weekend ai reads for 2025-05-16

“Best are voracious fan bases. Fan boys, fan girls,” Cunningham tells the group. “And an older demographic, where Aunt Carol doesn't really know how to use Facebook, and she's just likely to share everything.”

A mistake by an AI tool would not on its own be enough to trigger a payout under Armilla’s policy. Instead, the cover would kick in if the insurer judged that the AI had performed below initial expectations.

For example, Armilla’s insurance could pay out if a chatbot gave clients or employees correct information only 85 per cent of the time, after initially doing so in 95 per cent of cases, the company said.

We pay a great deal of attention to the words we use because they affect the way that we think. And the words that we use to frame a problem are some of the most important.

And yet, boredom: the great engine of creativity. I now believe with all my heart that it’s only in the crushing silences of boredom—without all that black-mirror dopamine — that you can access your deepest creative wells.

I hate the people who are building generative AI and trying to sell it to us. They're the same morons who tried to convince us that NFTs and the metaverse were the future, and they were wrong on both counts. Why am I meant to trust them now?

Will AI take your job? Not if it has something better to do. Call it Poleg’s Paradox: If AI is superhuman, it's a waste of energy to use it for tasks humans can do themselves. Ironically, the more powerful AI becomes, the more work it leaves for the rest of us.

So Ha set out to create a digital analog of a small firm. He built AI agents using Google’s Gemini 2.5 Pro model, grounding them in dozens of hours of recorded conversations between him and Borovich about their worldviews and investing goals.

…

After weeks of tweaking, Ha and Borovich had an army of agents at Antigravity that they collectively christened their “diligence engine” and that now reviews every startup they consider investing in.

That’s about when I realized the second big thing:

AIs are crappy architects.

It kept giving me stupid advice. For instance, every time it encountered an error, it would just catch it and print some logs. Uhhh… that’s bad. It would encounter a small problem, and design a big stupid solution instead of doing a small rearchitecture.

I'm here to help you brainstorm ways to use ChatGPT for Work! I also create custom-tailored prompts for your role. Get started by clicking the button below, then tell about your job and what company you work for. The more context, the better.

And remember, as AI democratizes the basic building blocks of software, the true differentiator isn’t shipping fast — it's shipping with conviction.

This is an area where many companies hit a stumbling block: they don’t know enough about processes at a granular level to begin to assess them, let alone quantify the potential benefits of inserting AI at critical junctures in those processes.

But there’s a way around this roadblock. Rather than manually sifting through countless business workflows, process discovery capabilities offer a more efficient way for organizations to pinpoint their most attractive AI opportunities.

With few guidelines besides blanket prohibition — which I don’t see as realistic — students are encouraged to develop their own ethical red lines surrounding AI. Consequently, a culture of silence and moral ambiguity has developed.

A homework help app developed by the Silicon Valley-based CourseHero provided instructions on how to synthesize flunitrazepam, a date rape drug, when Forbes asked it to.

Classic mistakes in this vein include thinking data is a moat when it isn’t; relying too much on weak data moats; confusing other moats (like scale) for data moats; misunderstanding which attributes of data contribute to its “moatiness”; failing to distinguish between software moats and data moats; and not realizing when a data moat has lost its effectiveness.

The findings contradict conventional wisdom that RAG inherently makes AI systems safer. The Bloomberg research team discovered that when using RAG, models that typically refuse harmful queries in standard settings often produce unsafe responses.

To build LegoGPT, the Carnegie Mellon team repurposed the technology behind large language models (LLMs), similar to the kind that run ChatGPT, for "next-brick prediction" instead of next-word prediction. To do so, the team fine-tuned LLaMA-3.2-1B-Instruct, an instruction-following language model from Meta.

The team then augmented the brick-predicting model with a separate software tool that can verify physical stability using mathematical models that simulate gravity and structural forces.

Our systems—of education, credentialing, hiring, housing, and electing officials—are dominated and managed by members of a “comfort class.” These are people who were born into lives of financial stability. They graduate from college with little to no debt, which enables them to advance in influential but relatively low-wage fields—academia, media, government, or policy work. Many of them rarely interact or engage in a meaningful way with people living in different socioeconomic strata than their own. And their disconnect from the lives of the majority has expanded to such a chasm that their perspective—and authority—may no longer be relevant.