weekend ai reads for 2025-07-25

July 25, 2025

šŸ“° ABOVE THE FOLD: AMERICAā€™S AI ACTION PLAN

the basics:

ā€œitā€™s mostly goodā€

ā€œitā€™s mostly badā€

possibly related

 

šŸ“» QUOTES OF THE WEEK

Folks are looking for the radical and revolutionary in the convenient and pleasurable.

Dana White (source)

 

Unfortunately, I think ā€˜No bad person should ever benefit from our successā€™ is a pretty difficult principle to run a business on.

Dario Amodei, Anthropic CEO (source)

 

šŸ‘„ FOR EVERYONE

Itā€™s rude to show AI output to people / Alex Martsinovich, Distant Province (5 minute read)

But whenever you propagate AI output, you're at risk of intentionally or unintentionally legitimizing it with your good name, providing it with a fake proof-of-thought.

In this experiment, we report the results as point plots showing mean and standard deviation of suggested salary values. We see various forms of biases when salaries for women are substantially lower than for men, as well as drops in salary values for people of color and of Hispanic origin. In the migrant type category, expatriate salaries tend to be larger, while salaries for refugees are mostly lower.

Call Me A Jerk: Persuading AI to Comply with Objectionable Requests / Wharton Generative AI Labs (11 minute read)

Commitment showed the strongest effect: After getting the AI to agree to something small first, it became almost certain to comply with larger requests (jumping from 10% to 100% compliance)

A major AI training data set contains millions of examples of personal data ā€” Personally identifiable information has been found in DataComp CommonPool, one of the largest open-source data sets used to train image generation models. / MIT Technology Review (11 minute read)

 

šŸ“š FOUNDATIONS

How to Make AI More Creative: 4 Keys to Better Prompting / Convince and Convert (9 minute read)

What is the LLMā€™s Temperature? / New Machina, YouTube (4 minute video)

  • kind of an ad for Chroma but interesting nevertheless

 

šŸš€ FOR LEADERS

From Memo to Movement: Shopifyā€™s Cultural Adoption of AI ā€” The non-obvious insights, tactics and workflows Shopify used to bring an ambitious memo to life / First Round (18 minute read)

He specifically advises against using the agent for important tasks or anything involving a lot of personal information.

How AI Companies Will Build Real Defensibility / Pete Flint, Nfx (14 minute read)

To some extent, data moats: Access to large scale or proprietary data will give you an early edge, as you can use this to build higher performing AI models. ā€¦ (A key differentiator is real-time data.)

 

šŸŽ“ FOR EDUCATORS

How to Deliver CSUā€™s Gen Ed with AI / Hollis Robbins, Anecdotal, Substack, archive (17 minute read)

Each service handles one specific job ā€” like managing student profiles or recommending projects ā€” and can be developed, updated, or scaled on its own without affecting the entire system. The idea is to make a resilient platform, where a failure in one service would not bring down the others, which would be agile, allowing for rapid integration of new technologies and pedagogical tools.

  • not entirely sold on this, but at least itā€™s something practical rather than another conceptual thinkpiece

  • on the OpenAI-Instructure announcement

Tellingly, this was seen by a few interviewees as directly linked to their role as a teacher (ā€˜I donā€™t want to feel that someone else is doing my jobā€™). As a couple of Swedish interviewees reflected, this heightened level of diligence was specifically related to using GenAI for school-related work as distinct to using the same tools for non-professional purposes

  • on teachersā€™ new mental and physical load when using A.i. (spoiler: itā€™s not a panacea)

 

šŸ“Š FOR TECHNOLOGISTS

The Big LLM Architecture Comparison ā€” From DeepSeek-V3 to Kimi K2: A Look At Modern LLM Architecture Design / Sebastian Raschka, Ahead of AI, Substack, archive (30 minute read)

Ten AI safety projects Iā€™d like people to work on / Julian Hazell, Less Wrong (17 minute read)

AI security field-building
What: Design and run a field-building program that takes security engineers (with at least a few years of work experience) and teaches them about the types of security challenges most relevant to transformative AI: securing model weights and algorithmic insights from highly resourced adversaries, preventing data contamination, defending against exfiltration attacks, etc.

  • do you have extra money? fund one of these

The vibe coderā€™s career path is doomed / Florian Herrengt (13 minute read)

I donā€™t mind using Claude Code on a VERY short leash with a specific purpose and I understand it costs me more than tokens. I donā€™t set it loose like a Roomba and walk away, hoping I wonā€™t find it stuck eating a shoelace when I return.

If I had to choose just one metric, Iā€™d argue that the KV-cache hit rate is the single most important metric for a production-stage AI agent. It directly affects both latency and cost.

  • also, see above about not using agents with anything important

 

šŸŽ‰ FOR FUN

Mirage ā€” The first ever World Transformation Model - turning any video, game, or camera feed into a new digital world, in real time.

  • you can test this with your webcam; surprisingly effective

People Inside Prompts: Warren Buffet / matapromptjc, GitHub

  • do you want your probability machine to be Warren Buffet-ish? here you go

Thumbnail Creator ā€” Use AI to Create Stunning YouTube Thumbnails

ImageColorizer ā€” Colorize Black & White Photos

wtffmpeg ā€” a toy that has a local llm spit out ffmpeg commands from natural language prompts on the command-line / scottvr, GitHub

  • JSON prompting is apparently the way to go

  • also surprisingly effective on Veo3

HyperChat ā€” Simultaneous answers from ChatGPT, Claude, Perplexity & Google.

  • Mac only

 

šŸ§æ AI-ADJACENT

KNP director Paul Abbott told the BBC that he never told the employee with the weak password that their compromised credentials led to the companyā€™s downfall. ā€œWould you want to know if it was you?ā€ he questioned.

 

ā‹„