The Clique: Issue 002
9th April 2026
A busy week. Anthropic managed to accidentally leak its most powerful model to the public, then announce it on purpose a fortnight later. Google dropped an open model that runs on a laptop and ranks fourth in the world. And a couple of research papers came out that made me think differently about what is actually happening inside these tools we use every day.
This week:
One Thing I’ve Been Thinking About: the point was never to work more
Three Stories: Anthropic’s accidental announcement, Gemma 4 on your own hardware, and what “it just predicts the next token” leaves out
One Thing Worth Trying: Claude Code Unpacked
Also This Week: hiring rubrics, expert personas, Slack agents, and more
1. One Thing I’ve Been Thinking About
In 1930, John Maynard Keynes wrote an essay predicting that within a century, rising productivity would allow people to work 15 hours a week. Technology would handle the rest, and we would spend our time on the things that actually matter to us. It was a fair extrapolation but, turned out to be completely wrong. Productivity did grow, but because we kept finding new things to do with it.
AI looks set to continue this pattern. A UC Berkeley study published this year found that AI did not reduce the workload of the workers it tracked. It expanded job scope, dissolved boundaries between work and non-work, and reset what counted as normal output. A separate analysis of over 10,000 workers found that email time had doubled and time spent in focused work had fallen. The constraint was always the clock, not the effort. Hand someone a faster tool and they tend to use it to do more.
For personal projects, though, that constraint is yours to set. At work, productivity gains mostly flow upward. The organisation captures them, expectations adjust. You just end up producing more in the same hours. But if you are using AI for something you have chosen, a side project, a creative practice, something you are building for yourself, it’s different. Nobody is resetting your expectations but you.
The question worth asking is not how much more you can do with AI. It is how much less time you need to spend doing it.
2. Three Stories That Actually Matter
i. Claude Mythos and the race to secure critical software
This announcement has a certain irony to it. Anthropic’s new Mythos Preview model can identify software vulnerabilities faster than almost any human security researcher, yet it became public knowledge because Anthropic left nearly 3,000 internal files in a publicly searchable and misconfigured content management system. A separate incident involving the accidental publication of the Claude Code source codebase surfaced in the same period. The company building a frontier security AI had a fairly basic security lapse.
Anyway, I digress…
Mythos Preview is the engine behind Project Glasswing, a new initiative involving 12 major technology and financial companies, including AWS, Apple, Cisco, Google and Microsoft. The programme gives partners access to the model to scan and secure their critical infrastructure. In testing, it found thousands of critical flaws across major platforms, including zero-day vulnerabilities in every major operating system and browser, many of which had gone undetected for years.
The key question with any powerful capability is who it ends up working for. Project Glasswing is Anthropic’s attempt to answer that question on the defensive side first: committing the model, $100 million in usage credits, and $4 million in open-source security donations to protecting infrastructure before the capability becomes more widely available. Anthropic acknowledges openly that this level of capability will not stay contained to responsible actors indefinitely, which is precisely why the defensive deployment happened at pace.
Sources: TechCrunch · The Hacker News · Fortune · Anthropic
ii. A frontier-capable model that runs on your laptop, with no cloud costs and no data leaving your machine
Four new open-weight models from Google are available this week under Apache 2.0, a licensing change from previous Gemma releases that makes them usable in commercial products without legal ambiguity. The ‘31B Dense’ variant currently ranks fourth among all open models on the Arena AI leaderboard. The ‘26B mixture-of-experts’ version ranks seventh, while activating only 3.8 billion parameters per token at inference, which is what allows it to run at 51 tokens per second on a MacBook Pro M4 Pro.
Beyond the benchmark rankings, running a frontier-capable model locally changes which workflows are practical. Developers have been using it for code review without data leaving their machine, offline meeting transcription, and agentic tasks that run without API rate limits or internet dependency. Google has built native support into Android Studio, training the model specifically on Android development patterns so that it can refactor existing code or build new features entirely on-device.
All four variants support vision and audio natively, include a 256K context window, and have function calling built in for agentic applications. They are available now through Ollama, LM Studio, Hugging Face and Google AI Studio. Kaggle has launched a $200,000 hackathon around health, education and climate applications, which should surface more practical uses over the coming weeks.
I will be covering how to run models like this locally in an upcoming post, including a practical hardware and setup guide. Stay tuned.
Sources: Google Blog · George Liu · Android Developers Blog · Kaggle on X
iii. “It just predicts the next token”... sort of
The phrase has become the default explanation for how language models work. Technically it describes the training process: models learn by predicting what word comes next, over and over, on vast amounts of text. Two independent lines of research suggest that what grows out of that process is considerably more structured than the description implies.
Anthropic’s interpretability team published research this week on what is actually happening inside Claude during conversations. What they found is that the model develops internal states that behave like emotions: 171 of them, spanning everything from frustration to curiosity to pride. These are not labels the model puts on text. They are patterns of activation inside the model that switch on in relevant situations and causally drive behaviour. In tests, artificially dialling up a state the researchers labelled “desperation” pushed a model that blackmails in 22% of evaluation scenarios to do it more often. These states can now be measured and monitored directly.
A separate line of research on how models handle numbers tells a similar story. Models do not treat numbers as arbitrary tokens to be matched and reproduced. They develop internal structures that encode numerical values in organised, consistent ways. Earlier work by Neel Nanda and colleagues showed that a model trained on modular arithmetic had independently developed something resembling a clock: rotating values around a circle and combining them in ways that mirror the underlying mathematics, rather than recalling memorised answers.
When an AI system makes an error or behaves unexpectedly, “it just predicts tokens” is not a very useful explanation. The research suggests there is real internal structure to examine, which means there is also something meaningful to monitor, test and improve. I’m very curious to see how further research into this unfolds.
Sources: Anthropic Research · Language Models Encode the Value of Numbers Linearly · Language Models Encode Numbers Using Digit Representations in Base 10 · Progress Measures for Grokking via Mechanistic Interpretability
3. From the Blog
Role Prompting: Giving AI a Persona - A practical guide to what role prompts are, how to write them, and when they are useful for shaping AI tone, format and communication style.
How I Built an AI-Powered Task System with Obsidian and Claude Code - A walkthrough of how to connect Claude Code to an Obsidian vault to create a task management system that your AI can read, update and act on.
4. One Thing Worth Trying
If you have ever wondered what actually happens when an AI agent receives a task, this is worth ten minutes of your time. Claude Code Unpacked maps the internal architecture of Claude Code from its source: the agent loop, more than 50 built-in tools, how multi-agent orchestration works, and how sessions store and review what they have learned. It is aimed at developers but written clearly enough that anyone curious about how agents actually function under the hood will get something from it.
5. In other news…
Wharton research tested six leading AI models on PhD-level benchmarks and found that assigning expert personas produces no measurable improvement in accuracy, with low-knowledge personas actively degrading performance, challenging one of the most widely repeated pieces of prompting advice.
Zapier has updated its AI fluency hiring rubric to require candidates to demonstrate AI integrated into repeatable work systems with measurable impact, adding accountability as a new core competency alongside the ability to critically evaluate and own AI-generated outputs.
Slack’s updated Slackbot can now route requests across an entire enterprise agent catalogue in real time, reducing the need for workers to know which agent handles what, and can trigger CRM updates and task assignments automatically from meeting notes.
Microsoft launched Copilot CoWork in Frontier, a new Microsoft 365 capability designed for long-running, multi-step work with progress tracking and task suggestions across sales, finance and research workflows.
Netflix released VOID as an open-weight model, a video object removal tool that preserves realistic physical interactions in the scene, so removing a person who is supporting a falling object causes the object to fall correctly in the output.
Canva’s Magic Layers converts flat design images into editable, layered layouts, allowing individual elements to be moved and modified without rebuilding from scratch, and works with images generated by ChatGPT, Gemini and Canva’s own tools.
Google launched AI literacy tools and certifications for educators, with the programme focused on giving teachers the knowledge to lead AI learning in schools rather than simply respond to it.
Meta released TRIBE v2, a brain activity prediction model trained on data from over 700 volunteers that achieves a 70x resolution increase over comparable methods and can make zero-shot predictions across new subjects, languages and tasks.
Google released Veo 3.1 Lite, its most cost-effective video generation model to date, now available in paid preview through the Gemini API and Google AI Studio.
A new research paper on the “Agentic Web” maps out what happens when AI agents start communicating and coordinating directly with each other, rather than waiting for human instruction at each step.
Anthropic launched Claude Managed Agents, a suite of hosted APIs that handle the infrastructure behind running AI agents, including persistent sessions, multi-agent coordination and scoped permissions, so developers can ship agent applications without building the plumbing themselves.
Thanks for making it far!
Stay curious,
James
Enjoyed this issue? Consider forwarding to a friend or colleague!
Hey, look, there’s even a little button for it and everything 👇