AI Breakthroughs: Math, Protein Folding & Efficiency
Key Questions
What is SubQ and its efficiency gains?
SubQ achieves sub-quadratic sparse attention supporting 12M token context with up to 1000x compute savings.
How did Confluence Labs perform on ARC-AGI-2?
Confluence Labs cracked ARC-AGI-2 with a 97.9% score, advancing progress on challenging AI benchmarks.
What did GPT-5.6 Sol Pro accomplish in optimization?
GPT-5.6 Sol Pro closed a 30-year gap in convex optimization using a detailed 10-page prompt and deep domain expertise.
How did K3 demonstrate self-improvement?
K3 autonomously optimized GPU kernels in a 15-hour run, halving compute time and building a compiler from scratch.
What is VideoChat3?
VideoChat3 is a fully open video MLLM with 4B parameters that outperforms larger models in video understanding tasks.
What valuation did Elorian achieve?
Elorian raised $55M at a $300M pre-seed valuation focused on visual AI.
How does DeepSeek V4 Pro compare to GPT-5.5 Pro?
DeepSeek V4 Pro beats GPT-5.5 Pro at 107x lower cost, highlighting efficiency gains in new models.
What concerns exist around synthetic data from T2I models?
New research shows modern text-to-image models suffer from aesthetic collapse that reduces diversity in synthetic training data.
SubQ achieves sub-quadratic sparse attention with 12M token context and 1000x compute savings. New image generator using coupled oscillators (UnconvAI) emerges. Mustafa Suleyman's MAI-image-2.5 ranks #2 for text-to-image. DeepSeek V4 Pro beats GPT-5.5 Pro at 107x cost. Confluence Labs cracks ARC-AGI-2 (97.9%). GLM-5.2 open weights top open-weight leaderboard. Recursive self-improving agents launched. Elorian raises $55M at $300M pre-seed for visual AI. GPT-5.6 Sol Pro used to resolve open statistics question on FDR control, demonstrating AI's role in advancing scientific methodology. VideoChat3: fully open video MLLM with 4B params outperforms larger models. New research challenges synthetic data quality from modern T2I models, showing aesthetic collapse reduces diversity. AI science summit showcases $19M for AI-driven genomics, quantum materials, geothermal. K3 autonomously optimized GPU kernels in a 15-hour run, halving compute time and building a compiler from scratch, signaling AI self-improvement in hardware optimization. GPT-5.6 used to close a 30-year gap in convex optimization, requiring a 10-page prompt and deep domain expertise. S1-Omni unified multimodal reasoning model outperforms GPT-5.5 and Gemini on scientific benchmarks. DSWorld data science world model accelerates RL training 14x by predicting state transitions before execution.