Today's AI news was a day that highlighted how the 'reach' of models keeps expanding while institutions and legal standards struggle to keep pace.
OpenAI Announces '10 Advances' in Mathematics and Theoretical Computer Science
OpenAI announced new results on long-standing open problems in mathematics and theoretical computer science, citing progress in areas such as geometry, cryptography, and complexity theory. The official announcement provides only limited information, and given the questions around the significance of each result and the room for verification, it's best to avoid definitive assessments. Still, it stands as a symbolic moment showing how the scope of frontier models is widening from text and code toward 'deep mathematical reasoning.'
OpenAI Reportedly Developing 'Astra,' a Family of Long-Horizon Reasoning Models
OpenAI is reportedly developing a new family of models called 'Astra,' in which multiple agents work together to tackle complex problems on the scale of hours to days. According to reports, CEO Sam Altman has already demoed it to policymakers in Washington, and the company has yet to decide whether to release it as GPT-6 or as a new GPT-5 variant. Long-horizon autonomous reasoning and agent coordination is an area that every major lab now positions as its next competitive axis. That said, this is report-based information rather than an official announcement, so the specs and release timing should be treated as undetermined.
EU Begins Mandatory Labeling of Realistic AI-Generated Content on August 2
Starting August 2, the EU will require labels on AI-generated content that is hard to distinguish from the real thing. It's a response to the societal impact of deepfakes and generated images and videos, and platform operators will also be required to comply. As generative AI spreads, how to safeguard 'transparency' has become the centerpiece of regulatory debates across major regions, and the EU's moves are likely to serve as a reference model for other regions.
Perplexity's Motion to Dismiss Rejected in Reddit Data Collection Lawsuit
Perplexity AI had filed a motion to dismiss the lawsuit Reddit brought over data collection (scraping), but the motion was denied, and the case will now proceed. For AI companies, how to justify data acquisition for training and answer generation remains a ongoing point of contention, and they continue to face scrutiny over legal standards from rights holders. This is only one step in the proceedings, but it has once again thrown the data procurement risks facing AI developers into sharp relief.
Bonus: Running Kimi K3 with Just 29GB of Memory
In the local LLM community, a method for running the giant Kimi K3 model with a mere 29GB of memory has been making waves. Called the 'Weight-Aware Streaming Tensor Engine,' the approach loads weights sequentially as needed, allowing massive models to run within a tight memory footprint. At roughly 0.50 tokens per second, it's far from practical, but it points in an interesting direction for working with ultra-large models on consumer hardware.