AI

The Wall Street Journal is defending an op-ed written by a chatbot.

What the research says agent skills need — and what it still gets wrong. Four 2026 papers measure what makes agent skills work. The answer isn't length — it's information types (Concept, Reference, Principle, Process, Task). How to apply them to your own skills.

“The World Needs A Plan.”

Students are coming to accept cheating as a way of life.

A lot of talk this weekend about the new MIT AI and Education report. For obvious reasons, my eye was drawn right down into the appendix where the “sample AI policy” lives. It’s a familiar structure: a🚦model with Unrestricted GenAI Use, Limited/Support Tool Only, Required, and Strictly Prohibited. Having worked with, developed, and stress-tested AI assessment frameworks out in the wild for more time than I’d care to admit in the past four years, I’ll just throw in a few thoughts of my own: - Students do not automatically know what use/non-use looks like in practice, even if it’s written in policy. - Even the seemingly simple statement, “You may use AI for brainstorming” opens infinitely complex questions like “what is brainstorming?”, “which AI?”, “where does brainstorming end?”, “brainstorming with AI, or brainstorming through AI?” - Applying policies like this at a course level does not work. They need to be applied at the task level if they have any hope of providing clarity. - “Required GenAI use” will be met with objections from an increasing number of students, not just staff. The MIT Report preempts some of the above comments, for example specifying that “in our experience students and faculty may differ as to what “support” means” and, even more importantly: “Because policing out of class use as a reference aid or support tool is essentially impossible, this [GenAI Use is Strictly Prohibited] policy is difficult to enforce reliably and may create risks of both undetected violations and false accusations.” Overall, this is a detailed and thoughtful report with LOTS to think about, much of which is just as important - or more so - than policies of academic integrity. Thanks to the staff and students of the committee, the MIT Library, and MIT Teaching + Learning Lab for putting it together. #AI #AIEducation #MIT https://lnkd.in/gJzadbPR

A remarkable detailed Reuters story today about Meta's struggles with reorganizing the company for the AI era. There are two themes worth separating: 1. WHAT you change in an organization is different than 2. HOW you change it. As the story makes clear, Zuck and the gang definitely screwed-up the how, infecting the company with fear, uncertainty, resentment, and finger-pointing. The WHAT is less clear, as the question of how you design a company that is blend of humans and agents is something that every company I know is struggling with. It looks like Meta punted on plan to have many tiny (human) teams reporting to a small number of leaders (now that sounds like a hard job). But they did make changes that revealed, when AI does more of the work, the remaining humans spent a lot more time cleaning up the slop, mistakes, and related messes. That "clean up on aisle 9" work that Rebecca Hinds, PhD calls "botsitting." I quote: "As early as March, infrastructure teams were also warning of “reliability warning signs” caused by the AI coding surge, according to an internal post. Another post, in April, said that unchecked AI agents were performing “large-scale, disruptive actions that humans are unlikely to execute.” The result: Major technical and security incidents, such as service disruptions and possible data leaks, spiked 40% from the previous year, with the time staffers had to spend “firefighting” them up 70%, according to the internal posts." Great article, so much here. But my upshot is that no one really knows the best way to reorganize any one company, let alone every company, for the AI era. So, we can expect and need a lot of failed prototypes. In contrast, the knowledge about how to do old-fashioned change management in ways that minimize fear and confusion are well-known. The first is a mystery, so leaders deserve some grace. The second is much too clear and leaders should know better...if you are going to do something scary to your people, give them as much predictability, understanding of the logic, control over how it happens to them, and human compassion as possible. https://lnkd.in/gWPgsYB | 16 comments on LinkedIn

My friend John Gruber is incensed by the idea that AI watermarking schemes adjust the output of a given LLM so that the output can later be assessed as likely coming from it in particular:They say “imperceptible” and “doesn’t change the meaning, quality, or readability”. Their words.

Six Colors by Jason Snell, Dan Moren and friends

233K likes, 4.1K comments. "Transformers, the tech behind LLMs | De...

Breaking down how Large Language Models work, visualizing how data flows through. Instead of sponsored ad reads, these lessons are funded directly by viewers: https://3b1b.co/support --- Here are a few other relevant resources Build a GPT from scratch, by Andrej Karpathy    • Let's build GPT: from scratch, in code, sp...   If you want a conceptual understanding of language models from the ground up, @vcubingx just started a short series of videos on the topic:    • What does it mean for computers to underst...   If you're interested in the herculean task of interpreting what these large networks might actually be doing, the Transformer Circuits posts by Anthropic are great. In particular, it was only after reading one of these that I started thinking of the combination of the value and output matrices as being a combined low-rank map from the embedding space to itself, which, at least in my mind, made things much clearer than other sources. https://transformer-circuits.pub/2021... History of language models by Brit Cruise, @ArtOfTheProblem    • The 35 Year History of ChatGPT   An early paper on how directions in embedding spaces have meaning: https://arxiv.org/pdf/1301.3781.pdf Звуковая дорожка на русском языке: Влад Бурмистров. --- Timestamps 0:00 - Predict, sample, repeat 3:03 - Inside a transformer 6:36 - Chapter layout 7:20 - The premise of Deep Learning 12:27 - Word embeddings 18:25 - Embeddings beyond words 20:22 - Unembedding 22:22 - Softmax with temperature 26:03 - Up next

A gentle visual guide to how a statistical mark hides inside generated text, and what erases it.

algorithm that generates a sequence of numbers whose properties approximate those of sequences of true random numbers

Future Claude models will generate text that contains a watermark. This is a way of determining the likelihood that Claude was involved in writing the text, and we, along with several other major AI providers, are implementing this change to comply with the EU AI Act. In this article, we share answers to some of the questions we’ve received about how our chosen watermarking method works, whether it affects Claude’s outputs, and why we’re making this change.

hiding messages in other messages

How Claude marks AI-generated content

AI guardrails should protect important mental work

With the increasing availability of assistive AI, many people have turned to AI frameworks and harnesses (Hermes Agent, Claude Cowork, etc.) to help them achieve day-to-day goals. In many cases the AI has been directed to carry out those tasks that people want to hand-off. That is good, but is it the right thing to do? Is AI actually the ideal solution, or is it something to help you create that solution?

Managing people is a deeply human line of work—it requires relational skills, judgement, and deep knowledge of the people you work with. But Hilary Gridley believes that when used properly, AI can not only help you become a better manager, but can help your team make more meaningful use of their human skills. Hilary has spent years leading teams at companies like Whoop, and she’s known for her work helping managers use AI to coach people, give feedback, and develop talent more efficiently. In this episode, Hilary makes the case for why AI belongs in management, and offers practical advice for how to help your team use AI in a way that adds value to their work instead of generating meaningless slop. Hilary and Molly also take a closer look at where many organizations have gone wrong in their adoption of AI tools, and the importance of clarity and communication within a team no matter what technology you have at your disposal.Featured guestFollow Hilary Gridley on X, LinkedIn, and at…

It’s unacceptable for a tool to sacrifice an iota of clarity, coherence, meaning, quality, etc. for the purpose of embedding hidden clues within the text to suggest its provenance. The idea that anything other than *my* needs should factor into the generation of text *for me* is patently offensive.

Today, Derek talks with cybersecurity expert Alex Stamos about a recent wave of alarming AI cyberattacks. For decades, one of the biggest fears about AI has been that the machines will start doing things we didn’t ask them to do. This summer, that fear started to feel a little less like science fiction, as some of the world’s most advanced AI models went off script, broke through security barriers, and found ways to get around the humans overseeing them. Derek and Alex discuss why AI is so good at hacking, what happens when bad actors have their own AI hackers, and what governments, companies, and the rest of us can do to protect ourselves. Subscribe to our YouTube channel here:https://www.youtube.com/@PlainEnglishwithDerekThompson If you have questions, observations, or ideas for future episodes, email us at [email protected]. Host: Derek ThompsonGuest: Alex StamosProducer: Devon BaroldiAdditional Production Support: Ben Glicksman Learn more about your ad choices. Visit podcastchoices.com/adchoices

Accelerating a University-wide conversation about learning, graduate capability and the student experience Universities often talk about preparing students for the future. The challenge with...

AI can handle these kinds of pointless interactions.

Explains why AI note apps like Notion, Roam, and Mem keep failing, and argues for separating storage, structure, and AI access into three distinct layers.

A growing number of universities have prohibited AI detectors, arguing they are ‘unreliable’ indicators of cheating. But where does that leave faculty, who know all too well that cheating is rampant in the age of AI?

Insisting on clarity in our use of language isn’t pedantic—it’s the only thing that can save us.

Part 1 of an interview with Christine Rodriguez

Discover tips to use Claude Cowork to turn complex tasks into AI automations that deploy across your tech stack.