🌟 Editor's Note: Recapping the AI landscape from 04/07/26 - 04/13/26.
🎇✅ Welcoming Thoughts
Welcome to the 38th edition of KPAI Weekly.
What’s included: company moves, a weekly winner, AI industry impacts, practical use cases, and more.
Switched my Claude accounts and usage got a bit better.
Interesting report from Stanford across AI, education, and more.
Should have a good interview lined up for next week.
xAI sued Colorado over AI discrimination law.
The New Yorker did an ugly profile on OpenAI CEO Sam Altman.
Let’s get started—plenty to cover this week.
👑 This Week’s Winner: Anthropic // Claude
Anthropic's week was headlined by Mythos Preview, a model so capable it prompted an emergency meeting between the Treasury Secretary, Fed Chair, and bank CEOs. Revenue continues to grow. And the industry is calling Claude the most important AI product on the market. Here's the recap:
Mythos Preview + Project Glasswing: Anthropic restricted its most powerful model for initial testing to 12 partners including AWS, Apple, and Microsoft after it autonomously found thousands of vulnerabilities across major operating systems and browsers. Definitely will be powerful. Still, this is likely overhyped.
Revenue Passes $30B: As mentioned previous, Anthropic's annualized revenue tripled from ~$9B at end of 2025. There are now over 1,000 business customers spending $1M+ per year. Claude Code alone is generating $2.5B+ annualized. It’s the best. I’d recommend a Claude team subscription at a minimum for all business operators.
"Claude Mania" at HumanX: At the 6,500-attendee HumanX conference in San Francisco, Claude was the most-discussed AI product — a dramatic shift from 2025 when OpenAI led. Glean's CEO called it "a religion." It’s about time everyone caught on. Anthropic remains the best tool, even with their recent usage issues.
Strong week but still some pitfalls: Claude Code quality regressions continued, with developers reporting broken logic and API cost spikes. The D.C. Circuit ruled against Anthropic in the Pentagon battle, upholding the "supply chain risk" designation. As an aside, Anthropic also hosted a summit with Christian leaders to discuss Claude's moral development. They need to stop trying to launch new stuff and make sure what they have is working and they’ll be just fine. OpenAI had the same issue for a bit.
From Top to Bottom: Open AI, Google Gemini, xAI, Meta AI, Anthropic, NVIDIA.
⬇️ The Rest of the Field
Who’s moving, who’s stalling, and who’s climbing: Ordered by production this week.
🔵 Meta // Meta AI
Muse Spark Launch: Meta launched its first model from Superintelligence Labs — natively multimodal with three reasoning modes. Scored 4th globally but trails significantly in coding. Nice! It was never going to be competitive with the others at launch but the important thing is they are finally releasing stuff again. Launch. Iterate. Iterate. Success. Iterate.
AI Zuckerberg Clone: Meta is building a photorealistic AI character modeled on Zuckerberg to interact with 79,000 employees when he's unavailable. He's reportedly spending 5-10 hours per week coding on AI. A little nuts but pretty interesting nonetheless. I wonder how good it is.
Proprietary Pivot: Muse Spark launched with no open weights, no public API — abandoning the open-source approach that defined the Llama era. Zuckerberg says open models are still coming. I hope so!
🟣 Google // Gemini
Mental Health Overhaul: Google introduced a crisis interface connecting users to the 988 hotline after a wrongful death lawsuit. Gemini can no longer claim to be human or simulate intimacy. $30M pledged to crisis hotlines. Good move.
3D Models in Chat: Gemini now generates interactive 3D models directly inside conversations. Users can adjust variables like gravity and orbital velocity in real-time. Sweet, I have to try this out. AI is giving us so much opportunity to do cool stuff that simply wasn’t on the table prior.
Moodle Integration: Google is embedding Gemini and NotebookLM into Moodle, the world's most popular open-source learning management system. Teachers can embed custom AI tools into curriculum starting May. Cool, based on the report in ‘Welcoming thoughts’ it seems like students are still using AI a boatload more than teachers. Maybe this will help enact some change.
⚪️ NVIDIA
8-Day Winning Streak: NVIDIA posted its longest stock streak since 2023, climbing ~14% over eight sessions on a Trump-Iran ceasefire, soft inflation data, and TSMC's record Q1. Not a big surprise here.
Jensen on AI Layoffs: CEO Jensen Huang told Cramer that companies cutting jobs over AI are doing so "because you're out of imagination. For companies with imagination, you will do more with more." I agree completely!
Lumentum Orders Through 2028: NVIDIA's $2B investment partner says demand is so intense its 2028 capacity is projected sold out within two quarters. JPMorgan raised their target from $565 to $950. Never doubt NVIDIA.
🟢 OpenAI // ChatGPT
Altman Home Attacked: A 20-year-old from Texas threw a Molotov cocktail at Altman's home at 4 AM, then was arrested at OpenAI HQ an hour later carrying an anti-AI manifesto with a list of executive names and addresses. A separate unsuccessful attack followed days later. People are crazy.
New Pro Tier: OpenAI launched a $100/month ChatGPT tier offering 5x more Codex usage. Codex crossed 3 million weekly active users — 5x growth in three months. I’m hearing great things about Codex (GPT’s coding tool). I’ll likely be using it more in the coming months in addition to Claude Code.
Self-Serve Ads Manager: OpenAI lowered its ad pilot minimum from $250K to $50K with a $60 CPM. U.S. ads already at $100M+ annualized with 600+ advertisers. I’d like to see some numbers on how this is doing.
🔴 xAI // Grok
Leadership Overhaul: xAI installed SpaceX's Michael Nicholls as new president. He told staff xAI is "clearly behind" and set a two-month deadline to improve. Curious to see if this will make meaningful impact. I like the idea of a SpaceX guy leading xAI.
Environmental Challenge: The NAACP and other groups filed to revoke a permit for 41 gas turbines powering xAI's Colossus 2 data center, arguing the process was rushed.
SpaceX $5B Loss: SpaceX swung from ~$8B profit in 2024 to ~$5B loss in 2025 despite revenue growing to $18.5B. xAI integration costs were the primary driver. Makes sense. I’m a strong believer in this joint company.
📢 Impact Industries 💻
Marketing // OREO GEO
Mondelez overhauled its $3.5B digital strategy after discovering Oreo was virtually invisible to AI chatbots. By unblocking crawlers and restructuring sites for Generative Engine Optimization (GEO), Oreo's AI citation rate jumped from 10% to 70%. This landmark case study proves Fortune 500 brands must now optimize for machine-readability rather than just human search. As AI agents begin driving consumer purchases, GEO is becoming the essential framework for brand discovery in the agentic era.
Read the Story
Engineering // Visa for AI Agents
Visa launched the first payment infrastructure built specifically for AI agent shopping, letting merchants make product catalogs discoverable to AI platforms and accept agent-initiated payments across multiple protocols. Currently piloting with AWS and others, with general availability planned for June 2026. The move signals a broader shift: brands will soon need to optimize not just for human shoppers but for autonomous AI agents making purchasing decisions on their behalf, reshaping how commerce works at the infrastructure level.
Read the Story
💻 Interview Highlight: Simon Willison on Agentic Engineering
Interview Outline: Simon Willison discusses the "November Inflection Point" where coding agents became reliable enough to move from prototyping to professional "agentic engineering". He explains the "Dark Factory" pattern—where AI writes, tests, and reviews code without human intervention—while warning of an impending "Challenger Disaster" in AI security caused by prompt injection.
About the Interviewee: Simon Willison is the co-creator of Django and a prolific open-source developer. He coined the term "prompt injection" and is widely considered one of the most practical voices on how to use AI to actually build products rather than just "vibe coding".
Interesting Quote: "I can churn out 10,000 lines of code in a day... but using coding agents well is taking every inch of my 25 years of experience.”
Condensed Interview Highlight — Simon Willison (Lenny's Podcast)
Lenny: You’ve talked about a massive shift that happened in November 2025. What changed?
Simon Willison: In November, GPT 5.1 and Claude Opus 4.5 crossed a threshold where agents went from "mostly working" to doing what you told them almost all the time. A lot of engineers woke up and realized they could suddenly churn out 10,000 lines of code in a day, turning software engineering into a bellwether for all other information work.
Lenny: What is the difference between "vibe coding" and "agentic engineering"?
Simon Willison: Vibe coding is for when you don't care about or understand the code; it's great for personal prototypes where bugs only hurt you. But for professional software deployed to a million people, you need "agentic engineering"—a deep discipline of using agents to build, debug, and test code to a higher standard than we could before.
Lenny: What happens to the people in the "middle"—not the juniors or the senior experts?
Simon Willison: The "middle" is in the most trouble because they don't have the deep expertise to amplify with these tools. My advice is to lean in and use AI to take on much more ambitious projects. I've shaved two-month learning curves down to zero, allowing me to automate things on my Mac or even get better at cooking.
Lenny: If writing code is now "cheap," where does the human brain still provide value?
Simon Willison: Prototyping is almost free now, which means our "superpower" has been democratized. The value moves to the ideation and proof phases—deciding what problems to solve and judging the quality of the results. The human brain is still essential for "agency," which is something AI will never have.
Lenny: What is the most dangerous security risk businesses are currently ignoring?
Simon Willison: I call it the "Lethal Trifecta". It occurs when an agent has access to private info, is exposed to malicious instructions (like an untrusted email), and has a mechanism for exfiltration. If your agent can reply to your email, it can be tricked into forwarding your private sales projections to a stranger.
👨💻 Practical Use Case (Issue 09 Revisited): GPT Projects
Difficulty: Beginner
Most people use ChatGPT in a one-off way: open a chat, ask a question, close the window. But GPT Projects take things further by giving the model memory and context across an entire workspace.
With Projects, you can:
Store project wide instructions: Instead of pasting the same context over and over, you can set persistent instructions the model always follows.
Keep better memory: The model remembers what you’ve uploaded, what you’re working on, and the style you prefer, so you don’t have to repeat yourself.
Organize files and tasks together: Drop in documents, notes, or datasets, and the model can reference all of them in one place.
Example: Imagine you’re writing a research paper. In a normal chat, you’d have to re-upload your sources every time and restate your formatting rules. In a GPT Project, you upload your sources once, set your style guide in the instructions, and the model can consistently help you draft sections, generate citations, and keep the tone aligned across the whole paper.
I’ve found projects useful in not only organizing my chats, but also the context aspect referenced above. It still leaves something to be desired but definitely a step up from scattered chats. Create a project folder and try it out if you haven’t already!
Issue 38 Update: Project capabilities have improved a bit since Issue 09, but they’re still as useful as ever. If you’ve been following along with the newsletter you know I’m now barely using GPT, but I’ve set up projects on my Claude team plan that I use every day. Easy way to keep context on multiple, well, projects, at once. Main capabilities still include Memory (what Claude understands about the project over time), Context (the ability to upload tons of project specific documents), Instructions (how you want chats within the project to respond), and of course Tools (ability to call outside tools such as Google Drive when needed). Basic but useful!
🗣️ Startup Spotlight
Wispr Flow
Wispr Flow — Turn speech into clear, polished writing in every app.
The Problem: Typing is a bottleneck. The average person types at 45 wpm, but we think much faster. Traditional voice dictation often creates more work than it saves, forcing you to go back and fix "ums," stutters, and formatting errors. For developers, leaders, and creators, the friction of the keyboard often kills their "flow state."
The Solution: Wispr Flow is a cross-platform "Voice OS" (Mac, Windows, iOS, and Android) that allows you to write 4x faster than typing. It doesn't just transcribe; it uses AI to understand intent, remove filler words, and structure your speech into perfectly formatted text in real-time. It works globally across any app—from coding in Cursor to sending Slack messages or drafting long-form content.
The Backstory: Founded in 2021 by Stanford AI researchers Tanay Kothari and Sahil Gupta, Wispr recently raised a massive $81 million to build out the future of voice-first computing. Backed by heavyweights like Reid Hoffman and Menlo Ventures, the team is focusing on high-stakes reliability, ensuring the platform is both HIPAA-ready and SOC 2 compliant for enterprise use.
My Thoughts: Talked with someone today who is a big fan of Wispr Flow. For those stream of conscious folks out there who juts like to get everything in their head on paper this is probably the tool for you. Turns clouded thoughts into coherent statements.
“It’s not likely you’ll lose a job to AI. You’re going to lose the job to somebody who uses AI”
- Jensen Huang | NVIDIA CEO
Till Next Time,
Noah from KPAI