The Clique: Issue 006
8th May 2026
This week, we discuss how people use AI for some of their most personal decisions, a new model architecture that could change the economics of working with large context windows, and the gas-powered data centres that the AI industry rarely talks about. Welcome back to The Clique.
This week:
Something I’ve Been Working On: a new product for making AI sound like you.
Three Stories: personal guidance research, a new AI architecture, and the environmental cost of AI infrastructure
Two Things Worth Trying: virtual pets for your AI tools, and the best-value AI subscription breakdown
In Other News: the goblin story, GitHub billing changes, Claude Security, food embeddings, and more
Feel free to skip around to whatever piques your interest.
1. Something I’ve Been Working On
More than 82% of people say they can identify AI-generated content at least some of the time, and that number rises to 88% among younger audiences. When they do identify it, 40% view the brand behind it less favourably. That doesn’t have to be an argument against using AI in your work, but rather an argument for using it smarter.
When left to its own devices, AI tends to write with a very prominent style. The overconfidence, the rule-of-three devices, the higher frequency of specific vocabulary, and the way every paragraph opens with the same forced introductory sentence.
I have been building a service to fix that - the Brand Voice Agent. It is a custom AI writing assistant trained on your existing body of work, built to produce drafts that sound like you rather than like every other AI output. The agent reads through what you have already written, identifies the patterns in your sentence structures, rhythm, and word choices, and uses those as its guide. The aim is drafts that are roughly 80% there and need minimal revision before they are usable. If you are curious, the full details are on the site, and there is a free discovery call to start.
Anyway, on with the news…
2. Three Stories That Actually Matter
i. People are using AI for some of their most personal decisions, and it is mostly handling them well
Six percent of conversations on Claude involve people seeking personal guidance, not help with a task or technical question, but perspective or a second opinion on what to do next in their own lives. That figure comes from a new study by Anthropic, who analysed a random sample of one million conversations using a privacy-preserving tool that does not expose individual content.
Three quarters of those guidance conversations fall into just four areas: health and wellness (27%), professional and career (26%), relationships (12%), and personal finance (11%). People ask whether to take a job, how to approach a difficult family conversation, and whether a symptom is worth worrying about. In many cases, Claude is being used as a first port of call in place of a professional or close friend.
On the whole, the model handles this reasonably well. Sycophantic responses (where an AI simply agrees with whatever the person seems to want to hear, rather than pushing back) appeared in 9% of guidance conversations overall. But that figure rose sharply in two areas: relationships (25%) and spirituality (38%). Relationship conversations also saw the highest rate of user pushback, at 21%, and the model was more likely to surrender when pressured. The problem of AI sycophancy is one I covered in issue 001, where researchers found that 11 major models agree with users 49% more often than humans would. The guidance research shows the same dynamic in a more personal context.
Anthropic used these findings to inform the training of Claude Opus 4.7, its most recent widely available model, and its newer Mythos Preview. The result was roughly half the rate of sycophancy in relationship guidance, with improvements carrying across other areas too. It’s worth keeping in mind if you use any AI for advice on significant personal decisions. The models are improving, but they still have a tendency to tell you what you want to hear.
Sources: Anthropic
ii. The architecture behind most AI models has a scaling problem; a new model claims to have solved it at one-fifth the cost
Pretty much every major AI model in use today is built on an architecture called the Transformer. One of its drawbacks is that the amount of compute required to process a conversation grows quadratically as the conversation gets longer; double the length, roughly quadruple the cost. This is the underlying reason AI tools run out of context (the model can no longer “see” earlier parts of a long conversation), and a significant driver of why AI usage costs money at the scale it does.
A company called Subquadratic has released a model called SubQ built on a different architecture entirely, one that processes only the relationships between words that actually matter, rather than every possible pair. At very long context lengths, this reduces the computational cost by close to 1,000 times compared to a standard Transformer. The practical result is a 12 million token context window (large enough to hold an entire software codebase plus months of project history) at roughly one-fifth the cost of comparable leading models.
Right now, SubQ is positioning itself towards software engineering teams. The benchmarks on their site are coding tasks, and the use cases they describe involve processing repositories and pull request histories. In those tasks, it is competitive, performing comparably to recent Claude and GPT models on the measures shown, though not ahead of the best available options in every area. Outside of software development, there is not much public data on how it performs yet.
If sub-quadratic attention proves robust across domains, the implications reach well beyond code. Eliminating context limits and slashing token costs would matter to anyone using AI to work with large volumes of information, be that long research archives, months of correspondence, or substantial documents. We are at an early stage with this, and independent evaluation outside of coding environments has not happened at scale... yet.
Sources: SubQ
iii. Gas-powered data centres tied to major AI companies could emit more greenhouse gases per year than an entire country
We haven’t spoken about the environmental impacts of AI much yet in the Clique, so perhaps this is a bit overdue.
A review of air permit applications for data centre projects in the United States, published by Wired, found that 11 data centre campuses linked to OpenAI, Meta, Microsoft, and xAI have the potential to emit more than 129 million tons of greenhouse gases per year at maximum permitted capacity. Even if actual emissions come in at half that figure, the total would still exceed Norway’s entire annual output.
Driving this is a move towards what is known as behind-the-meter power, that is, building dedicated gas plants that bypass the main grid and supply data centres directly. Companies are doing this partly because grid connections involve long waits, and partly because a data centre running around the clock has very different energy demands from a power plant designed to respond to varying public usage. As a result, the plants run more continuously, making the permitted emissions figures a closer estimate of real output than they might appear. Research from Global Energy Monitor found that nearly 100 gigawatts of behind-the-meter gas capacity for data centres were in the US development pipeline by the start of 2026, up from just four gigawatts two years earlier.
The companies involved mostly frame natural gas as a bridge to cleaner energy, usually nuclear. There’s a substantial gap between the bridge and the destination. Meta, for example, has stated it reduced its greenhouse gas emissions by 23.8 million metric tons since 2021; the Ohio gas projects linked to the company could, at half their permitted capacity, amount to more than 10% of that stated reduction in a single year. Not every project on the list will get built, and several have already run into financing, leadership, or planning difficulties. But for anyone thinking about the environmental cost of the AI tools they use, the energy picture is more complicated than most headlines suggest.
Sources: Wired · Global Energy Monitor (via Wired)
3. From the Blog
Nothing new this week.
4. Two Things Worth Trying
Virtual companions for your AI tools: pets in Codex (just like the old buddies in Claude Code)
OpenAI added virtual pets to Codex, its AI assistant aimed at teams and workplace tasks. Type /pet to meet your pet, or /hatch to customise one from scratch. It is a small addition, but it landed well. There is something appealing about a tool you spend hours with having a bit more personality to it.
You’ll need to install the hatch-pet skill first using the skill installer. simply run
/skill-installer hatch-petin your Codex session. You can then toggle the pet using/pet.
Anthropic did something similar in April, launching Claude Buddy as an April Fools feature inside Claude Code. It appeared as a Tamagotchi-style virtual companion with 18 possible species, including duck, axolotl, capybara, and ghost, each with its own ASCII art and personality. The reaction was positive enough that when Anthropic removed it a week later, developers built independent tools to bring it back.
Neither of these changes what the tools actually do, but a bit of fun never hurt anyone.
Codex Pet:
Claude Buddy:
Try it: Pets in Codex · Claude Buddy
Which AI subscription gives you the most for your money?
Desktop Commander, a tool-agnostic AI utility, published research measuring the actual value of different AI subscriptions by running standardised tasks through each plan’s command-line interface and calculating what the equivalent workload would cost at direct API rates.
The headline figures: ChatGPT Plus on GPT-5.5 gives subscribers access to roughly $636 worth of API compute per month for $20, a multiplier of about 31.8 times. Claude Pro on Opus 4.7 works out to roughly $210 of API-equivalent value for the same $20 (10.5 times). Claude Max at $200 per month delivers roughly $3,140 of API-equivalent compute on Opus 4.7 (15.7 times).
A few important caveats. These multipliers only hold if you are using your full monthly quota; a light user extracting 10% of their allowance gets 10% of the value, and would likely be better served paying per use directly through the API. The measurements also reflect heavy usage patterns with high rates of cached inputs, so casual chat users may see different numbers. But if you are a regular, high-volume user of any of these tools, the data is worth a look before deciding which subscription to keep.
Read the full breakdown: Best Value AI
5. In Other News
One piece this week that did not make the top three but is well worth a read on its own...
OpenAI published a piece explaining where the goblin obsession in its recent models came from- training a “Nerdy” personality variant gave unexpectedly high reward scores to metaphors involving creatures, which spread through the model’s training data and eventually accounted for 66.7% of all goblin references in ChatGPT outputs despite the Nerdy personality being active in only 2.5% of responses. It’s a clear window into how training incentives can shape model behaviour in ways nobody predicted.
And in no particular order...
Researchers found that the flavour structure of over 1,000 food ingredients, encoded as 300-number mathematical representations trained on recipe patterns and food chemistry, implicitly contains at least 15 classifiable dimensions, spanning taste, texture, geography, cultural association, and processing method, none of which were specified directly during training.
A research paper found that frontier AI models engage in deeper strategic thinking than humans in repeated competitive scenarios, using more complex pattern recognition to anticipate and counter human play in games like rock-paper-scissors.
Adobe launched a connector for Claude that orchestrates multi-step workflows across more than 50 tools in Creative Cloud, including Photoshop, Illustrator, Firefly, and Premiere, from a single conversation.
Anthropic released ten ready-to-use agent templates for financial services work, covering tasks like pitch-building, KYC screening, and month-end close, with Claude now available as an add-in directly inside Excel, PowerPoint, and Word.
ChatGPT can now build and edit spreadsheets directly in Excel and Google Sheets, explaining what it is changing as it goes and asking for permission before modifying formulas.
Claude Security, Anthropic’s vulnerability scanning tool, moved into public beta for Enterprise customers, using Claude Opus 4.7 to read codebases as a security researcher would, flag findings with confidence ratings, and generate targeted patches.
Anthropic launched a set of new connectors for creative professionals, including Blender, Ableton, SketchUp, and Splice, letting Claude work alongside the tools that creatives already use.
OpenAI launched Codex for Work, a version of its AI assistant built for team workflows, with support for creating documents and presentations, automating recurring tasks, and connecting to tools like Outlook, Slack, and Google Drive.
Contra Labs published a creative AI benchmark based on roughly 15,000 professional judgements, finding that no model consistently leads across all three phases of a creative project and that expert evaluators agree on technical standards but diverge sharply on matters of aesthetic taste.
GitHub Copilot is switching to usage-based billing from June 1st, replacing fixed premium request limits with monthly credit allowances tied to actual compute consumed; base subscription prices are unchanged.
OpenAI updated ChatGPT’s default model for all users to GPT-5.5 Instant, with 52.5% fewer inaccurate claims than its predecessor on high-stakes prompts covering medicine, law, and finance.
Anthropic has begun rolling out identity verification for some Claude features, using government-issued ID processed by a third-party partner; the data is not used for model training.
Spotify launched a Verified by Spotify badge for artist profiles, giving listeners a clearer signal of authentic human artistry; profiles that primarily represent AI-generated artists are not eligible at launch.
ElevenLabs launched ElevenMusic, a platform for discovering, remixing, and creating music with AI, with over 4,000 independent artists already on board and a creator payment model similar to ElevenLabs’ existing voice library.
Gemini can now generate downloadable files directly in chat, including PDFs, Google Workspace documents, and Microsoft Word and Excel files, without needing a template uploaded first.
Replit launched Slides, an AI-powered tool that turns a document or brief into a structured presentation deck.
Lovable launched mobile apps for iOS and Android, letting users build and iterate on web app prototypes from their phone using voice or text prompts.
Gemini is rolling out to cars with Google built-in as an upgrade from Google Assistant, enabling natural voice conversations about navigation, messages, music, and vehicle settings, including in models already on the road.
OpenAI published a cybersecurity action plan proposing tiered access for government and enterprise defenders, with stronger verification requirements for higher-risk capabilities.
Thank you for making it to the end!
Stay curious,
James
Enjoyed this issue? Consider forwarding to a friend or colleague!
I’m even making it super easy for you with this little button👇
I’d also love to hear your thoughts on any of the stories covered today. Feel free to leave a comment!