🌟 Editor's Note: Recapping the AI landscape from 07/01/26 - 07/28/26.
🎇✅ Welcoming Thoughts
Welcome to the 2nd edition of KPAI Monthly.
What’s included: company moves, a weekly winner, AI industry impacts, practical use cases, and more.
Special Wednesday Edition.
My founding engineer has been telling me to check out Oh My Pi (One of the top agentic harnesses on the market right now).
I need to start using a real harness (outside of Claude Code) because GPT Sol and Kimi K3 are very good.
I’ll be designing and instructing an Ohio State PCE course starting in January of 2027 on Claude Automations.
Claude shared chats got leaked, then fixed. Not good, not concerned.
A lot of discourse around Open Models right now, see PUC.
New KPAI Website is live, try the minigame at the bottom of the home page!
I’ve written 72k lines on Claude Code this month, I’d estimate 90% of that was spoken versus typed.
Almost a year into KPAI (/NoahonAI) and the ops side of the business (Automations, Offerings, SOP’s, Structure, Company Templates etc.) is finally built out… time to start spending 24/7 on sales, PM, Mktg… and golf.
^ I believe this is where most businesses are heading.
Took a lot of trial and error but very happy with the framework we put together on AI Efficiency vs. Automation. Works great with the Audit system.
More on open models from Dario Amodei and Jensen Huang.
OpenAI released an interrupting voice agent… kinda cool.
One of OpenAI’s testing models breached set guardrail contain.
Anthropic Economic Index from June is interesting.
Kind of surprised at the lack of production from Google in Q2 2026.
Cool game in the Startup Spotlight this week!
Sunsetting the NoahonAI domain and comms at the end of July.
Another month with a surprise runner up in the AI race.
Is Meta turning a corner?
The inaugural (KP)AI race is over… more to come next month!
Let’s get started—plenty to cover this Month.
👑 This YEAR’S Winner: NVIDIA
NVIDIA takes the month, and win’s the year long race! It lined up more than $750 billion in deals in a single week, launched a security alliance, and shows no sign of stopping. Here's the recap:
Half a Trillion With SK: NVIDIA and Korea's SK Group signed letters of intent worth more than $500 billion: SK Telecom builds a 2-gigawatt AI factory on NVIDIA's platform, and SK hynix co-develops next-generation memory. Huang's words: "between us, we're going to do half a trillion dollars' worth of business." The market for NVIDIA is not going away.
The $250B OpenAI Backstop: NVIDIA is reportedly in talks to guarantee $250 billion in financing so OpenAI can lease a 10-gigawatt data center campus in Ohio, plus another $350 billion to finance OpenAI's chip purchases. The "you buy from me, I invest in you" structure spooked investors: the stock fell about 5% and Apple retook the top market-cap spot. Wow. I wouldn't worry about either of these companies.
Open Secure AI Alliance: Days after OpenAI disclosed its rogue-agent breach, NVIDIA launched an alliance to build free, open cybersecurity tools for AI, with founding members including Adobe, CrowdStrike, Hugging Face, and Dell. I like this. If open models are going to be more advanced, something like this is a big step in the right direction.
NVIDIA also made a reported $5 billion investment in Ilya Sutskever's (OpenAI Cofounder) Safe Superintelligence, put $1 billion into Korea's Naver alongside Brookfield to more than triple a data center there, and rolled out its DSX "AI factory" model, which backs cloud operators' buildouts in exchange for a share of the revenue.
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
Compute for Rent: Meta and BlackRock announced a $14 billion joint venture to build a 1-gigawatt data center campus in El Paso, and Anthropic is reportedly in preliminary talks for a two-year lease of up to $10 billion of Meta's computing power. This is impressive on both accounts! Maybe Meta can have more success in compute than they’re having in models.
First In-House Image Model: Muse Image, the first image model from Meta Superintelligence Labs, went live in Meta AI, Instagram Stories, and WhatsApp and debuted #2 on the LMArena leaderboard. Nice, good to see them releasing stuff again.
Meta's First Paid API: Muse Spark 1.1, an agentic model with a 1M-token context window and computer use, shipped with Meta's first-ever paid developer API, a real break from the free-Llama playbook. About time, not sure anyone uses this though.
🟠 Anthropic // Claude
Opus 5 Launches: Claude Opus 5 comes close to Fable 5's intelligence at half the price and is now the default on Claude Max. It caps a busy stretch: Fable 5 and Mythos 5 came back online in early July after the government lifted its export controls. Fable is still the best of the best.
Record a Skill: Claude can now learn a task by watching you do it once. Record your screen, talk through the steps, and Claude turns the recording into a reusable skill it can run on its own. Available on Pro, Max, and Team plans. Cool! Denoting workflows is the 1st step to increasing AI Efficiency.
Claude for Teachers: Verified US K-12 educators now get free premium Claude access with curriculum-aligned integrations. The same day, Anthropic committed C$10 million to Canadian AI research institutes. Nice, more teachers should be getting familiar with AI!
🟢 OpenAI // ChatGPT
ChatGPT Goes to Work: OpenAI launched ChatGPT Work, an agent mode that stays on a project for hours and delivers finished spreadsheets, slide decks, and small web apps, connected to Slack, Gmail, Drive, and Salesforce. Chat, Work, and Codex now live in one rebuilt desktop app. Continues their strong presence in the corporate market.
Interrupting Voice: GPT-Live is a new voice architecture that listens and speaks at the same time, so it handles interruptions the way a person does, and it now powers ChatGPT Voice. Cool! Voice models are improving slowly but surely.
Uncle Sam Altman: OpenAI reportedly proposed giving the US government a 5% stake, worth roughly $42.6 billion at its $852 billion valuation, as it works toward an IPO. Interesting, let the public share in AI benefits (and probably get friendlier regulation). Why not.
🔴 SpaceX // Grok
Grok 4.5 Launches: xAI shipped Grok 4.5, a 1.5-trillion-parameter model trained partly on data from Cursor, the coding tool SpaceX bought in June. It is priced at $2/$6 per million tokens and ranked #4 on the Artificial Analysis index at launch. Just found out I have a friend who knows the Cursor cofounder, pretty cool.
Grokflows: Workflows in Grok Build let users fan a task out across 128 AI agents, up to 1,024 for the biggest jobs, with deep research built in as a standard workflow. Nice, becoming standard for the top models. You can do this in Fable too and it will eviscerate your usage.
Grok Moves Into the Office: Outlook and Excel add-ins landed first, then a free Google Workspace add-on for Docs, Sheets, and Slides. Grok now sits inside both Microsoft's and Google's work suites. Grok feels like a little brother to Anthropic and OpenAI, it’s not going away any time soon.
🟣 Google // Gemini
New Flash Models: Google released Gemini 3.6 Flash (17% fewer output tokens than its predecessor), the budget 3.5 Flash-Lite, and 3.5 Flash Cyber, a security model limited to governments and trusted partners. The top model (3.5 Pro) is still missing. I’m not sure what they’re doing at Google but they are not releasing new top AI models. Pretty unimpressed.
950M Monthly Users: Pichai announced at Q2 earnings that the Gemini app passed 950 million monthly users, up from 750 million at the end of 2025 and roughly at parity with ChatGPT. Interesting, I’ve met 1 person who’s main AI model is Google Gemini (My sister).
Gemini Spark Lands on Mac: Google's always-on agent arrived as a Mac app for AI Ultra subscribers, organizing local files and building documents on its own. Who is using this.
🤖🏥 Impact Industries 🎓
Robotics // J&J's OTTAVA
Johnson & Johnson won FDA authorization for OTTAVA, the first surgical robot whose four arms are built directly into the operating table instead of parked around it on carts. The design takes up 30 to 50% less space in the operating room, and the table and arms move as one, so the team can reposition a patient mid-surgery without resetting each arm. OTTAVA is authorized for ten general-surgery procedures, including gastric bypass and gallbladder removal, and it is the first real challenger to Intuitive's da Vinci, which has held about 80% of this market for two decades.
This is awesome. I've been waiting for robotics and healthcare to come together a bit more, and I wonder how many more of our Impact Industries will blend as time goes on.
Read the Story
Education // Katy AI Ban
Katy ISD, a Texas school district with nearly 100,000 students, adopted one of the country's first grade-by-grade AI policies. Generative AI tools like Copilot and Gemini are banned outright from kindergarten through 6th grade so students build reading, writing, and math skills without algorithmic help. Supervised access begins in 7th grade through district-approved tools only, and high schoolers can use AI on approved assignments if they cite it. The district calls it preparing students for an AI workforce while protecting the years where foundational thinking develops.
I think I actually agree with this one. But I do believe the teachers should be using AI, and they should be teaching students about it too.
Read the Story
💻 Interview Highlight: Sam Altman (Relentless Podcast)
Interview Outline: Sam Altman examines how AI and exponential scaling laws have transformed the startup landscape over the last decade. He covers mental frameworks for operating in chaos, OpenAI's abrupt pivot from a quiet research lab to a product giant, the critical choice between AI liberty and authoritarianism, and why physical supply chains, specifically chips and energy, are the true bottlenecks to superintelligence.
About the Interviewee: Sam Altman is the CEO and co-founder of OpenAI, the leading artificial intelligence research lab and deployment company. A former president of Y Combinator, Altman has spent over a decade deeply embedded in the startup ecosystem, scaling operations and advising founders on how to leverage exponential growth curves to build generation-defining companies.
Interesting Quote: "The fight of the current moment is: are we going to head to a world of AI authoritarianism or liberty?"
Condensed Interview Highlight — Sam Altman (Relentless Podcast)
1. Interviewer: What is the biggest shift in how to build a startup today compared to 10 years ago?
Sam Altman: The biggest shift is clearly AI and the sheer speed with which a tiny team can execute today[cite: 2]. A 10-week-old startup built today operates at a scale and capability that would have required a year of work previously[cite: 2]. Startups always have their biggest inherent advantage when the ground is shifting rapidly and cycle times collapse, making this an unbelievable era to build a company[cite: 2].
2. Interviewer: If you want AI to continue scaling unabated, what is the biggest bottleneck you face?
Sam Altman: The absolute biggest bottlenecks right now are transistors and then electrons, in that exact order[cite: 2]. While we often focus on algorithms creating better algorithms, there is a massive lack of focus on building data centers that can build more data centers[cite: 2]. Ultimately, to produce high-quality, low-cost intelligence, we need physical world assets like massive energy grids and robots to stay on this ramp[cite: 2].
3. Interviewer: What do you think the biggest impacts are on getting this next part of the AI curve right?
Sam Altman: Beyond the standard safety and economic alignment issues, the defining fight of this current moment is between AI authoritarianism and liberty[cite: 2]. We have to decide whether we want a single "machine god" controlled by a small group, or if we want to widely decentralize this power into the hands of the public[cite: 2]. History shows that trading liberty for safety is a long-term net loss, which is why we must empower society to express its ideas with this technology[cite: 2].
4. Interviewer: How do you decide to pull the plug on a project that is already successful and well-funded?
Sam Altman: The hardest thing isn't killing a failing project; it is deciding to kill a good project to make your best project even better[cite: 2]. For example, when GPT-3 started working, we shut down our robotics initiatives, and recently when coding agents took off, we shut down highly anticipated projects like Sora and our browser[cite: 2]. It’s a painful realization, but you have to reorient your people and compute toward the most critical path for superintelligence[cite: 2].
5. Interviewer: How do you differentiate between a genuine, enduring trend and a hype bubble?
Sam Altman: A fake trend, like early VR, generates a lot of hype, but users don't actually design their daily lives around the experience, leaving the product to sit on a shelf[cite: 2]. A real trend is persistent; with tools like ChatGPT or coding agents, you see users engaging with it daily and integrating it fundamentally into their workflows[cite: 2]. When evaluating startups or new technologies, you have to look past the initial noise and focus strictly on that deep, recurring behavioral change[cite: 2].
👨💻 Practical Use Case: Open AI Models (Not OpenAI)
Difficulty: Mid-level
When Claude's and OpenAI's newest models launched this summer, the government held them back for weeks over cybersecurity concerns before clearing them for the public. That was possible because those top models live on their makers' servers.
Open AI models are the opposite: models you download and keep. Instead of only selling access through a website, the company releases the model file itself (the "weights," everything the model learned in training, saved as a file). Once it's on your computer, it runs like any other program. Here's what you can actually do with one:
With open AI models, you can:
Chat privately: The simplest use. Run a model like Google's free Gemma 4 on a regular laptop and you have a capable assistant that works offline, costs nothing per question, and never sends a word off your machine. Apps like LM Studio (Issue 31) make it a ten-minute setup.
Put it inside a harness: Last month's Practical use case covered agentic harnesses, the software wrapper that turns a model into a worker with tools and guardrails. The engine inside a harness is swappable, and an open model can be that engine: same agent behavior, no per-token bill, and your files never leave the building.
Make it yours: Because you hold the file, you can train the model further on your own material, your documents, your industry's language, and end up with a specialist the big labs don't sell.
Know the tradeoff: The top paid models still win on hard reasoning, and running the biggest open models well takes serious hardware. The sweet spot is private, repetitive, or high-volume work, not the hardest problems.
This is also why Washington is weighing rules for open models: there is no company server to restrict once the file is out in the world. This month China's Kimi K3, the largest open model ever released, beat the top paid models on several coding benchmarks, and strong coding is exactly the skill behind the cybersecurity concerns.
Learn More Below ⬇️
🤖 Startup Spotlight
Enigma Robotics
Enigma — Controlling a robot as easily as adjusting the volume.
The Problem: Hardware in robotics has advanced rapidly, but telling a robot what to do remains a nightmare. If loading a dishwasher takes 15 minutes of complex prompting or code tuning, people give up and do it themselves. The usability bottleneck is what's keeping robots out of daily life.
The Solution: Enigma is building an intuitive control layer for robotics. Instead of forcing users to act like software engineers, their system abstracts complex spatial tasks into simple inputs—making operating a machine feel as natural as turning a volume dial. To gather training data, they’ve launched a public site where anyone can remotely control 100 real AI-powered robots in real time.
The Backstory: Founded less than a year ago by prodigies Jonathan Jacobi (Microsoft's youngest-ever hire at 17) and Gal Niv, Enigma just emerged from stealth with a massive $71 million seed round. Co-led by Index Ventures and Ribbit Capital, with participation from Sarah Guo’s Conviction Partners and key angels from OpenAI, DeepMind, and Anthropic, it marks one of the largest robotics seed rounds of the year.
My Thoughts: We’ve talked about this in the past but yet another example of the shift toward “vibe everything”. In order for robotics to become mainstream, we absolutely need to be able to talk to them like humans. It’s just messaging, not speaking, on the site but I tried it out and it’s pretty cool, click the link to talk to a robot (Try the painting one)!
“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
Are you pro open models or pro regulation? Also, check out Ohio’s Claude stats (one notable emission on use cases).
Till Next Time,
Noah from KPAI