🌟 Editor's Note: Recapping the AI landscape from 03/24/26 - 03/30/26.
🎇✅ Welcoming Thoughts
Welcome to the 35th edition of KPAI Weekly.
What’s included: company moves, a weekly winner, AI industry impacts, practical use cases, and more.
Leaked Slack Messages from Sam Altman show him saying he tried to “save Anthropic” on the government deal that went south.
OpenAI Foundation pledged $1B in nonprofit spend over the next year
Grok 5 Missed its Q1 Deadline
New lawsuit over Meta’s smart glasses
Claude adjusted its usage limits and they’re very limiting
Hoping they roll them back a bit
Anthropic IPO could come as early as October
The AI landscape continues to emulate the TV show ‘SIlicon Valley’ (See TurboQuant below).
Let’s get started—plenty to cover this week.
👑 This Week’s Winner: Google // Gemini
Google had the most well-rounded week in the race. Between a lab breakthrough that rattled chip stocks, an aggressive play to poach users from ChatGPT and Claude, and a major robotics partnership, they made moves across research, product, and enterprise. Here's the recap:
TurboQuant: Google Research unveiled a memory compression algorithm that shrinks AI working memory by at least 6x with near-lossless quality. Still a lab result, not yet in production. Very impressive, working memory is a huge commodity right now.
Chat History Import: Google shipped migration tools letting users port their full conversation history and personalized memories from ChatGPT and Claude directly into Gemini. Was looking for this from GPT → Claude a few months ago. Smart move from Gemini.
DeepMind + Agile Robots: Google DeepMind partnered with Agile Robots to integrate its Gemini Robotics foundation models into 20,000+ deployed robotic systems globally, starting with high-value manufacturing. Very cool.
That’s not all: Google also upgraded Lyria 3 Pro, their AI music generator that can now produce full songs up to 3 minutes long. They also updated how AI articles appear in search, dropping AI content farms in traffic by 60-80%. Finally, Gap partnered with Gemini to let shoppers browse and buy products directly inside the AI assistant.
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.
🟠 Anthropic // Claude
Pentagon Ruling: A federal judge blocked the Pentagon's "supply chain risk" designation in a 43-page opinion, calling it "classic illegal First Amendment retaliation." Good for Anthropic.
New Model Leak: A data leak exposed a draft describing Claude Mythos as a "step change" over Opus 4.6 with dramatically higher scores across the board. This model would be the best in the world. I’m excited.
Computer Use on Mac: Claude can now open apps, click buttons, type, scroll, and navigate macOS as a full desktop agent. Similar to Cowork but available in Claude Code too!
🟢 OpenAI // ChatGPT
Sora Shutdown: OpenAI discontinued Sora after burning an estimated $15M per day against just $2.1M in total lifetime revenue. Disney's $1B licensing deal collapsed with the move. This is the right move for OpenAI. Sora was unnecessary money pit.
SoftBank $40B Loan: SoftBank secured a $40 billion unsecured, 12-month loan from JPMorgan, Goldman Sachs, and four Japanese banks to fund its $30B commitment to OpenAI's $110B round. Strong IPO signal here.
Shopping Overhaul: OpenAI pivoted from its Instant Checkout feature to a new agentic protocol. Walmart launched an in-ChatGPT app, joined by Target, Sephora, Best Buy, and Home Depot. OpenAI still trying to figure out in-app commerce.
⚪️ NVIDIA
Stock Slides: NVIDIA fell roughly 3% as the Nasdaq posted its worst weekly decline since April 2025. I’m sure it will bounce back (not financial advice).
Antitrust Probe: Senators Warren and Blumenthal launched an antitrust probe into NVIDIA's $20B Groq deal, accusing the company of structuring a "reverse acquihire" to evade review.
China H200 Restart: NVIDIA confirmed it is restarting H200 manufacturing for China after a 10-month freeze, with initial orders from ByteDance, Tencent, Alibaba, and DeepSeek. Nice - 25% of China sales still going to US Govt.
🔴 xAI // Grok
Baltimore Deepfake Lawsuit: Baltimore became the first U.S. city to sue xAI, alleging Grok generated over 3 million sexualized images, including 23,000+ depicting minors, in a single 11-day span. Same problem, more lawsuits.
Co-Founders All Gone: The final two xAI co-founders departed this week, completing a total exodus of all 11 original founders. SpaceX and Tesla executives have been brought in to restructure.
Doubling Down on Video: Hours after OpenAI killed Sora, Elon posted "we are doubling down" on Grok Imagine video generation. Aurora engine already generated 1.245 billion videos in January alone. Door is wide open in the video-gen market.
🔵 Meta // Meta AI
Safety Verdicts: A Santa Fe jury ordered Meta to pay $375 million, the first time a state prevailed at trial against a major tech company over child safety. The next day, an LA jury found Meta liable in the first-ever social media addiction trial. Good for society, bad for Meta.
More Layoffs: Meta cut roughly 700 employees across Reality Labs, Facebook, recruiting, and sales in its second 2026 layoff round. Continued aggressive AI hiring. Continues to be all-in on AI.
El Paso Data Center: Meta increased its El Paso data center commitment from $1.5 billion to $10 billion, targeting 1 gigawatt of capacity by 2028. More compute. More compute. More compute.
🤖 Impact Industries 🚑
Robotics // San José Airport Humanoid
San José Mineta International Airport deployed "José," a multilingual humanoid robot built by local startup IntBot, at Terminal B as a four-month pilot ahead of the 2026 FIFA World Cup. The robot uses vision, audio, and language processing to assist travelers in 50+ languages — one of the first humanoid deployments in a major U.S. airport. A small preview of what airports, stadiums, and public spaces could look like in years to come.
Read the Story
Healthcare // 5-Second Heart Detector
Noah Labs received FDA Breakthrough Device Designation for Vox, a tool that detects worsening heart failure up to three weeks before hospitalization by analyzing a 5-second daily voice recording. No hardware, no implants, no in-person visit. Validated across 5+ multicentre clinical trials with Mayo Clinic, UCSF, and Charité Berlin using the world's largest exclusive dataset of 3+ million heart failure patient voice samples.
Read the Story
💻 Interview Highlight: Olivia Moore on The Deep View
Interview Outline: Olivia breaks down the latest a16z research on the Top 100 GenAI apps, highlighting how AI has shifted from a standalone destination to an integrated OS-level utility. She explores why persistent memory is the new competitive moat for startups and how AI coding agents have created a "YouTube Moment" for software development.
About the Interviewee: Olivia Moore is a partner at Andreessen Horowitz (a16z) specializing in consumer AI. She tracks real-world adoption data to identify which platforms are actually retaining users and how the "Big Three" (OpenAI, Anthropic, Google) are beginning to diverge in their technical personas.
Interesting Quote: "Memory is the most underrated feature in consumer AI right now; switching costs are finally coming down because models actually know you."
Condensed Interview Highlight — Olivia Moore (a16z)
Host: What do you mean when you say 2026 is the year of the "Big Tech Unlock"?
Olivia Moore: For years, AI was something you had to "visit" in a separate app. This year, the major platforms have finally integrated these models deep into the OS layer. AI is now an invisible utility powering your browser, email, and calendar without requiring manual prompts for every task.
Host: Why do you think memory is the most underrated feature in AI right now?
Olivia Moore: Historically, every AI conversation was a "first date." Now that models have persistent memory, switching costs are increasing. If an AI already knows my writing style, business goals, and schedule, I’m much less likely to leave for a slightly smarter model that doesn't know me at all.
Host: You mentioned that software is having a "YouTube Moment." Can you explain that?
Olivia Moore: Before YouTube, making a video was hard. Afterward, anyone could be a creator. We are seeing that same shift in software. With AI coding agents, one person can now build and ship an app that used to require ten engineers, leading to an explosion of niche software built by non-programmers.
Host: Are we seeing the major AI platforms start to specialize, or are they all still racing for the same goal?
Olivia Moore: They are diverging. ChatGPT is the "everything app" for general utility, while Claude has carved out a niche for creative writing and deep technical analysis. Meanwhile, Gemini is winning on integration because it’s baked into the entire Google Workspace ecosystem people already use for work.
Host: If AI is doing the execution, what is left for the human?
Olivia Moore: Humans are for ideas, and AI is for execution. We’re seeing a shift where the "value" of a person isn't their ability to do the work, but their ability to define the problem and judge the quality of the output. The "managerial" layer of society is expanding because everyone now has a team of AIs to manage.
👨💻 Practical Use Case: Dual Prompting
Difficulty: Basic
I thought I covered this a while back, but when I went to search for it to recap, it wasn’t there. Dual prompting is novel but something I do fairly often, using one LLM to write a prompt for another. Instead of spending 10 minutes crafting the perfect prompt yourself, ask Claude or GPT to write it for you, then paste that prompt into whichever model you want to do the actual work.
Where this shines:
You know what you want but can't articulate it cleanly
You want to take advantage of one model's strengths while prompting another
You need a structured prompt with examples, constraints, and formatting rules but don't want to build it from scratch
For example, I'll ask Claude to write me a detailed Gemini prompt for a specific task, then drop that into Gemini, or vice versa. The output quality jumps noticeably because the prompt itself is already well-structured. It takes 30 seconds and usually gets you 80% of the way to a perfect result.
Simple, free, and one of the most underrated ways to get better output from any AI tool.
Learn More Below ⬇️
📊 Startup Spotlight
Interloom
Interloom — The "Context Graph" for Enterprise Memory.
The Problem: Most AI agents are "technically" capable but "contextually" blind. They can read a manual, but they don't know the "tacit knowledge," the undocumented judgment and weird edge cases that an employee with 15 years of experience just knows. Roughly 70% of operational decisions are never written down, which is why most AI agents fail when they hit real-world complexity.
The Solution: Interloom builds what it calls a Context Graph. It’s essentially a Google Maps for company operations. By ingesting millions of real-world emails, support tickets, and call transcripts, it maps out how experts actually solve problems rather than how the manual says they should. This creates a persistent "corporate brain" that AI agents can plug into to make smarter, human-like decisions.
The Backstory: Founded in 2024 in Munich by Fabian Jakobi, the company emerged from stealth with an initial $3M round from Air Street Capital. They just closed a $16.5M Seed round (March 2026) led by DN Capital. They already count industry giants like Volkswagen, Commerzbank, and Zurich Insurance as customers, using the tech to bridge the gap between documentation and reality.
My Thoughts: This is similar to RAG but not quite. RAG (Retreival Augmented Generation) is a more linear process of recalling data, while Interloom uses context grahs to map relational responses and structures. Moreso an enterprise tool, but a look at how recognition of prior documentation is evolving with the move into agentic AI.
“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
36 Weeks of AI race tracking in Figma ⬇️.
Till Next Time,
Noah from KPAI