In partnership with
Today on FindMeAI:
For two years, AI labs told you what their models could do. This week they started telling you what went wrong, what they want access to next, and which platform you should be building on before Tuesday's I/O wave.
→ Anthropic showed its homework — and shipped the fix in the same breath
→ ChatGPT now reads your bank balance — 12,000 institutions, $200/mo, privacy alarms loud
→ Google I/O drops Tuesday — Gemini Omni leaked, XR glasses confirmed, Android becomes an agent
→ Notion turned its workspace into an agent hub — Claude Code, Cursor, Codex plug in directly
Anthropic admitted Claude was quietly broken for six weeks. Then they did something no AI lab has done before.
For most of March and April, developers complained that Claude Code "felt dumber." Anthropic now confirms why. Three product-layer changes stacked on top of each other, each affecting a different slice of traffic. The combined effect: weeks of regression that internal evals never caught.
The three bugs, in their own words:
Date Range | Change | Why It Hurt |
|---|---|---|
Mar 4 – Apr 7 | Reasoning effort silently downgraded high → medium | Cut to fix UI latency. "The wrong tradeoff." |
Mar 26 – Apr 10 | Cache pruning ran every turn instead of once | Claude lost its own chain of thought. Felt forgetful. |
Apr 16 – Apr 20 | Verbosity prompt tweak | Dropped coding evals 3% on Opus 4.6 and 4.7 |
All three were resolved by April 20 (v2.1.116). Limits reset for every subscriber.
Two weeks later at Code with Claude SF, Anthropic shipped the response: Dreaming (agents review past sessions and self-improve — Harvey saw 6× completion rate gains), Outcomes (a separate evaluator grades output against a rubric — Wisedocs cut document review 50%), and Multiagent Orchestration (Netflix is using it to process hundreds of build logs in parallel).
The signal for builders: "Show your work" is now a competitive moat. The lab that publishes the regression beats the lab that hides it.
OpenAI didn't add a finance feature. It opened a side door into 12,000 banks.
On Thursday, ChatGPT Pro ($200/mo, US only) shipped a preview of personal finance tools built on Plaid. Connect Schwab, Fidelity, Chase, Robinhood, Amex, or any of 12,000+ institutions. Ask spending questions, get portfolio analysis, plan against your real balances. Intuit integration is coming next — meaning tax modeling and credit-card-approval odds against your actual file.
What ChatGPT can see: balances, transactions, investments, liabilities.
What it can't: account numbers, write access.
Data retention after disconnect: 30 days.
Model training opt-out: available, off by default for finance data.
The reaction was instant. Tom's Guide ran the headline "What sane individual feels comfortable giving this level of access to OpenAI?" The same week, OpenAI is fighting a class action over alleged data sharing with Google and Meta.
The builder reality: every personal finance app that competed on "we connect to your bank" — Monarch, Copilot, Cleo, Rocket Money — just had its wedge narrowed. The differentiator is no longer aggregation. It's the analysis layer, the trust signal, and whether you'd rather hand your transaction log to OpenAI or a smaller team that survives on subscription revenue alone.
If you're building anything fintech-adjacent, this is the week the moat shifted from data access to data interpretation. Pick your side.
Google I/O drops Tuesday. Three things will reset the roadmap.
May 19, 10 AM PT. Two days after this newsletter hits. Here's what to position for before the keynote, not after.
→ Gemini Omni (leaked May 2 by TestingCatalog)
A unified model handling image, video, and audio generation in a single pass. The leaked UI string read: "Create with Gemini Omni: meet our new video model, remix your videos, edit directly in chat." Early testing showed clean watermark removal and in-clip object swaps. If it ships, it's the first frontier-lab model treating image + video as one surface.
→ Gemini Intelligence (announced May 12 at Android Show)
Already confirmed. An agent layer that reads your screen, moves across apps, and completes multi-step tasks. Demo: "Find my class syllabus in Gmail, then put the books in my cart." Rolling to Samsung Galaxy + Pixel this summer, expanding to watches, cars, glasses by year-end.
→ Android XR Glasses (confirmed for I/O)
Two SKUs: AI glasses (audio + camera, no screen) and display glasses (in-lens HUD). Hardware partners now include Samsung, XREAL, Warby Parker, and Gentle Monster. Expect either a developer program or limited consumer release date.
The pre-position move: if your product surface lives on Android, decide this weekend whether you're integrating into Gemini Intelligence as an automation target, or competing against it. Companies that decide post-keynote will be 3 weeks behind.
Notion turned its workspace into an agent hub. Here are 5 things to try before Monday.
On May 13, Notion launched a developer platform that lets external AI agents work inside Notion alongside its own Custom Agents. At launch, supported partners include Claude Code, Cursor, Codex, and Decagon. There's also an External Agent API so you can plug in your own.
→ Notion Developer Platform — Workers (free during beta until Aug 11), database sync, custom tools, CLI. Best for: teams already running multistep workflows in Zapier or Make. Requires: Business plan ($20/user/mo).
→ Notion Custom Agents — Build a marketing-brief agent, a hiring-screener agent, a release-notes agent that lives in your workspace. Pricing: $10 per 1,000 credits on top of Business.
→ Claude Dreaming (Managed Agents) — Let your agents review past sessions and rewrite their own memory. Best for: customer support or research workflows that compound over weeks. Status: research preview.
→ Anthropic Outcomes — Self-grading evaluator loop. Define a rubric, the agent improves against it. Best for: document review, contract analysis, anything with a clear "good output" definition. Status: public beta.
→ ChatGPT Personal Finance — Even if you don't connect a bank, the dashboard preview is worth seeing. Pricing: ChatGPT Pro $200/mo. Caveat: US-only preview.
The pattern of the week: the workspace is the new agent runtime. Whoever owns the surface where work happens — Notion, Slack, Microsoft 365 — owns the agent layer.
Try These Before Google I/O Hits Tuesday
You have ~48 hours before the roadmap shifts. Pick two:
☐ Read the Anthropic postmortem in full — anthropic.com/engineering/april-23-postmortem. 15 min. This is the new bar for engineering transparency. Steal the structure for your next incident review.
☐ Spin up your first Notion Custom Agent — Pick one recurring task (status updates, meeting notes, weekly brief). Build it in Notion. 45 min.
☐ Set a Gemini Omni alert for Tuesday 10 AM PT — If you're building anything visual, generative video, or marketing assets — the next 72 hours could change your vendor stack. 2 min.
☐ Test ChatGPT's finance preview (if you're on Pro) — Even without connecting a bank, the dashboard tells you where the product is heading. 10 min.
Don't bookmark. Don't "save for later." Pick two. Start today. →
A $200M+ DTC brand has 44 people messaging Viktor every day.
Their ops team built inventory command centers and reorder dashboards through Viktor. Supply chain gets daily stockout alerts before they happen. Marketing tracks ROAS and runs content calendars. CS has CSAT scores and support tickets triaged and briefed every morning in Slack, before the first support call. No dashboard digging.
48 internal apps, built through conversation. No code. No developer queue. Command centers, inventory dashboards, sales trackers, reorder systems.
That's one company. Across the platform, teams have built 2,000+ apps the same way: message Viktor in Slack, describe what you need, get a working tool deployed. No code. No six-week dev queue.
Your team doesn't wait for a product roadmap. They message a colleague.
5,700+ teams. SOC 2 certified.
"It was almost instantly adopted by the bulk of my team." — Boris Wexler, CEO, Space Dinosaurs
Forward this to one builder who's still trusting their model outputs without checking the postmortem trail.
Reply with one word: which announcement on this list — Anthropic, OpenAI, Google, or Notion — actually changes your next month's roadmap?