one of the best things AI has done for me is give me the ability to quickly visualise any idea that comes to my mind. I was preparing breakfast today and while I was cooking, I could just ask @ChatGPT to visualise what a better solution could look like. As a former designer, this is too much fun!

All deep research apps should have a toggle for “language complexity” where we can choose all content to be written for a certain school grade, how GPT had in its Canvas Mode. Because all Deep Research uses such tough vocabulary by default. One small feature can make a huge difference. @GoogleDeepMind

1/3 Just watched @AnshMehraaa yt video on “How to Save Real Intelligence in 2027?” Episode 5. The main idea: AI can create the output. But your internal benchmark decides whether it is actually good. Taste + judgement still matter.

@AnshMehraaa Love the direction you’re taking with these videos, bro.⚡✌️ How to Save Real Intelligence in 2027? youtu.be/ynrJC7X7clM?si=zDYC… via @YouTube

SpaceXAI dropped Grok 4.7, and it's a big upgrade over 4.6 for coding and agentic work. According to Musk, it has 2.1T parameters, about 40% larger than Grok 4.6, while keeping the same $2/$6 pricing. But the bigger change is how it works. Grok 4.7 is trained to spend longer on difficult tasks, check its own work, and handle long-running coding and agentic workflows. It was also reportedly trained on supplemental SpaceX engineering data, including Starlink satellite telemetry, manufacturing logs and failure data. And you can see the jump in SpaceXAI's own benchmarks. Terminal-Bench went from 20.3% with Grok 4.6 to 38% with 4.7, while EEBench jumped from 53% to 64%. It also handles its 500K-token context window more reliably. Grok 4.7 is available in Cursor, Grok Build and the Grok API. And Musk has already teased Grok 4.8, a 2.5T-parameter model trained with a new C++ software stack.

Odyssey-3 just launched, and this might be one of the biggest steps yet toward general-purpose robotics. It’s a world model that can: → Drive a car in India with just 20 hours of simulated driving data → Control robot arms with only tens of hours of demonstrations → Power humanoids and adapt to changes like different lighting → Pilot drones and navigate around obstacles → Play GTA V and transfer learned skills to other games But the really interesting part is that it doesn’t need thousands of examples for every new task. It already has a learned understanding of physics, movement and cause-and-effect, so a few hours of new experience can be enough to teach it how to apply that knowledge to a new task. It can even generate simulated worlds where other AI agents can practise, fail and learn. Basically, AI learning inside AI-generated worlds. And this could be huge. If world models can scale, they could make robots far more adaptable and accelerate physical AI across industries.