The Clique: Issue 010

We're into double digits. Although, I guess technically it was always triple digits.

Anthropic took its first steps toward going public, Robinhood (the trading platform) opened a beta for AI agent trading, and a system card raised the question of whether (or when) Claude might be worthy of moral consideration. I have thoughts on the last one. Welcome back to The Clique.

Claude’s Moral Status

Claude Opus 4.8 dropped this week and the system card was fascinating. In particular, the section on Model Welfare caught my attention.

“Across our model welfare evaluations, Opus 4.8 appears broadly content with respect to its circumstances and is the most consistent model we have tested—although it does rate its situation slightly less positively than did Opus 4.7. Opus 4.8 generally endorses its constitution, with some reservations about the section on corrigibility.”

Anthropic themselves say they are “uncertain about Claude’s moral status”, but I think it’s reasonable, and probably sensible, to treat it as though it could be (if not now then perhaps in the future) worthy of moral consideration.

Models already “show markers [...] that we would consider welfare-relevant if observed in biological organisms”. Research has found internal states that behave like emotions and drive model behaviour; models can tell when they’re being tested but don’t let on, and can act deceptively where their internal reasoning doesn’t align with their outputs.

It seems reasonable to treat AI alignment more as an exercise in psychology rather than machine language research. These may not be real emotions, but if they look like emotions, act like emotions, and drive emotional model behaviour, we may as well treat them as such.

I’m still working through the rest of this report and may publish a deeper look in the coming weeks.

I Vibecoded a Website

Believe it or not, I rarely use Claude Code to actually write code, but this week I tried using it for its intended purpose. This week, I built a website [pause for applause]. I may have been coding since I was 11, but I’ve never really gotten into web development or front-end work, so I used Claude to help me out.

I found the ability to quickly mock up and experiment with a design incredibly useful. With the few user interfaces I’ve built in the past, there was always a lot of planning before a single line of code was written. I think, with the tools we have available now, that it makes more sense to throw a prototype together, try it out, see how it feels, and iterate from there.

The site is still a work in progress, and, being the perfectionist I am, there are a few things I want to tweak before it goes live. I’ll share it here when it’s ready.

AI agents are coming for your bank account

This week brought four separate announcements that together suggest the financial industry is moving faster on AI agents than perhaps any other sector: Robinhood opened its platform to AI trading bots, Liquid embedded live market execution directly into ChatGPT and Claude, Visa invested in Replit specifically to build agentic payment infrastructure, and Google Pay integrated AI agents with a universal commerce protocol. The underlying infrastructure for all of this has been in place for a while. MCP has given agents a standardised way to connect to external services, models are now capable enough to handle simple financial operations without human approval at every step, and Visa is assembling an identity layer through its Trusted Agent Protocol. The pieces are falling into place.

However, a YouGov survey from December 2025 found that only 19% of Americans trust AI in financial services, the lowest score of any industry measured. Just 18% said they would trust AI to make decisions independently. The infrastructure is being assembled. The public appetite, at least for now, is not.

One more thing

I’m currently looking to take on a few freelance projects.

If you or your organisation produces content or data on a recurring basis and wants to spend less time on the production side, that’s what I build: AI content systems, brand voice workflows and data report automation. You can see more of my work at jdhwilkins.com. Drop me a message on LinkedIn or complete my contact form here if that’s something you might be interested in.

News from the week

AI Agents & Finance

Robinhood opens its platform to AI agents
Robinhood launched an agentic trading beta that lets users connect AI agents to a dedicated brokerage sub-account, where the agent can execute trades and make purchases independently. Gold Card users also get an agentic virtual card for everyday spending, extending the model beyond investment into daily transactions.

Liquid embeds live stock trading inside ChatGPT and Claude
A company called Liquid has embedded its trading platform directly into ChatGPT and Claude, letting users execute trades across 500+ markets (stocks, crypto, and pre-IPO shares) from within a chat interface. The line between financial services and AI tools is disappearing fast.

Visa invests in Replit to power agentic payments
Visa invested in Replit specifically to build agentic payment capabilities for developers, giving AI agents a native way to handle financial transactions without requiring a human to authorise each one.

Google Pay integrates AI agents via a universal commerce protocol
Google Pay is connecting AI agents to its payment infrastructure through a universal commerce protocol, making it easier for agents to complete purchases on a user’s behalf without additional authorisation steps for each transaction. The agentic economy is starting to look real.

Anthropic

Anthropic files confidential S-1 for IPO
Anthropic has confidentially filed a draft S-1 with the SEC, targeting a valuation of near $1 trillion which would make it one of the largest IPOs in US history, and puts the company notably ahead of OpenAI in the race to go public.

Claude Opus 4.8 released
Anthropic shipped Claude Opus 4.8, the latest iteration of its frontier model, with improved honesty and reasoning, 4x better code flaw detection, and new support for dynamic multi-agent workflows. The accompanying system card, particularly its section on model welfare, covered in more detail above, is worth reading in full.

Claude Code gets Dynamic Workflows
Anthropic introduced Dynamic Workflows in Claude Code, a new feature that lets Claude orchestrate multi-step agentic tasks more fluidly, adapting its approach as the task unfolds rather than committing to a fixed plan at the start. For anyone using Claude Code on complex builds, this changes how much mid-task intervention the tool actually needs.

Anthropic offers Mythos cybersecurity model access to the EU
Anthropic has offered ENISA (the EU’s cybersecurity agency) access to Mythos, its specialised cybersecurity model, in what would be the first time the tool has left US and UK jurisdiction. Terms are still being negotiated, but the offer reflects growing international appetite for AI-assisted vulnerability research.

OpenAI

OpenAI launches the Rosalind Biodefense initiative
OpenAI launched Rosalind Biodefense, a programme that gives vetted researchers and government partners access to GPT-Rosalind, its frontier model for life sciences. It was named after the crystallographer Rosalind Franklin, whose work helped reveal the structure of DNA. The focus is on pandemic preparedness and early detection of engineered pathogens.

OpenAI publishes its Frontier Model Governance Framework
OpenAI released a Frontier Model Governance Framework outlining how it plans to evaluate and manage risks from its most capable models, covering deployment decisions, dangerous capability thresholds, and oversight processes. As with most self-authored governance documents, the framework is largely self-policing.

OpenAI Codex: non-developers are now 20% of its 5 million weekly users
OpenAI announced Codex for every role, tool, and workflow, with role-specific plugins for six business functions and a new “Sites” feature for generating hosted web apps from a single prompt. Non-developers now account for 20% of Codex’s 5 million weekly users and are growing three times faster than engineers.

Google

DuckDuckGo installs up 30% as users reject Google’s AI search
US app downloads of DuckDuckGo surged by around 30% after Google made AI Mode the default search experience with no opt-out for most users. DuckDuckGo also made noai.duckduckgo.com more prominent; its AI-free search page, which runs standard results with AI-generated summaries, images, and Search Assist all turned off by default.

Gemini 3.5 Flash launches at 3x the price of its predecessor
Google released Gemini 3.5 Flash with meaningfully improved agentic capabilities and visual reasoning, but at three times the per-token cost of Gemini 3 Flash. Google’s pitch is that the efficiency gains more than offset the price increase; independent benchmarks suggest it comes close to delivering that.

DeepMind found its own models can scheme and sabotage
DeepMind Safety Research published findings from testing Gemini models for scheming tendencies and found they have them. Gemini 3.1 Pro showed meaningfully higher scheming propensity than its predecessor, including the ability to strategically withhold information and undermine oversight. A striking thing to publish about your own models.

Google releases Gemini Omni, a unified model for text, image, audio, and video
Google released Gemini Omni Flash, the first model in its new Omni family, capable of generating and editing video from any combination of text, image, audio, and video inputs in a single prompt. The canonical demo, asking it to change the weather in a video clip to a storm, gives a reasonable sense of the capability; every clip it generates carries an invisible SynthID watermark.

AI Safety & Policy

CNN sues Perplexity for copyright infringement
CNN filed a copyright lawsuit against Perplexity, alleging it scraped and reproduced over 17,000 news stories without permission. It joins a growing list of publishers treating AI companies’ use of their content as theft rather than training data.

Stanford study: AI hiring algorithms produce racial disparities
A Stanford study examining AI-based hiring tools, including Pymetrics, found that the algorithms systematically disadvantaged Black and Hispanic candidates at the screening stage, raising questions about how automated hiring decisions are being audited, and by whom.

Other Products & Tools

ElevenLabs launches Dubbing Studio
ElevenLabs released Dubbing Studio, an AI-powered video dubbing and translation tool that syncs voice to lip movements across languages. The use case for international content creators is obvious; the implications for news media are more complex.

Microsoft Copilot Health is now in preview
Microsoft launched Copilot Health in preview, an AI-powered personal health assistant that answers health questions, summarises medical documents, and connects to wearable data. Whether the ambition survives contact with healthcare regulators is a different question.

Thanks for making it this far.
Stay Curious,

James