weekend ai reads for 2025-05-23

One of the most unexpected findings of Wondercraft’s report was that only 41.8 percent of respondents under the age of 25 said that they used AI throughout their workflow — significantly lower than respondents in older age groups, according to Serrander.

Even when users interacted with the AIO, they didn’t go far. While 88 percent tapped “Show more,” the median scroll depth was just 30 percent. Most (86 percent) only skimmed the content, and very few reached the bottom of the answer. Trust and visibility were concentrated in the top third—similar to how people treat traditional snippets in search.

People who don’t pause exist more in their head than their body. The mind is top-down, rigid, quick, enforcing an established view. The mind is waiting for the other person to be done so they can say what’s rattling around inside. The body is slower, needs more time, and then words bubble up organically, one after another, without planning. People who exist more in their body are generally better at connecting emotionally with others.

Gross oversimplification, but older people use ChatGPT as a Google replacement. Maybe people in their 20s and 30s use it as like a life advisor or something. And then people in college use it as an operating system.

An MCP registry on Windows will act as the secure, trustworthy source for all MCP servers that AI agents will be able to access. “Agents can discover the installed MCP servers on client devices via the MCP registry for Windows, leverage their expertise, and offer meaningful value to end users,” says Davuluri. MCP servers will be able to access things like the Windows File System, windowing, or the Windows Subsystem for Linux.

All information retrieval systems, going back to analog libraries and forward to search engines, have biases. What’s different here is that lies are being baked into foundational technologies, with no other perspectives available.

Most agents fail due to feature incompleteness and weak integration. They may complete a first task but fail on follow-ups due to limited API access. Companies like Ramp solve this by giving agents full UI access, enabling them to operate the product like a human.

Our experiments confirm that all the top open- and closed-weight LLMs we test exhibit significantly lower performance in multi-turn conversations than single-turn, with an average drop of 39% across six generation tasks. Analysis of 200,000+ simulated conversations decomposes the performance degradation into two components: a minor loss in aptitude and a significant increase in unreliability. We find that LLMs often make assumptions in early turns and prematurely attempt to generate final solutions, on which they overly rely. In simpler terms, we discover that when LLMs take a wrong turn in a conversation, they get lost and do not recover.

Rather than testing one idea at a time, like a traditional company might, Daydream typically tests 15 to 20 ideas in parallel. “These AI tools are phenomenal at helping you find out if you're wrong faster, and helping you get to right faster,” he said.

Along with skills, data is also an issue -- and AI leaders have their work cut out. Over the past year, their organizations ran an average of 45 experiments with generative AI. However, only an average of 20 experiments (44%) are expected to go into production.

Talent shortages and the need for clean data are potential roadblocks.

1. AI training

2. Employee-driven innovation

3. AI and human collaboration

Now that I’ve changed my approach, I’m not frustrated with LLMs anymore. I have once again very low expectations, so when they do something well it’s a nice surprise. I’m trying to be smart about how I use them, they're such a great tool for learning for example.

This changes the purpose of docs: they’re no longer just for human readers, but also for agent consumers. In this new dynamic, the documentation interface becomes something like instructions for AI agents. It doesn’t just expose raw content, but explains how to use a system correctly.

AI can Talk! I spent 2 hours playing with Veo 3 @googledeepmind and it blew my mind now that it can do sound! This is all Generative AI text to video out of the box... it comes with dialogue, sound design and music

It’s been the backbone of the web since Al Gore flipped the switch, and it’ll still be here long after your trendy framework is rotting in a GitHub graveyard. So take your smartass logic and shove it. HTML’s king, and you're just a peasant with a keyboard.