Koda Intelligence Briefing

Google Stitch is free and it's already breaking things: Google launched Stitch this week — a free AI design tool that builds working, board-quality reports and interfaces from plain-language prompts. No wireframes. No Figma. No design agency. The author tested it on real financial data: uploaded a photo of hand-drawn numbers, typed one prompt, and got back a premium executive financial report complete with a clean headline, 12-month trend charts with green/red performance indicators, a comparison table, and three FP&A driver callouts. Output quality equivalent to three hours of analyst work on a Sunday night.

Figma lost 12% the day Stitch launched: That market signal is the story. Google just made free what professional teams used to pay $50-100 per seat per month for. This pattern does not stay in design software. It moves to every category of professional work. Finance reporting is next. Legal document generation after that.

The prompt is the FP&A judgment: The key insight in this issue is not about the tool. It's about how you prompt it. "Don't ask for a design. Ask for FP&A judgment built into a visual format." That is a completely different instruction. Generic AI outputs look generic because the prompt was generic. Specificity in the prompt produces specificity in the output. The author provides a copy-paste prompt template for generating CFO-ready reports.

The professional work disruption curve is accelerating: Stitch is a signal, not an endpoint. AI is systematically making professional tools free. Each disruption follows the same pattern: tool launches free, incumbents drop in market value, early adopters get an unfair advantage for 6-12 months before the skill becomes table stakes. You are inside the 6-12 month window right now for finance AI tools.

You can see the live demo at google-stitch-finstory.netlify.app

Stitch critiques its own work in real time and responds to voice commands

One-click deploy to production — no engineering required

BYOC (Bring Your Own Cloud) now in public preview on AWS, GCP, and Azure: This is a major enterprise unlock. BYOC lets you run Pinecone's data plane inside your own cloud account. Your vectors, metadata, and queries never leave your environment. Setup is fully self-serve via a Pulumi-based wizard with pull-based operations that execute locally in your cluster. No SSH, no VPN, no inbound network access required. This removes the #1 objection for enterprise AI deployments: data sovereignty.

HIPAA compliance for Standard plan at $190/month: Previously available only on Enterprise. Starting now, Standard plan customers can add HIPAA compliance including encrypted data storage, audit logging, enhanced security controls, and BAA execution support. This opens the healthcare AI market to mid-market companies that couldn't afford Enterprise pricing.

Garbage Day — How Pinecone Safely Deletes Billions of Objects: Deleting data at scale is harder than adding it. Pinecone's Janitor garbage collection system safely removes billions of objects from immutable blob storage without disrupting live queries. The post covers why deletion in distributed systems requires more than a delete call, and how Pinecone keeps storage costs low without sacrificing query reliability. Essential reading if you're running a production vector database.

When "Performance" Means Two Different Things: Speed and accuracy are both called "performance" in AI, but optimizing for one trades off against the other. This post breaks down the distinction for production systems: when you need latency-optimized retrieval vs. when precision matters more. A practical framework for setting the right performance target before you build.

Upcoming webinar: Pinecone Platform Series — Slab Architecture (March 24)

Upcoming webinar: Nuances of Hybrid Search (March 26)

Vector Search for Agentic AI hands-on introduction series starting April 23

Pinecone + n8n official node: The official Pinecone Assistant node for n8n is live, enabling no-code vector search workflows in automated pipelines.

10M-vector benchmark guide: New guide covers how to benchmark Pinecone at 10M+ vectors with real performance data.

Meta is building a CEO agent: Mark Zuckerberg is building an AI agent to help him run Meta. He envisions a future where everyone has their own AI agents handling tasks, communication, and decision support. This is not a chatbot experiment. This is the CEO of a $1.5T company betting his own workflow on autonomous AI.

Amazon's AI-native phone: Amazon is developing a phone built around AI from the ground up, not AI bolted onto an existing OS. If it ships, it would be the first major consumer hardware designed for agentic workflows by default.

SpaceX + Tesla chip factory: Elon Musk's "Terafab" AI chip factory will combine SpaceX and Tesla resources to manufacture custom AI silicon. The goal is to reduce dependency on NVIDIA and build chips optimized for Tesla's self-driving and xAI's Grok infrastructure.

85% of YC's W26 batch is AI-first. 56 of 198 companies are building fully autonomous agents positioned as AI employees.

Gemini (crypto exchange) hit with class-action lawsuit over IPO disclosures.

Terafab is Musk's play for AI hardware independence: The Rundown breaks down how the factory would produce custom chips for both Tesla FSD training and xAI's Grok models. Current NVIDIA dependency is a bottleneck for both companies. Vertical integration is the move.

Google Stitch website redesign tutorial: Step-by-step guide on using Google Stitch to completely redesign a website from scratch using only prompts. The tool generates responsive HTML/CSS and deploys in one click. This is the practical how-to the finstory AI newsletter teased.

Plastic-based treatment for Parkinson's: Scientists developed a plastic material that can deliver dopamine-producing cells directly to the brain, bypassing the blood-brain barrier. Early results show significant motor function improvement in animal models. This could change the treatment paradigm for neurodegenerative diseases.

AI in scientific discovery is accelerating: This week's issue highlights how AI models are now being used to predict protein folding, drug interactions, and treatment outcomes at speeds that compress years of research into weeks. The pattern: AI handles the combinatorial explosion of possibilities, humans handle the judgment calls.

Use AI daily, do not just study it: Dharmesh Shah (HubSpot co-founder) argues that hands-on experience teaches more than reading research papers. The most important thing is not tracking every development but actually using AI in your day-to-day work. He recommends tools like Company Research Agent, Prospect Extractor, and ICP Builder on agent.ai.

Schedule quarterly AI opportunity reviews: Every quarter, take your top 3 growth obstacles and check if new AI tools can now address them. Keep a "not yet, but soon" list. Companies that evaluated AI once and gave up fell behind competitors who revisited regularly.

Try new tools with specific problems in mind: Problem-first evaluation beats aimless exploration. When Claude 3.7 Sonnet launched, Dharmesh tested it specifically for code review efficiency and found hybrid reasoning was remarkably effective at understanding complex codebases.