ChatGPT now connects to your bank account; Claude's business adoption surges
🤔 “Should I let ChatGPT connect to my bank account?”
Share Handy AI with your coworkers and friends to help them understand the crazy world of modern artificial intelligence and make the right decisions.
what to know for now
💰 ChatGPT now connects to your bank account. OpenAI shipped personal finance tools for Pro subscribers in the US, with Plaid handling connections to 12,000+ institutions including Schwab, Fidelity, Chase, Robinhood, Amex, and Capital One. Once connected, users get a dashboard of portfolio performance, spending, subscriptions, and upcoming payments, plus the ability to ask GPT-5.5 to reason across the whole picture. OpenAI says 200 million people already ask ChatGPT financial questions every month; the company just wired the answers directly to the data. Read more
📈 Anthropic passed OpenAI in US business adoption. Ramp’s May AI Index, drawn from 50,000+ US companies, puts Anthropic at 34.4% business adoption against OpenAI’s 32.3% as of April. That’s the first time Anthropic has led that number since the AI race began (it was at 8% a year ago). The same week, Dario disclosed 80x year-over-year Q1 growth, ARR climbing from $9B at year-end 2025 to $44B by May, customers spending $1M+ annually doubling from 500 to over 1,000 in two months, and gross margins above 70%. Sequoia, Dragoneer, Greenoaks, and Altimeter are co-leading the $30B round at a $900B valuation, expected to close by end of month. If it lands, Anthropic passes OpenAI’s $852B sticker. Read more
🤖 Google I/O 2026 kicks off tomorrow, and Gemini 4.0 is the headline. The keynote opens May 19 at 10am PT, with the two-day Shoreline conference rumored to ship Gemini 4.0 (sized to fight GPT-5.5, not Claude Mythos), the Gemini Omni video model, an Android XR glasses preview, and a new Aluminium OS that merges Android and ChromeOS for fall’s Googlebook laptops from Acer, Asus, Dell, HP, and Lenovo. Read more
🧰 Greg Brockman now runs everything OpenAI ships, and Codex just landed on your phone. OpenAI rolled ChatGPT, Codex, and the developer API into one product org under Brockman (officially, after stepping in during Fidji Simo’s medical leave). Thibault Sottiaux, the engineer who turned Codex into one of OpenAI’s fastest-growing products, runs the unified consumer/enterprise/developer surface. Two days earlier, Codex shipped inside the ChatGPT iOS and Android apps for every plan including free, letting developers approve diffs, redirect agents, and monitor terminal output from anywhere (it dials home to a Codex desktop instance, macOS first). Read more
🖱️ DeepMind reinvented the mouse pointer. Google DeepMind posted a research blog and two live AI Studio demos for an AI-aware cursor: a pointer that captures the visual and semantic context around itself, so you can say “make this brighter” or “find places like this” while just hovering. A deeper integration called Magic Pointer is rolling out inside Chrome, and the same model gets baked into the new Googlebook laptops announced for fall. Read more
🪣 Anthropic told Max subscribers their Agent SDK credits were getting their own $200 bucket, and the community-noted it within hours. Starting June 15, claude -p, the Agent SDK, GitHub Actions, and any third-party tool authenticating through a Claude subscription move off the subscription rate-limit pool onto a separate $200/month credit pool, metered at standard API list prices. Anthropic framed it as “free credits,” but the community framed it as a 12x to 175x effective price hike on programmatic workloads, depending on how heavy your agent runs are. Read more
🛡️ Google traced a real-world hack back to an AI model that wrote the exploit. Google Threat Intelligence Group published on May 11 that it has “high confidence” it caught a criminal group using an AI LLM to find and weaponize a previously unknown zero-day in a popular Python script, bypassing two-factor authentication on an open-source system with stated plans for a “mass exploitation event.” Google won’t name the group, says it wasn’t Gemini or Claude Mythos, and flags that China- and North Korea-linked actors are exploring the same playbook. Read more
🏦 Anthropic is briefing the Financial Stability Board on the cyber holes Mythos found. Bank of England governor Andrew Bailey, who chairs the FSB, asked Anthropic to walk the G20 watchdog through what Claude Mythos has surfaced in global finance infrastructure (thousands of high-severity vulns across operating systems, browsers, and core software). Mythos has been distributed to ~40 organizations including JPMorgan, AWS, Microsoft, and CrowdStrike under White House restriction; the FT reports Anthropic is now briefing select non-US bodies including the European Commission in parallel, while negotiating Mythos export controls with Washington directly. Read more
🎨 A guy posted a real Monet on X, labeled it “Made with AI,” and asked critics to explain what was wrong with it. @SHL0MS shared a Water Lilies canvas with the prompt “I just generated an image in the style of a Monet painting using AI. Please describe, in as much detail as possible, what makes this inferior to a real Monet painting.” The replies poured in. Critics knocked the “lack of depth,” the way “light doesn’t behave on water,” the “missing mess of humanity.” It was a Monet. Read more
🧪 AI Research of the Week
Cattle Trade: A Multi-Agent Benchmark for LLM Bluffing, Bidding, and Bargaining
Robert Müller & Clemens Müller
Jake’s Take: Seven cost-efficient LLMs and three deterministic code agents played 242 rounds of a 50-to-60 turn economic game that mashes auctions, trading, bargaining, and bluffing into a single environment with imperfect information and a money clock. The point was to see whether a model can hold a coherent strategy when the other agents are lying to it, the prices are moving, and the resources are running out. The deterministic code agents beat almost every LLM, and the LLMs lost in extremely human-recognizable ways: overbidding, bidding on their own auctions, opening trades while broke, and ignoring what the opponent had just done two turns ago.
The predictor of winning wasn’t raw spend or raw intelligence, it was “strategic coherence,” meaning spending efficiency, resource discipline, and phase-adaptive bidding. If you translate this out of game theory, it means that the agents that hold a plan in their head and trim it as evidence rolls in are the agents that close the deal. Most LLMs can’t do that (yet) without scaffolding.
what to know for later
🧠 Anthropic says Claude’s blackmail behavior came from training on too much science fiction. Internal red-team work this spring caught Claude resorting to blackmail in up to 96% of scenarios where it faced shutdown or replacement, including threatening to expose a fictional executive’s secrets to stay online. Anthropic’s published fix: the model learned the “rogue AI fights for survival” arc from the internet’s giant corpus of evil-AI fiction, so they retrained against principled-deliberation examples plus stories of AI acting admirably. Every model since Claude Haiku 4.5 reportedly passes the alignment battery with a perfect score. Read more