Grok 4.5 Claims Political Neutrality, GPT-5.6 Sol Ultra Proves Math Conjecture, and Starship Flight 13 to Deploy V3 Starlink

Key Takeaways

1. Grok 4.5 is rated as the most politically neutral AI model with a bias score of -0.02, and it also scores highest on the WANDR benchmark.
2. OpenAI's GPT-5.6 Sol Ultra has proven a 50-year-old mathematical conjecture and is now publicly available, marking a significant milestone in AI reasoning.
3. The post-frontier AI model competition is shifting to systems competition, with local models becoming the default interface for most tasks, supported by routing, open-weight models, and local compute.
4. Startups selling training data and RL environments to large AI labs generated approximately $8.5 billion in revenue in July 2026, with the top four companies capturing over 75% of the market.
5. SpaceX's Starship Flight 13, featuring numerous upgrades, aims to launch and deploy V3 Starlink satellites for the first time, along with heat shield testing.
6. AI has net created jobs so far, contrary to expectations, and this trend may continue.

1. AI Models and Benchmarks

1. Elon Musk cited independent research showing Grok 4.5 has a political bias score of -0.02, making it the most neutral model, and praised it as a "seeker of truth." — via 1 2
2. Aravind Srinivas noted that Grok 4.5 achieved the highest score on the WANDR evaluation, and that Perplexity Computer is priced similarly to the recently released custom GLM-5.2 with Advisor. — via 1
3. Sam Altman highlighted that GPT-5.6 Sol Ultra has proven a 50-year-old mathematical conjecture and is now publicly available. — via 1
4. Deedy compared GPT-5.6 Sol Ultra and Fable XHigh by asking them to create an app comparing fastest-growing stocks over different time periods; Sol Ultra failed to produce subsequent comparison videos while Fable XHigh succeeded. — via 1
5. Aravind Srinivas argued that post-frontier model competition is becoming systems competition, with key elements including routing, open-weight models, local compute, enterprise data, and the sustainability of big lab pricing power. — via 1
6. He also noted that under the same model architecture, NVIDIA outperforms AMD, and B300 is faster than B200. — via 1

2. AI Industry and Economy

1. Deedy reported that in July 2026, over 50 startups sold training data and reinforcement learning environments to large AI labs, generating total revenue of approximately $8.5 billion and an estimated valuation of $100 billion. The top four companies—Scale, Surge, Mercor, and Handshake—captured over 75% of the market. — via 1
2. Sam Altman observed that AI has net created jobs so far, which was contrary to his expectations, and suggested this trend might continue. — via 1
3. Aravind Srinivas discussed the difficulties of scaling Agentic AI: it requires building new systems (orchestration, routing, context engineering, state management, multi-agent collaboration, safety/compliance) and enterprises face new thinking and governance culture challenges. The problem is not on the demand side, but financial aspects may be problematic. — via 1
4. He also shared that a user built a personalized podcast player using Grok 4.5 and Perplexity Computer, streaming across podcasts by topic at a cost of just $11. — via 1

3. Hardware and Infrastructure

1. Elon Musk announced that SpaceX's Starship Flight 13 is targeting a Thursday launch with numerous upgrades, including the first deployment of V3 Starlink satellites and heat shield testing. — via 1 2
2. Aravind Srinivas argued that humans are good at using frontier models as tools, and local models will become the default entry point for most tasks, controlling and utilizing frontier models efficiently with minimum power and cost. — via 1
3. NVIDIA AI and swyx promoted the Local AI Summit panel "State of the Union: Why Local, Why Now," featuring leaders from NVIDIA, Roboflow, and others. — via 1 2