Anthropic's Self-Building AI insights, Squarespace Layoffs

Noted

Your daily brief in one short note from us to you.

· November 3, 2026

Your Daily Note,

AI is moving beyond answering questions. It’s helping build the next generation of AI models, taking on more of our daily work, and even attempting to predict how people behave. But as these systems become more capable, questions about cost, control and reliability are becoming harder to ignore.

Today, Anthropic is warning that AI could soon reach a new stage of self-improvement, a startup wants to predict human behaviour, and one founder is spending up to $200,000 a year on AI tokens.

Here are five stories worth understanding, plus a few more developments across AI and technology.

In today’s Noted:

–Anthropic says AI could be approaching the point where it helps build itself

–Y Combinator’s Garry Tan spends up to $200,000 a year on AI tokens

–Google brings Dreambeans to India, turning personal data into AI-generated stories

–Squarespace lays off about 19% of its workforce, around 300 people

–42AI comes out of stealth with $62.5 million to predict human behaviour

ANTHROPICself-building AI →

Anthropic says AI could be approaching the point where it helps build itself

We noted:

Anthropic has warned that it may be getting close to a stage where AI significantly accelerates the development of more capable AI systems. The company says none of its current models have crossed that threshold yet.

•The concept. In its latest regulatory filing, Anthropic discussed recursive self-improvement (RSI), where AI becomes capable enough to help develop and improve its own successors.

•Already in use. Anthropic disclosed that it uses AI systems in its research and development, and that AI-assisted work has previously improved model capabilities in ways it did not recognise at the time.

•The threshold. Its internal estimates suggest it may imminently meet its Responsible Scaling Policy’s definition of “dramatically accelerated progress attributable to AI.”

•Not there yet. None of its current models have been designated as exceeding that threshold, and the company is still assessing the evidence.

•Human oversight. Anthropic does not expect to safely let Claude design and build successive frontier models without meaningful human supervision or understanding.

Why it matters

AI will likely take on more work inside AI research itself, including coding, experimentation, testing and model development.

This could help AI companies move faster, but it also creates a difficult challenge: ensuring that safety and oversight keep pace with improvements in capability.

Y COMBINATORtoken bill →

Y Combinator’s Garry Tan spends up to $200,000 a year on AI tokens

We noted:

Y Combinator CEO Garry Tan says he spends between $100,000 and $200,000 annually on AI tokens to power the agents he uses for work and personal tasks.

•Where he said it. Tan shared the figure in an upcoming episode of Walter Isaacson’s podcast The Builders.

•What he uses agents for. Coding, research, preparing for interviews and managing everyday work.

•Shared tools. He has published 23 Claude Code skills, called gstack, on GitHub to help others see how agents can be used.

•YC’s own agent. An internal agent called Stray Light has context on founders the accelerator has backed, people who have contacted it and startup applications.

•The obstacle. Tan says spending this much is a privilege most people cannot afford, and token costs need to come down before personal agents go mainstream.

Why it matters

AI agents could increasingly move beyond individual prompts and become ongoing assistants that remember projects, access relevant information and carry out tasks across different applications.

However, the economics will matter. More capable agents may also require significantly more tokens, making usage costs an important consideration.

GOOGLEstories from data →

Google brings Dreambeans to India, turning personal data into AI-generated stories

We noted:

Google’s experimental app Dreambeans is expanding to India, using information from connected Google services to create personalised visual stories about users’ interests and everyday lives.

•Where it works. Dreambeans is available on Android and iOS in India, Google’s first market for the app outside the United States.

•How it works. It uses connected services such as Gmail, Google Photos, Calendar, Search, YouTube and Gemini to generate illustrated stories without a prompt. Google recently added animation using its Veo model.

•Languages. It supports 13 local languages, including Hindi, Punjabi, Tamil, Telugu and Bengali, plus English, but only one language per story.

•Privacy controls. Users choose which services to connect, can disconnect them and can delete their Dreambeans data. It works only with personal Gmail accounts.

•Early US numbers. Sensor Tower estimates shared with TechCrunch put US downloads at nearly 55,000 since June, down around 84% month over month in October.

Why it matters

Google is exploring more actionable features that go beyond stories and recommendations, so the app may do more with connected data over time.

If the India launch finds a larger audience, proactive apps that draw on the email, photos and calendars people already keep inside Google could become more common.

SQUARESPACEunder pressure →

Squarespace lays off about 19% of its workforce, around 300 people

We noted:

Website builder Squarespace is cutting about 19% of its global workforce, roughly 300 employees, as it restructures its organisation.

•The cuts. Squarespace is reducing its global workforce by approximately 19%, around 300 people, as it restructures to simplify how it operates.

•How it was reported. The Verge reported the layoffs first, and a company spokesperson confirmed them to TechCrunch.

•Not disclosed. The report does not say which roles, teams or locations are affected, or whether more cuts are planned.

•The market. Wix bought Base44 for $80 million last year, Lovable raised a $400 million Series C in August at a $13.3 billion valuation, and Google launched Stitch, a tool for generating UI designs.

•Its own AI. Squarespace launched Blueprint in 2024 and has announced Forma, an AI-native platform that currently has a waitlist.

Why it matters

Website building looks set to lean further on AI. Rivals are raising large rounds, and Squarespace is pointing to Forma as its next step.

Many small businesses run their sites on Squarespace, so how Forma develops, and how the company’s plans change after the restructuring, will matter to them.

42AIpredictive →

42AI comes out of stealth with $62.5 million to predict human behaviour

We noted:

Israeli startup 42AI has come out of stealth with $62.5 million in seed funding to build what it calls a behavioural world model, designed to predict how individuals will act.

•How it works. 42AI says it turns people’s psychological tendencies into mathematical traits, then a transformer-based model outputs probabilities for possible decisions.

•Who uses it. The company says it works with 10 large brokerages, using the system for liquidity, margin and hedging.

•Data sources. Onboarding, KYC, support calls, voice and sentiment analysis, transactions and trading activity.

•Company-reported results. It cites a 98% accuracy figure, which its CEO says measures whether trading flows fall within a predicted range, not whether each trade is right. These figures are unverified.

•Privacy. The company says the model never knows who a person is, though TechCrunch notes it still builds individual behavioural profiles from linked customer activity.

Why it matters

The company plans to expand into gaming, fraud prevention and cybersecurity, with a longer-term aim of helping robots and AI agents understand people. Its CTO acknowledged that each new domain needs its own training.

Its CTO also called the possibility that brokerages use predictions against their own customers a valid concern. The founders pointed to regulation, monitoring and an internal ethics committee, but TechCrunch notes they did not explain what would prevent misuse.

Products worth trying

Xmind

An AI-powered mind-mapping app whose new agent works alongside you on your maps.

Xmind 2027 adds Xmind Agent, which can work with your existing maps and local documents to research topics, compare ideas and reorganise information through conversation, and turn ideas into tasks and project plans. MCP support lets you bring your map context into tools such as ChatGPT and Claude. Free options are available, and Android support is live with iOS coming.

UnleashX

A platform for building AI employees that handle specific business roles and workflows.

Choose pre-built AI employees or create your own, connect them to existing business tools across more than 100 integrations, and set rules for how they operate, including human escalation for exceptions. Typical tasks include lead qualification, support follow-ups and CRM updates. A Product Hunt launch offer gives 50% off platform fees plus 1,000 credits.

Beacon

An open-source shared memory layer for AI coding agents.

Beacon, from Asymptote Labs, captures what happens in sessions across Claude Code, Cursor, Codex and OpenCode and turns useful discoveries into reusable knowledge and skills, so agents build on earlier sessions and you can see what they have been doing. It is free to get started.

Everything else in AI

Trajectory raises $45 million to let non-technical teams train AI models

Trajectory, a platform that helps businesses improve AI models for specialised work, has raised a $45 million Series A at a $300 million valuation, led by Sequoia with Nvidia, Bessemer Venture Partners and Bain Capital Ventures also taking part. Companies start with an open-source model, connect it to their apps and collect feedback on how it performs, and Trajectory uses that feedback to retrain the model over time. Harvey and Rogo are among its customers, and the platform is now publicly available.

Hugging Face proposes shared integration tests for its Transformers library

Hugging Face has introduced Model Integration Contracts, a way to package small, fast tests for the Transformers library, which supports more than 750 model types. Its example is a bug that sat in the main branch for almost two months before it was fixed. The new check currently reports failures without blocking changes, and Hugging Face lists open questions, including how many real bugs it will catch.

Want a note in your inbox?

Join Noted for 5–7 things worth understanding in AI, delivered daily.

Free, daily, unsubscribe anytime.

See you tomorrow
same time, different note.

~ Stackari

P.S. Twenty-three skills, one GitHub repo and a $200,000 token habit: Garry Tan’s gstack is public if you want to see what a heavy agent setup looks like.

Want to get in front of AI-curious founders?
Sponsor a note →

Was this note forwarded to you? Get your own →

Stackari · More news