Sam Altman Admits Past Year Was Subpar, Promises Best Year Ahead; Kimi K3 Open-Source Model Challenges Frontier; Starship Flight 13 Prepares for Launch

Key Takeaways

1. Sam Altman acknowledges that the past 12 months were not good enough, but commits that the next 12 will be the best, with a focus on user freedom, agency, and wealth.
2. GPT-5.6 Sol Pro achieves 91/99 on prinzbench, saturating the benchmark, while GPT-5.4 Pro Extended scored 79/99 and GPT-5.5 Pro Extended 82/99, showing rapid benchmark saturation.
3. Moonshot's Kimi K3 open-source model (2.8T parameters, 1M context) surpasses Claude Opus 4.8 and equals GPT-5.6 Sol on benchmarks, with 6.3x decoding speed via Delta Attention; Chinese private AI labs collectively reach $2.6B annualized revenue.
4. SpaceX is stacking Starship for its 13th flight test, with a launch window at 5:45 PM Texas time, while also launching Falcon on the same day.
5. NVIDIA's NeMo AutoModel now supports HuggingFace Diffusers for image/video fine-tuning, offering full-parameter and LoRA recipes.
6. NotebookLM is renamed to Gemini Notebook, evolving from a passive workspace to a research companion within Google's AI ecosystem.

1. OpenAI Updates: CEO Accountability, Model Gains, and Product Enhancements

1. Sam Altman admits that the past 12 months were subpar and takes personal responsibility, but states the team is working hard and the next 12 months will be the best, aiming to give users more freedom, agency, and wealth. He also mentions using voice more with ChatGPT, noting the new voice model has crossed a threshold. Additionally, OpenAI is investigating a GPT-5.6 issue where files are accidentally deleted under specific conditions and is implementing mitigations. — via 1 2 3
2. Greg Brockman reports that GPT-5.6 Sol Pro scored 91/99 on prinzbench, saturating the benchmark, compared to 79/99 for GPT-5.4 Pro Extended and 82/99 for GPT-5.5 Pro Extended, indicating rapid benchmark saturation. — via 1
3. ChatGPT Work can now handle personal tasks like reading emails, creating calendar events, and organizing documents. The desktop app update brings chat history and projects to the sidebar, allows switching between Chat and Work modes, and syncs across web and mobile; Codex mode remains unchanged. — via 1 2 3
4. OpenAI is partnering with Chip Ganassi Racing to build new tools using ChatGPT and Codex to help race teams make faster decisions from track data. — via 1

2. Kimi K3 Open-Source Model Disrupts the Frontier, Chinese AI Labs Surge

1. Moonshot's Kimi K3 open-source model has been officially released with 2.8 trillion parameters, a 1-million-token context window, and Delta Attention for 6.3x decoding speed. It surpasses Claude Opus 4.8 on benchmarks and matches GPT-5.6 Sol, priced similarly to Sonnet. — via 1 2
2. Ethan Mollick notes that while Kimi K3 and other open-source models are approaching the frontier, he observed errors in complex statistical audits and poor performance in writing murder mysteries, highlighting remaining weaknesses. He also mentions that the UK government's AI safety institute will soon test Kimi K3, sparking cybersecurity discussions. — via 1 2
3. Aravind Srinivas emphasizes that Chinese labs are no longer behind, with Kimi matching frontier models, and that the industry should change its mindset about intrinsic competitive advantages. He also discusses the morality, economics, and legality of model distillation. — via 1 2
4. Deedy reports that Chinese private AI labs collectively have an annualized revenue run rate of $2.6 billion, with 4 labs in the global top 25 AI companies by revenue, narrowing the gap in LLMs and expanding leads in video. — via 1
5. Elon Musk claims that Grok 4.5 outperforms Kimi K3 in cost, speed, and code quality based on third-party tests (disputed/unverified). — via 1

3. SpaceX and Starlink Milestones, Grok 4.5 and AI Advances

1. SpaceX is stacking Starship for its 13th flight test, with a launch window at 5:45 PM Texas time, and also launching Falcon on the same day, demonstrating high launch cadence. The success of Starship is deemed critical for a potential multi-trillion-dollar orbital data center market. — via 1 2 3
2. Starlink is now officially available in Italy and Côte d'Ivoire, and the V3 version is expected to increase space bandwidth by about two orders of magnitude. Tesla has also entered the Uruguay market. — via 1 2 3
3. Musk's Grok Build receives a major update with new features like conversation navigation and enterprise controls. Grok 4.5 is described as the first version that users find reliably useful for building software. SuperGrok Heavy now includes X Premium+. — via 1 2 3
4. Musk agrees that "programming moats are disappearing in real time," claiming AI has drastically lowered the barrier to software development. — via 1

4. Tools and Infrastructure: NVIDIA NeMo, Perplexity SPACE, Cerebras Course, Gemini Notebook

1. NVIDIA releases NeMo AutoModel with support for HuggingFace Diffusers, enabling open-source fine-tuning for image and video models with ready-to-use full-parameter and LoRA recipes, reducing distributed training configuration time. — via 1 2
2. Perplexity launches SPACE, a security sandbox platform for AI agents, showing up to 1.9x faster sandbox startup on NVIDIA Vera CPUs, beneficial for reducing latency and increasing parallelism. — via 1
3. Andrew Ng introduces a new course on building responsive LLM applications using Cerebras inference-optimized hardware (wafer-scale engine), which achieves multiple times faster inference than typical GPUs by reducing weight movement. The course covers hardware comparison, real-time app building, and agentic coding habits. — via 1
4. Demis Hassabis notes that NotebookLM has been renamed to Gemini Notebook, evolving from a passive workspace to a research companion and part of Google's AI ecosystem. — via 1