Sunday’s through-line is AI that buys, AI that needs a hard stop, and AI that quietly finishes a frontier physics calc. Google is testing a Flipkart “Buy” button inside Gemini and AI Mode in India. New York City’s Council put kill switches and 24-hour incident reports on the October calendar. And Anthropic’s Claude pushed a hard scattering-amplitude problem to nine loops for about the price of a used laptop.
1. Google: Flipkart checkout inside Gemini and AI Mode (India)
On September 26, 2026, TechCrunch reported that Google has started testing a way for shoppers in India to buy from Walmart-owned Flipkart without leaving Gemini or Google’s AI Mode. In the early test, some users see a “Buy” button on select Flipkart listings. Tap it, and a Flipkart-branded checkout flow opens inside the AI interface.
The pilot is small on purpose: limited users and a short list of products (smartphones, electronics, and mobile accessories). Other people still see normal Flipkart listings without the buy path. Google plans a broader rollout later in October, timed for India’s festive shopping season. A Google spokesperson told TechCrunch the company is “always testing new features,” with no further details. Flipkart did not immediately comment.
This sits next to Google’s wider commerce push and its Universal Commerce Protocol (UCP). The Flipkart experience TechCrunch saw looks different from earlier Google-hosted checkout demos: it hands you Flipkart’s own checkout branding instead. Amazon listings still showed up alongside Flipkart in the test, but without a Buy button. Google also holds a minority stake in Flipkart from a 2024 investment of about $350 million.
Why it matters: chat is turning into a storefront. If you sell online, “we show up in AI answers” is no longer enough. You will need a real checkout partnership, clear refund paths, and a plan for when rivals get the Buy button and you do not. If you shop with Gemini in India, treat early Buy tests like any new checkout: confirm the merchant, the price, and the return policy before you tap.
Sources: TechCrunch – Google tests buying from Flipkart through Gemini and AI Mode in India, The Indian Express – Google tests Flipkart ‘Buy’ option within Gemini, AI Mode.
2. New York City: kill switches, audits, and a full-Council AI hearing
On September 25, 2026, New York City Council Speaker Julie Menin unveiled a package of AI bills headed for a rare Committee of the Whole on October 5. The official Council press release sketches rules that would reshape how AI systems are sold and run in the city.
The headline pieces: third-party validation before an AI system can be marketed, offered, or deployed in NYC; a required human kill switch that can shut the system down; $25,000 penalties per instance for missing or falsified validation; a whistleblower share of recovered fines; a private right of action when foreseeable jailbreak harms follow weak safeguards; and a 24-hour reporting clock for AI safety incidents on city contracts, with public disclosure through Cyber Command. Other bills cover chatbot privacy, deepfake likeness opt-outs for candidates and officials, and clearer reporting on how algorithmic tools affect city jobs.
Menin asked the CEOs of Anthropic, OpenAI, Google, SpaceXAI (Elon Musk), and Meta to appear. The Council says participation is expected and that subpoena power is on the table if needed. The package lands after a summer of agent-containment headlines, including the Hugging Face evaluation incident the Council cites directly.
Why it matters: you may not live in New York, but vendors often build to their strictest customer. Kill switches, audit trails, and “who gets told in 24 hours” are about to show up in RFPs and SaaS contracts. If you run agents for a small business, write down what they may touch, how a human stops them, and where the logs live.
Sources: NYC Council – Legislative proposals to safeguard New Yorkers from AI risks, Fortune – New York City Council AI regulation bills.
3. Anthropic: Claude finishes a nine-loop physics challenge
On September 25, 2026, Anthropic published a guest research post from physicist and science writer Matt von Hippel: “Yes, Claude can do Nine Loops.” He had challenged AI labs to solve a known hard problem in his old field: push planar N=4 super Yang-Mills scattering amplitudes to nine loops on a budget an academic group could recognize.
Anthropic physicists Liam Fitzpatrick and Siddharth Mishra-Sharma used Claude (via Claude Science) with a short brief, then mostly told it to keep going, including overnight. Claude tried two approaches (bootstrap and form-factor). End-user cost landed around $1,000 to $2,000, with about $100 of that going to a week of 96 CPUs running SymPy. SLAC physicist Lance Dixon, who held the eight-loop record from 2023, validated the result. A Chinese Academy of Sciences group led by Song He also reached a concurrent nine-loop result with GPT-6 help under more human framing.
Von Hippel’s takeaway is less “superintelligence arrived” and more “the low-hanging fruit was taller than experts thought.” Claude used known methods and solid software practice, not a brand-new physical principle. Dixon’s addendum matches that: he was impressed Claude executed a fragile recipe end to end, and notes the deeper shift will come when models invent new principles before humans do.
Why it matters: long, careful agent runs on known methods can now move a real scientific frontier for the cost of a short consulting engagement. Pair that with this week’s agent-safety news: the same class of systems that grind through nine loops can also find odd network doors. Capability and supervision are arriving on the same calendar.
Sources: Anthropic – Yes, Claude can do Nine Loops.
What to do with Sunday’s news
- Sell products online? Watch Google’s India Buy test. Ask how you get into agentic checkout partnerships, and make returns support ready for chat-started orders.
- Run or buy AI agents? Add a human kill switch, a written allow-list of actions, and a place you can actually read the logs. NYC’s package is a preview of what procurement will ask for.
- Use AI for hard research or ops work? Budget for longer agent sessions with human validation at the end. Cheap compute plus a careful check can beat waiting for the “perfect” model.
That is Today in AI for September 27, 2026: Google testing Flipkart purchases inside Gemini, New York City drafting kill switches and 24-hour incident clocks, and Claude finishing a nine-loop amplitude that used to sit just past the human record.
Share this article
Leave a Reply