Amodei Loses His Sh!t; OpenAI's GitHub Snub? Bots v Alzheimer's

Today's AI Outlook: 🌥️

Anthropic And OpenAI Rivalry Goes Nuclear

The simmering rivalry between Anthropic and OpenAI just erupted into the open. A leaked internal memo from Anthropic CEO Dario Amodei blasted OpenAI’s recent Pentagon partnership, calling it “maybe 20% real and 80% safety theater.” The 1,600-word memo, circulated internally before surfacing publicly, reads like a pressure valve finally blowing after years of quiet tension between the companies.

Amodei, who left OpenAI in 2020 before founding Anthropic, did not hold back. The memo accuses Sam Altman and OpenAI leadership of misrepresenting discussions around the Pentagon deal and “gaslighting” competitors about safety concerns. It also references a $25M political donation by OpenAI co-founder Greg Brockman while emphasizing Anthropic’s refusal to engage in what Amodei described as “dictator-style praise” toward government officials.

The drama unfolded against a strange backdrop. The Pentagon reportedly labeled Anthropic a potential “supply chain risk,” and shortly afterward OpenAI finalized its own defense partnership under similar terms. Days later, Amodei struck a softer public tone, saying the company likely shares “much more in common than differences” with the U.S. government.

Why it matters

Frontier AI competition is no longer just about models. It is about government contracts, geopolitical influence, and control over the AI infrastructure stack. Defense partnerships could shape how powerful AI systems are deployed and regulated, making these deals enormously strategic.

The leaked memo also highlights just how personal the rivalry has become. Anthropic and OpenAI share roots, talent, and investors, but they increasingly compete across the same markets: enterprise AI, government contracts, and frontier research.

The Deets

• Amodei called OpenAI’s Pentagon messaging “80% safety theater.”
• The memo accuses Sam Altman of misleading competitors about negotiations.
• Anthropic was reportedly flagged by the Pentagon as a supply chain risk.
• The document also referenced political donations tied to OpenAI leadership.

Key takeaway

The AI industry’s most important rivalry is now playing out in public, and the stakes go far beyond chatbots.

🧩 Jargon Buster - Safety Theater: Policies or messaging designed to appear responsible or safe without significantly changing the underlying risk.

⚡ Power Plays

OpenAI Building Its Own 'GitHub,' Poking the Microsoft Bear

OpenAI is reportedly developing its own internal code repository platform, potentially replacing Microsoft-owned GitHub for its engineering teams. The project reportedly began after repeated GitHub outages disrupted workflows during the platform’s ongoing infrastructure migration to Azure.

According to internal discussions, GitHub leadership warned employees that the full migration could take up to two years, requiring major engineering resources to maintain stability. Frustrated by the delays, OpenAI teams began exploring whether they should simply run their own platform.

The project could eventually expand beyond internal use. Some OpenAI employees have suggested opening the system to outside developers and integrating it with Codex coding agents, effectively turning the platform into an AI-native development environment.

Why it matters

If OpenAI launches a public version, it could put the company in direct competition with Microsoft’s GitHub, despite Microsoft being OpenAI’s largest financial backer.

Developer platforms control massive ecosystems. GitHub alone hosts code for more than 100M developers, making it one of the most strategic assets in the software industry.

The Deets

• The project started after GitHub outages during Azure migration.
• OpenAI engineers began building an internal alternative.
• Some employees propose offering it as a paid external platform.
• Integration with AI coding agents like Codex is under discussion.

Key takeaway

OpenAI is steadily building an entire AI-native developer stack. That strategy could eventually put it head-to-head with its own investors.

🧩 Jargon Buster - Code Repository: A centralized platform where developers store, manage and collaborate on software code.

🎬 Big Picture

ByteDance Video Engine Undercuts Hollywood, Competitors On Pricing

New pricing from ByteDance’s Volcano Engine suggests AI video generation is about to become dramatically cheaper. The company revealed API costs for its Seedance 2.0 video model, showing that generating video can cost around $0.13 per second.

The math behind the number is straightforward. The system charges roughly $6.40 per million tokens, and a typical 15-second video consumes about 300,000 tokens, bringing the total cost close to $2 per clip.

That is significantly cheaper than many existing tools. Platforms like Runway and Pika often reach $0.20 to $0.50 per second, particularly when creators need multiple attempts to get usable results. Compared with Hollywood production, the difference is even starker. A finished minute of traditional film production can cost tens of thousands of dollars when crews, actors, and locations are involved.

Why it matters

Cheap AI video does not replace film studios overnight, but it dramatically lowers the cost of producing visual content. That shift could reshape advertising, social media video, indie filmmaking, and online education.

The falling price of AI generation also signals a broader trend. Creation tools that once required massive resources are rapidly becoming accessible to individuals and small teams.

The Deets

• Seedance 2.0 pricing implies $0.13 per second of generated video.
• A typical 15-second clip costs about $2 to generate.
• Competing tools often cost 2–4x more per second.
• Hollywood production still costs thousands per finished minute.

Key takeaway

The economics of video production are shifting fast. AI is pushing visual creation toward the same cost curve that transformed digital photography and publishing.

🧩 Jargon Buster - Tokens: Units used by AI systems to measure how much text or data is processed during generation.

🧠 Research & Models

Google Pushes AI Pricing Lower With Gemini 3.1 Flash Lite

Google has launched Gemini 3.1 Flash Lite in developer preview, positioning it as the cheapest model in the Gemini lineup. The model is optimized for high-volume workloads, prioritizing speed and cost efficiency rather than advanced reasoning.

The pricing is aggressive: Input tokens cost about $0.25 per million, while output tokens cost around $1.50 per million, pushing AI inference closer to commodity infrastructure pricing.

Early Arena benchmark tests show the model outperforming GPT-5 on some text tasks, while maintaining faster response speeds than many competing fast-tier models.

Why it matters

AI model pricing is collapsing rapidly. As costs drop, the competitive advantage moves away from the models themselves toward distribution platforms, AI agents, and application layers built on top.

The Deets

• Gemini 3.1 Flash Lite targets high-volume API workloads.
• Input costs are roughly $0.25 per million tokens.
• Output tokens cost about $1.50 per million.
• Benchmarks show competitive performance despite the low price.

Key takeaway

The race to cheaper AI is accelerating. Soon the most valuable AI companies may not be those with the smartest models, but those controlling how people use them.

🧩 Jargon Buster - Inference: The process where a trained AI model generates answers or predictions from new data.

🔬 Research & Education

OpenAI Builds Framework To Test If ChatGPT Actually Improves Learning

OpenAI has introduced a new research framework designed to measure whether AI tutoring tools actually improve learning outcomes over time. Developed with researchers from Stanford University and Estonia’s University of Tartu, the system tracks metrics such as knowledge retention, motivation, and persistence rather than just immediate test scores.

In an early trial involving more than 300 microeconomics students, participants using ChatGPT’s study mode scored 15% higher than those studying without it. Results in other subjects were less statistically significant, underscoring the complexity of measuring AI’s educational impact.

Estonia is now running the largest test yet, monitoring 20,000 high school students across the country for a full semester.

Why it matters

AI is rapidly entering classrooms around the world. The big question is whether it improves learning or simply makes completing assignments easier.

Large-scale studies like this could shape how schools adopt AI tutors, how teachers design coursework, and how policymakers regulate educational AI tools.

The Deets

• The system is called the Learning Outcomes Measurement Suite.
• Early trials involved 300+ students.
• Microeconomics scores improved by 15%.
• Estonia is running a national test involving 20K students.

Key takeaway

Education may become one of the most important proving grounds for AI’s real-world value.

🧩 Jargon Buster - Learning Outcomes: Measurable knowledge or skills students gain after completing a course or training program.

🤖 Robotics

Robots Move From Hospitals To Nuclear Plants

Robotics innovation is expanding into dramatically different environments, from medical treatment to nuclear cleanup.

Now Jacksonville startup MMI received FDA clearance to begin human trials of microscopic surgical robots designed to treat Alzheimer’s by clearing lymphatic drainage pathways in the neck. The company, backed by $220M in funding, will test the procedure on 15 patients in hopes of proving safety before seeking broader approval by 2027.

At the opposite extreme, TEPCO unveiled a massive 22-meter snake-like robotic arm built to explore the damaged Fukushima Daiichi nuclear plant. The 4.6-ton robot is designed to navigate narrow radioactive passages while capturing inspection data and supporting debris removal from melted fuel areas.

Meanwhile, robotics is expanding into everyday industries. Samsung announced plans to convert its global factories into AI-powered autonomous manufacturing systems by 2030, using digital twin simulations, AI agents, and eventually humanoid robots. Xiaomi has already begun testing humanoid robots in its car factory, where prototypes successfully completed three hours of assembly work with 90.2% accuracy.

Why it matters

Robotics is moving beyond research labs into real-world infrastructure. Healthcare, energy, agriculture and manufacturing are all beginning to adopt machines that combine AI perception with physical action.

The result is a rapidly growing category of physical AI systems designed to operate where humans cannot or should not.

The Deets

• MMI received FDA clearance for Alzheimer’s treatment trials using micro robots.
• TEPCO deployed a 22-meter robotic arm inside Fukushima’s reactor site.
• Samsung plans AI-driven autonomous factories by 2030.
• Xiaomi humanoid robots achieved 90.2% assembly success rates.

Key takeaway

The next wave of AI will not stay inside software. It will increasingly take physical form in robots operating across industries.

🧩 Jargon Buster - Digital Twin: A virtual simulation of a real-world system used to test changes and optimize performance before deploying them physically.

⚡ Quick Hits

• Nvidia CEO Jensen Huang said the company’s $30B investment in OpenAI will likely be its last before OpenAI eventually goes public.
• Seven major tech companies signed a U.S. pledge agreeing to fund energy upgrades for AI data centers.
• Meta signed a multi-year AI licensing deal with News Corp worth up to $50M annually.
• Chicago banned commercial sidewalk delivery robots, citing accessibility concerns.
• Qualcomm says robotics chips could become a meaningful revenue stream within two years.

🛠️ Tools Of The Day

Adapt – An AI “computer for teams” that connects company data, automates workflows, and lets teams build applications directly from Slack conversations

Glaze – A new Raycast tool that converts AI chats into fully functional local desktop applications.

Enia Code – A proactive coding agent that automatically detects bugs and suggests architectural improvements as developers write code.

Picsart Persona & Storyline – A creator tool that generates consistent AI characters and video narratives, designed for faceless content channels.

MuseMail.ai – Generates fully designed marketing emails from a single prompt, helping teams produce on-brand campaigns quickly.

Today’s Sources: The Rundown AI, AI Secret, Robotics Herald