We are officially past the chatbot phase. @thsottiaux just had his work AI coordinate with his personal AI over email to handle taxes with a human CPA. Took 20 mins
everyone is obsessing over gpt-6 astra and opus 5.5 getting smarter, but chatbots are boring now.
the real shift is AI becoming the actual operating system.
who builds it first? apple, openai or anthropic?
SpaceX is taking another step toward entering the mobile network market.
The company has agreed to acquire licenses for low-band radio spectrum in the 800 MHz range in a deal reportedly worth around $8 billion. The acquisition is intended to strengthen Starlink Mobile and help SpaceX compete with traditional mobile carriers.
Low-band frequencies can travel long distances and penetrate buildings more effectively than higher frequencies, making them useful for expanding mobile coverage.
SpaceX aims to combine satellite connectivity with terrestrial infrastructure to provide a more comprehensive mobile service.
However, challenging established operators such as AT&T, Verizon, and T-Mobile will require substantial investment, infrastructure, and time. The deal is an important step, but building a competitive mobile network remains a major challenge.
I tried building a game based on the new Tesla Roadster with Gemini 4 Argon. It took 11 detailed prompts, each addressing its previous mistakes, to get the best version. By prompt 12, Argon had changed the mechanics, turned the car sideways, and ruined the gameplay. Prompts 13 and 14 initially produced nothing, while prompt 14 eventually broke almost everything, replacing the mechanics we had built. Prompt 15 couldn't fix anything.
First image: the broken version after prompt 12.
Second image: the best working prototype from around prompt 8.
They really just got on stage and said this newly designed 4000+ pound car can do flips and go half the speed of sound (380+mph) and will ship in 2028. This is a prank, right?
Claude Opus 5.5 vs Gemini 4 Argon: build a mobile Lode Runner game.
Claude did it on the first try with a tiny prompt and zero extra explanations.
Argon’s first attempt was terrible. But after detailed instructions, its second version came surprisingly close to Claude’s
Today, I ran 8 experiments on Gemini 4 Argon to test reasoning, accuracy, hallucinations, and self-correction.
Results: 3 error-free, 3 partially successful, and 2 with significant errors. The biggest weakness? Staying logically consistent and catching its own mistakes.
Open weights are where things get interesting. A powerful model is impressive, but making it accessible for developers to inspect, adapt, and build on could matter even more. Now let's see how Muse Spark Open performs outside the benchmark table
Today we’re releasing Muse Spark Open, the most capable US open-weight model to date.
Built for coding and agentic work, it excels at long-horizon tasks, complex software engineering, and multi-step reasoning.
Key Model Specs:
→ 816B total / 42B active parameters
→ 1M token context window
→ Natively multimodal across text, image, video, audio
benchmarks show what a model can do at its best. the real test is whether it can do it reliably when tasks get messy, complex, and unpredictable. what was the biggest gap between Argon's early demos and the model you actually shipped?
A conversation with Sergey Brin, @joshwoodward, @tulseedoshi, and @melvinjohnsonp about Gemini 4 Argon, how we got here, and the path ahead!
70% of calls resolved sounds impressive, but how many customers actually got their problem solved, and how many simply gave up on the AI?
