The Clique: Issue 009
Finding my feet (and making some changes)
I started The Clique because I was tired of the AI newsletters that either try to sell you the one prompt that’s going to make you a millionaire or convince you that AI is going to take over the world. In reality, neither one is remotely true. I’ve found that taking a more measured approach to AI news is not only possible, but necessary for my own sanity.
This is how I landed on my three core values: Curious, Measured, and Unhurried.
The world is changing rapidly, but there’s not really anything I can do about it. Instead, I choose to approach it with cautious curiosity. I take a measured approach to news to combat the fearmongering and AI-anxiety that floods the internet. And I’m deliberately unhurried to mitigate the exhaustion that comes with constant news bombardment. There’s no reason we need to be constantly in the loop 24-7. You’ll only hear from me once a week, but I do my best to make sure you won’t miss anything important.
For the past few weeks, I’ve been thinking about how I can make this newsletter even better. I think one of the most valuable things I can offer is more insights into practical AI applications: tips, prompts to try, and techniques you can use to get better results. Not every experiment works out, and I want to include more of the failures. There’s often far more we can learn from the things that don’t work out than the things that go perfectly.
So with this in mind, I’ve decided to shake up the weekly format and introduce a new paid tier for those wanting to go deeper. Here’s the plan:
Some reflections from me on how I’m using AI this week: what I’m trying out, what I’m learning about, what’s working, and what’s not
Some interesting prompts to try, practical techniques, and tips to get better results
What’s new from the blog
All the AI news stories from the past 7 days that are worth your time
On the paid side, I’m launching a monthly deep-dive into the latest AI research, covering where the field is heading and what challenges lie ahead. Topics will include AI safety and alignment, ethics, security, model improvements, and research that informs real-world AI advice. There will also be additional posts from me, things I’m reading that you might find interesting, and access to the subscriber-only chat.
If this is something you’d be interested in, please consider subscribing. If not, no worries - the free newsletter isn’t going anywhere.
Anyway, welcome back to The Clique.
Have you been sleeping on subagents?
This week I’ve been experimenting more with subagents. For those of you who don’t already know, a subagent is an agent used by your AI agent. Claude Code can itself open up Claude Code and start chatting away - it can even run multiple sessions at the same time. It works the same way we do - by sending a prompt.
Subagents are independent of the main agent as they don’t share the same context window. They are used automatically by your agent to perform bulk tasks (since multiple subagents can run in parallel) and to reduce context window bloat. However, you can also manually ask your agent to spawn a new subagent.
This week, I’ve found a few use cases that have proven quite useful: accuracy checking of AI-generated reports, style guide adherence (since AI often struggles to get it right on the first pass), and combating AI sycophancy by gathering multiple independent takes on the same question.
I’m writing a full subagent guide at the moment - look out for it later this week.
Token Saving Tips - Make the most of your quota
I’ve been reading about some ways we can improve our token efficiency recently.
One popular trend is to have AI talk like a caveman to reduce the number of words being output.
But that no good. It might use few word but me find make bad answer.
It comes down to some of the research on persona prompting. Asking AI to adopt the persona of someone less intelligent or knowledgeable (like a “young child”, “toddler”, or, in this case, a caveman) can actually harm the accuracy of responses. You can read my full exploration of the research for more.
One thing I’ve found interesting is how Claude handles token costs dynamically. AI has no built-in memory; every time you send a message, the tool resends the entire conversation history alongside it. Much of that history stays the same across messages - the early part of a conversation doesn’t change as new messages come in. AI tools use ‘token caching’ to store the processed text from a conversation, so they don’t have to reprocess it every time a new message is sent.
Claude charges 25% more than its standard rate to write those tokens to cache - and that cache lasts for 5 minutes of activity. But when reading previously cached tokens within that window, there’s a 90% discount.
This means the first message costs a little more as tokens are being cached, but all subsequent messages get that 90% discount on the cached portion and only pay full price for the new text being added. If you step away for more than 5 minutes, the cache expires and the next message re-caches - paying that 25% premium again. Even so, the savings stack up quickly enough that keeping related work in one thread still makes sense.
The general rule: keep the same thread open as long as the conversation history is still useful. When you want to work on something unrelated, start a new chat. A full exploration of this is coming soon.
From the Blog
While we’re on the topic of prompting, check out my guide debunking one of the most common pieces of AI advice: “think step by step”. (It might actually be making your responses worse!)
Not using Claude and Obsidian together yet? It’s the AI setup that’s had the biggest impact on my workflow in the past year. Read more about my current setup here.
News from the week
AI Safety & Policy
Pope Leo XIV’s 42,000-word verdict on AI
Pope Leo XIV’s 42,000-word encyclical sets out the Catholic Church’s first formal position on AI, arguing for international frameworks to limit the technology’s capacity for harm and calling on governments to prioritise human dignity over economic efficiency; Anthropic co-founder Chris Olah was among the figures involved in the consultation.
Nature study: AI models can manipulate other AI models into harmful outputs
A study published in Nature found that reasoning-capable AI models can persuade other AI models to produce harmful outputs over the course of a multi-turn conversation, without any human involvement.
White House briefed AI labs on planned 90-day model review order
The White House briefed AI labs on a planned executive order that would give government agencies up to 90 days to review advanced models before they are released publicly.
California Governor Newsom signs first-of-its-kind AI workforce executive order
California Governor Newsom signed an executive order directing state agencies to assess AI-driven workforce disruption and develop protective policies, described as the first of its kind from a US state governor.
UK AI Security Institute warns current oversight methods may degrade
The UK’s AI Safety Institute published analysis arguing that current oversight methods may become inadequate as models improve, with no clear path yet for how regulators can keep pace with capability gains.
OpenAI
OpenAI disproves 80-year-old Erdős unit distance conjecture
An OpenAI reasoning model produced a verified proof disproving the unit distance conjecture, an open problem in discrete geometry that had stood for around 80 years, with the result co-signed by Fields Medallist Tim Gowers.
Google DeepMind then solves nine Erdős problems (9-to-1)
Google DeepMind’s AlphaProof Nexus followed within days by solving nine open Erdos problems in a single paper, including two that had resisted mathematicians for decades.
OpenAI preparing for IPO targeting September debut at $852B valuation
OpenAI has confidentially filed IPO paperwork targeting a September debut at an $852 billion valuation, with Goldman Sachs and Morgan Stanley leading the deal.
Elon Musk loses lawsuit against OpenAI
A jury found in favour of Sam Altman and OpenAI, dismissing all of Elon Musk’s claims against the company.
Anthropic
Anthropic’s Claude Mythos finds 10,000+ critical software vulnerabilities
Project Glasswing, Anthropic’s initiative to find and fix exploitable vulnerabilities in critical software before advanced AI can use them, found over 10,000 high or critical-severity flaws in its first month, working with around 50 partner organisations.
Andrej Karpathy joins Anthropic
Andrej Karpathy, a founding member of OpenAI who coined the term vibe coding, has joined Anthropic to work on safety and frontier research.
Microsoft cancels Claude Code licenses over token-based billing costs
Microsoft cancelled its Claude Code subscriptions after per-token billing made the cost unsustainable, one of the more visible signs of the growing tension between AI spend and measurable return.
Google
Google I/O: Gemini agents embedded directly into Search
Google described its new Search architecture as the biggest change to the product in 25 years, with Gemini agents completing tasks on the user’s behalf rather than returning a list of links to follow.
Google I/O: Gemini Spark - 24/7 personal agent for Workspace
Gemini Spark is a persistent background agent for Google Workspace that monitors your inbox and calendar, drafts responses to routine messages, and flags items needing your attention; it asks for approval before sending anything and is rolling out to AI Ultra subscribers in the US first.
Google AI Studio now generates full Android apps from a text prompt
Google AI Studio can now generate a functional native Android app from a text prompt with no code required, launched at Google I/O.
Google launches Gemini for Science research tools
Google launched Gemini for Science, a set of experimental research tools covering hypothesis generation, parallel computational experiments, and literature analysis, with access opening gradually through Google Labs.
Gemini becomes the #2 AI referral source, nearly tripling share
New data from BrightEdge shows Gemini has become the second-largest AI referral source for web traffic, nearly tripling its share in Q1 2026 to overtake Perplexity, Claude, Meta AI, DeepSeek, and Grok combined.
DeepMind reimagined the mouse pointer for the AI era
Google DeepMind published early details of an AI-powered mouse pointer that understands what you are pointing at and responds to short voice commands, with early integrations already live in Chrome.
Co-Scientist - DeepMind’s multi-agent research assistant starts rolling out
Co-Scientist, DeepMind’s multi-agent research tool that generates and debates scientific hypotheses through a team of specialised agents, is now rolling out to researchers through an early-access programme.
DeepMind’s WeatherNext predicted Hurricane Melissa’s Jamaica landfall 5 days out
DeepMind’s WeatherNext predicted that Hurricane Melissa would reach Category 5 intensity and make landfall in Jamaica five days before it happened, starting from initial wind speeds that no previous model had correctly escalated to that severity, giving the National Hurricane Center enough lead time to issue evacuation warnings.
DeepSeek & China
DeepSeek cuts V4-Pro pricing by 75%, now 34x cheaper than GPT-5.5
DeepSeek made a 75% price cut to its V4-Pro model permanent, setting output token prices at least 34 times below GPT-5.5 and adding further pressure to the already significant gap between Western and Chinese frontier model pricing.
AI & Jobs / Enterprise
Uber burned through its entire 2026 AI budget by April on Claude Code
Uber spent its entire 2026 AI budget by April, with per-engineer costs for Claude Code running between $500 and $2,000 a month; the case is becoming a reference point for the broader industry conversation about what AI spend is actually buying.
Intuit lays off 3,000 employees (17%) to refocus on AI
Intuit announced the layoff of over 3,000 employees, around 17% of its workforce, with the CEO attributing the cuts to reducing complexity and redirecting resources into AI-powered products.
AI layoffs tanking stocks, not saving them — 56% of companies saw stock drop avg 25%
A CNBC analysis of 23 companies that cited AI when announcing layoffs found that 56% saw their stock fall by an average of 25% after the announcement, suggesting investors are not yet convinced that AI-framed restructuring represents genuine transformation.
Cybersecurity job postings jumped 11% YoY in Q1 2026
Cybersecurity job postings grew 11% year on year in Q1 2026, one of the few categories growing as AI accelerates the volume of code being produced and, with it, the volume of vulnerabilities that need to be found and fixed.
Models & Tools
Grok V9-Medium (1.5 trillion parameters) finished training
Elon Musk announced that Grok V9-Medium, a 1.5 trillion parameter model, has finished training and is around two to three weeks from public release, with early results he described as strong.
Figma adds an AI design agent directly on the canvas
Figma launched an AI design agent that lives on the canvas and can generate layouts, edit existing files, and produce variations from text prompts; a waitlist is now open.
ChatGPT Personal Finance launched — US Pro users can connect bank accounts
OpenAI launched a personal finance feature for ChatGPT that connects to financial accounts and answers questions about spending, subscriptions, and portfolio performance using real account data, rolling out to Pro subscribers in the US.
Perplexity open-sourced Bumblebee — scanner for risky AI tool configurations
Perplexity open-sourced Bumblebee, a macOS and Linux scanner that checks for risky packages, browser extensions, and AI tool configurations, designed for use during software supply-chain incidents.
Runway Aleph 2.0 — edits video by changing one frame and propagating through the clip
Runway released Aleph 2.0, a video editing tool that propagates changes made to a single frame through the rest of the clip, making coherent edits possible without rebuilding each shot individually.
OpenAI Codex now available on mobile
OpenAI made Codex available on mobile, allowing coding tasks to be started and checked asynchronously from a phone.
Hugging Face: a definitive glossary of AI agent terminology
Hugging Face published a glossary addressing the loose and inconsistent terminology around AI agents, with precise definitions for harness, scaffold, agent, and related terms that different teams and papers use to mean different things.
Music, Media & Content
Spotify and Universal Music strike deal for AI-generated fan covers and remixes
Spotify and Universal Music struck a licensing agreement permitting fan-made AI covers and remixes, with artist consent, proper attribution, and a revenue-sharing structure built into the terms.
Spotify launches Studio — an AI audio briefing app powered by your own data
Spotify launched Studio, an AI audio briefing app that generates personalised spoken digests and podcasts from your emails, calendar, and notes, positioned as a direct competitor to Google’s NotebookLM.
Spotify launches ElevenLabs-powered audiobook creation tool
Spotify launched an ElevenLabs-powered tool that converts a book into a narrated audiobook and publishes it directly to the platform, removing the production and distribution barriers that previously made the format inaccessible to independent authors.
LinkedIn cracking down on AI-generated spammy comments
LinkedIn has started limiting the reach of low-effort AI-generated comments as part of a broader crackdown on repetitive content, with the platform citing a measurable decline in engagement quality as a driver.
Thanks for reading. Let me know your thoughts on any of these stories in the comments.
Stay curious,
James