Keep pulling the thread on Sebastian Raschka & Nathan Lambert.
OpenAI's O1 model was famous for introducing inference-time scaling, where a model generates hidden thoughts or intermediate steps before providing a final answer.
Google announced the launch of Gemini Diffusion, which it claims can generate text much faster than its Nano 2 model while achieving the same quality on most benchmarks.
The headline feature of GPT-5 is a router that directs user queries to different models, which is expected to save OpenAI significant money on GPU costs.
In July 2025, Meta publicly stated it was re-evaluating its relationship with open source AI.
OpenAI's research division has consistently produced definitional AI concepts and products, including Sora and O1 Thinking Models.
Recent large-scale deals in the AI sector, such as Grok's reported $20 billion deal and Scale AI's nearly $30 billion valuation, are often structured as licensing agreements rather than full acquisitions.
Anthropic lost a court case and was ordered to pay $1.5 billion to authors for copyright infringement related to its training data.
It is reported that XAI is expected to reach a one-gigawatt compute scale in early 2026 and a full two-gigawatt scale by the end of that year.
The National Science Foundation awarded the Allen Institute for AI (AI2) a $100 million grant over four years, the largest computer science grant the NSF has ever given, to support its work on open AI models.
Google's ability to develop its entire hardware and software stack, from data centers to TPUs, gives it a historical advantage over competitors who must pay high margins for NVIDIA GPUs.
The court found Anthropic culpable in a copyright lawsuit because the company used torrented books for training its AI models.
Cursor's "Composer" model, a fine-tuned version of a large Chinese mixture-of-experts model, updates its weights every 90 minutes based on real-world user feedback.