💸 The $70M Gamble, 😫 AI Burnout is Real, ⚔️ The Great Coding Split
Plus: Svedka's AI Super Bowl ad, Waymo 'hallucinates' traffic, and the end of abstract math.
🎵 Podcast
Don’t feel like reading? Listen to it instead.
📰 Latest News
This week’s image aesthetic (Flux 2 Pro): Romantic Hudson River School
OpenAI and Anthropic Just Split the Engineering World in Half
The battle for coding dominance escalated sharply when the industry’s two leading labs launched their flagship engineering models within minutes of each other. OpenAI released GPT-5.3-Codex, a “hands-on” agentic model designed to operate the computer itself—navigating terminals and GUIs to debug, test, and deploy software with a reported 75.1% score on the Terminal-Bench 2.0 benchmark. Simultaneously, Anthropic countered with Claude Opus 4.6, debuting a massive one-million-token context window and a new “agent teams” feature that enables swarms of AI instances to collaborate on complex codebases in parallel. While OpenAI’s tool is currently gated inside GitHub Copilot and a new desktop app, Anthropic has made its API available immediately to developers.
Why it Matters
Engineering teams now face a stark strategic choice between two divergent philosophies of automation. OpenAI is betting on depth of agency, building a “cyber-operator” that can actively drive the operating system to fix vulnerabilities and execute deployments like a human site reliability engineer. Anthropic is betting on reasoning depth. The combination of massive context and adaptive thought has drawn praise from competitors like xAI co-founder Igor Babuschkin. He noted its capabilities in physics suggest a significant moment for scientific discovery is near. The market has split between a tool that acts as a tireless junior developer (OpenAI) and one that functions as a high-level researcher (Anthropic).
The $7 Million Experiment: Svedka Airs the Super Bowl Ad Made Entirely by AI
The real competition at Super Bowl LX wasn’t on the field. The commercial breaks were dominated by a fierce battle between Silicon Valley giants like Google, Amazon, Meta and OpenAI. They spent millions to position tools such as Gemini and ChatGPT as essential household utilities rather than niche novelties. The most significant moment came from Svedka Vodka which aired the first national advertisement produced entirely by generative AI. In contrast Anthropic used its airtime to attack the business models of its rivals by positioning itself as the only privacy-focused and “ad-free” alternative.
Why it Matters
This signals that AI has graduated from a technical curiosity to a mass-market commodity. The massive capital expenditure proves that tech firms are in a desperate land grab for mainstream loyalty before the market calcifies. Svedka’s experiment is perhaps more significant for the industry itself. By handing the creative reins to an algorithm for a national spot the brand has crossed a line for the advertising sector. It suggests a future where high-budget campaigns are generated rather than filmed. This threatens to cut human production teams out of the loop entirely regardless of whether the audience actually likes the result.
The $70M Gamble: Crypto.com CEO Buys ‘AI.com’ from a Malaysian Ponzi Link
The most expensive domain in history just launched during the Super Bowl. AI.com is a new platform from Crypto.com CEO Kris Marszalek. It debuted during Super Bowl LX with a commercial that signalled a massive push into the consumer market. Marszalek reportedly paid a record-breaking $70 million in cryptocurrency for the URL. He is positioning it as a user-friendly “app store” for autonomous agents. The service allows users to generate personal AI agents that can execute tasks across apps without any coding. Examples include trading stocks or managing calendars. The core promise is a “decentralised network” where agents autonomously build new features. They then share those upgrades with the rest of the swarm which theoretically makes every user’s bot smarter over time.
Why it Matters
This is a brute-force attempt to buy the “front door” of the AI internet. Marszalek is betting nearly nine figures on the domain alone. He believes the winner of the agent wars will be decided by branding rather than just technology. The backstory of the domain is as strange as its price tag. The seller was reportedly Arsyan Ismail, a Malaysian tech figure who bought the domain as a 10-year-old in 1993 for $100 simply because “AI” matched his initials. Ismail, who was previously linked to the collapsed BitKingdom Bitcoin ponzi scheme, walked away with a 700,000x return on the sale.
The launch itself proved rocky. Despite burning $10 million on the Super Bowl ad spot, the website crashed instantly with a 504 Gateway Time-out error under the weight of the traffic. Critics have also pointed out that the platform appears to be a thin wrapper for OpenClaw, an open-source agent framework, despite its lofty slogan of “accelerating the arrival of AGI”. It marks a pivot from AI as a developer tool to a mass-market utility. Users simply “hire” an agent to do their digital chores. The high-profile market entry combines massive capital injection with deafening marketing. It creates a competitive push for accessibility in the AI agent space.
The Efficiency Paradox: New Data Proves AI is Making You Work Longer, Not Smarter
The promise that AI would liberate us from the grind has backfired. A new study in the Harvard Business Review tracked 200 tech employees over eight months and found the technology acted as an accelerant for burnout. Instead of using efficiency gains to rest, employees used the tools to expand their roles and take on tasks they previously would have delegated. The research reveals that AI induced a state of hyper-multitasking. Staff voluntarily worked longer hours and filled their breaks with “quick” AI-assisted tasks simply because the tools made it easy to do so.
Why it Matters
This phenomenon is known as “workload creep.” The friction of starting a task has dropped to zero so employees are drowning themselves in busywork. It also introduces a dangerous “competence trap” where non-experts use AI to generate mediocre work in fields they do not fully understand. This forces actual specialists to spend their time cleaning up the mess rather than doing their own jobs. The result is not the leisure utopia we were promised but a recipe for cognitive fatigue, lower quality output and higher staff turnover.
📝 More from Harvard Business Review
Axiom Just Turned Abstract Mathematics into Executable Code
Axiom has effectively automated the role of the mathematician. Its new proprietary system AxiomProver has solved multiple open mathematical problems including Fel’s conjecture without human assistance. The AI works by translating natural language into Lean which is a strict formal programming language for logic. It then generates a proof strategy that is mathematically verified line-by-line by the software itself. The company has released the resulting proofs for human review.
Why it Matters
This validates the “math-as-code” thesis. By converting theoretical problems into executable software the system shifts the burden of verification from humans to machines. This has immediate applications outside of academia. The same formal methods can be used to mathematically prove that software in self-driving cars or medical devices is bug-free. It promises a future where critical infrastructure is not just tested but logically guaranteed to work. It allows scientists to focus on high-level creative connections while the AI handles the rigorous logic.
Waymo is Using AI to ‘Hallucinate’ Traffic Nightmares
Waymo has built a “Matrix” for its robotaxis. The Waymo World Model is a new simulation engine powered by Google DeepMind’s Genie 3 that generates interactive, photorealistic 3D environments from text prompts. Instead of just replaying recorded logs, the system creates entirely new scenarios—complete with synthetic camera and lidar data—allowing engineers to test the software against hazards that don’t exist in the real world. The technology is currently deployed internally to train the company’s autonomous driving stack.
Why it Matters
This solves the “long tail” problem of autonomy. Physical testing is bound by the laws of probability; you cannot schedule a tornado or a sudden deer crossing to test how the car reacts. By generating these edge cases on demand, Waymo moves from driving physical miles to simulating infinite variations of reality. It implies that the final hurdle for self-driving cars isn’t logging more hours on the road, but surviving the millions of synthetic “nightmares” generated in the lab.
Last week’s newsletter:
Great find: "AI is Making You Work Longer, Not Smarter" - I can relate to it when self-proclaimed vibe-coding newbies send you pull requests: Along the lines of "Look 'ma - see what I can do now!"