Newsletter, 6 November 2026
More people are asking ChatGPT, and Google is passing on fewer visitors
This week: ChatGPT's growth and what it means for search traffic, two different ways to give engineers AI coding tools, a state gaming commission asks operators about AI, and Meta's Muse agent is already taking orders.
Our four views on how AI is changing iGaming are on the blog: Acquisition is changing, A new front door, How products get built and businesses run and Regulation. This week's news sits underneath them.
More people are asking ChatGPT, and Google is passing on fewer visitors
Our view: Acquisition is changingOpenAI says ChatGPT has 1.2 billion weekly users, up from 1 billion two months ago, and that its volume of search-like queries is now roughly one-fifth of Google's. The figures come from ChatGPT consumer product lead Adam Fry, speaking to Alex Kantrowitz (Big Technology, 3 November).
Reddit shows the other side of it. Its shares fell around 30% after its Q3 results, even with revenue up 53% on the year (Kyle Byers, LinkedIn). CEO Steve Huffman told analysts that Google search referrals had gone down meaningfully in the quarter, which he linked to AI Mode and AI Overviews (Search Engine Roundtable, 2 November). Movable Ink's 2027 predictions (2 November) add that customers can compare and narrow down their options with an assistant before they ever reach a brand.
Two big firms, two ways to give people AI coding tools
Our view: How products get built and businesses runRevolut's David Abitbol, who leads AI operations, presented at AI Engineer NYC on 13 October that about 2,000 engineers now default to self-hosted GLM on 56 Nvidia B200 GPUs. He puts the cost at around $0.30 per million tokens, roughly 90% below before, with Claude Code and Codex kept for exceptions (Runtime Wire, 3 November).
Goldman Sachs went a different way. Bloomberg reports (2 November) that it has taken Claude Code and GitHub Copilot back from bankers and traders, who had used them for months, and kept them for roughly 12,000 engineers, following a review of productivity and risk. A spokesperson said the firm wanted the tools with the people best suited to use them over the long term. The reports don't say the tools failed to deliver value, and Nate Andorsky's LinkedIn commentary reads it as a coordination question: many people building their own tools at once is hard to keep consistent (AI Weekly).
Massachusetts asks sportsbooks how they use AI
Our view: RegulationThe Massachusetts Gaming Commission set 3 November for sportsbooks to answer its 9 October letter on AI and machine learning. The questions cover promotions, marketing, fraud, risk and account management, which is most of the AI an operator runs, not only chatbots. Replies are voluntary for now and kept confidential, and whether this becomes a rule will likely depend on what the disclosures show (Gambling.com, 15 October). No outcome has been reported yet, and we will follow it.
A short list of each AI system, the data it uses and the human oversight around it makes questions like these quick to answer. Our Personalised Player Journeys, Bonus Recommendation Engine and Retention Engine use cases are designed with consent, human approval and an audit trail from the start.
Meta's Muse agent is already taking orders
Our view: A new front doorOlivia Moore of Andreessen Horowitz (a16z), who follows consumer AI closely, reports from Meta's Q3 call (LinkedIn, 3 November) that 25% of early Muse users have completed a purchase through the agent, with an average over $200, and that shopping is the top activity.
She also shared a free way to see what an agent sees (LinkedIn, late October). Create a dummy account on an agent platform, ask it to find something you would want to be found for, and ask how it chose. In her test, an agent blocked by one retailer moved on to another, and picked the one with no bot challenges, verifiable stock and same-day pickup.
Other things we've been reading
- -Three levels of AI productivity. Jeroen Tas describes individual, process and enterprise levels, and says most people use 2026 tools inside 2015 job structures (LinkedIn, 3 November).
- -Support tickets. Cibe Hariharan shared ReDefiNER, a 0.8B model that runs on your own machine without a GPU. A plain-English rulebook sets what it pulls from a ticket, which could suit routing support queries (LinkedIn, 5 November).
- -Small models for narrow jobs. Rob Litterst on Coinbase training a small open model on its own data for fraud prevention, and CEO Brian Armstrong's view that specialists win narrow tasks (LinkedIn).
- -Personal software. John Ginsberg is 90% sure the future of software is personal, with SaaS providing the foundations and guardrails and agents building around each organisation's needs (LinkedIn).
- -Agents as teammates. Stuart Winter-Tear on new Stanford, MIT and Oxford research: agents that understand the other side can still concede too easily in negotiation, so working well with others is worth measuring on its own (LinkedIn).
- -Ontologies for reliable AI answers. Bharath S. explains how an ontology adds business meaning and rules to a knowledge graph, so answers can be checked before they are given (LinkedIn).
- -Rebuilding SaaS. Eric Fraser on OccupySaaS, a group of engineers using open-source AI tools to rebuild established SaaS products (LinkedIn).
- -AI in security. Einat Shimoni on a16z's survey of CISO priorities for 2027 and how her team at Lusha has put AI into security operations (LinkedIn).
Work with us
If one of these is useful to you, send us a message or visit bettingproduct.ai/products. We start with a fixed-scope AI Discovery and Pilot: one problem, one pilot, one decision.
Ian and Vlad | BettingProduct.ai