before the PR

My https://austegard.com/bsky/ after the simple change to the web component

1

Oskar @austegard.com · 3d

jules.google doesn't get enough attention. It's really quite good, and nicely implemented, the way it runs in a vm with a clone of your project, thus able to locally test its contributions before issuing the PR. Not sure when they added memory functionality? I tend to use it for clean up/docs

Jules amending a webcomponent script used as a dynamic content index on my austegard.com website

1 2

Reposted by Oskar

Grace @gracekind.net · 3d

Give me a single reason why Sora2 should exist.

YouTube video by Hank Green

youtu.be

3 3 29

Oskar @austegard.com · 3d

How feasible is a small model with strong reasoning and contextual understanding, but minimal embedded knowledge? It seems ideal for many RAG applications and continued "learning" (i.e. new facts, but not improved reasoning skills).

1