The Sequence Radar #807: Last Week in AI: From Mega-Rounds to Mathematical Breakthrough
Summary
The AI industry recently experienced significant advancements, marked by substantial financial investments and new model releases, particularly emphasizing agentic systems and enhanced reasoning capabilities. Anthropic secured a record-breaking $30 billion Series G funding round, valuing the company at $380 billion post-money, driven by the success of Claude Code, which boasts a $2.5 billion annualized revenue run rate. OpenAI introduced GPT-5.3-Codex and its ultra-fast variant, GPT-5.3-Codex-Spark, achieving over 1,000 tokens per second inference speeds on Cerebras WSE-3 hardware. Zhipu AI launched GLM-5, a 744B-parameter Mixture-of-Experts model trained on Huawei Ascend chips, achieving 77.8% on SWE-bench Verified. MiniMax M2.5, a 230B-parameter MoE model, also achieved a state-of-the-art 80.2% on SWE-Bench Verified. Google DeepMind unveiled Aletheia, a mathematical research agent powered by Gemini Deep Think, which achieved 95.1% accuracy on IMO-Proof Bench and solved four open Erdős Conjectures.
Key takeaway
For AI Architects evaluating next-generation model deployments, consider the shift towards agentic systems and specialized hardware. Your strategy should prioritize models like GLM-5 or MiniMax M2.5 for their high efficiency and performance in complex tasks, especially if seeking open-weight solutions or optimizing for specific hardware like Huawei Ascend chips. This trend suggests that integrating self-correcting, agentic AI into your workflows will be crucial for tackling long-horizon problems and achieving publication-grade research.
Key insights
AI is transitioning towards agentic systems capable of complex, long-horizon tasks, supported by significant investment and specialized model releases.
Principles
- Agentic systems enhance AI's ability to tackle complex problems.
- Inference-time scaling unlocks scientific problem-solving.
- Self-correction improves AI's reliability and research capabilities.
Method
DeepMind's Aletheia uses a "Generator-Verifier-Reviser" agentic harness to recognize and correct its own hallucinations through internal natural language verification, enabling professional-grade mathematical research.
In practice
- Deploy high-efficiency MoE models for agentic workloads.
- Utilize AI for debugging and automating internal code reviews.
- Explore LLM-as-a-judge protocols for evaluating generated content.
Topics
- AI Agentic Systems
- Large Language Models
- AI Funding
- Mathematical AI
- AI Infrastructure
Best for: AI Architect, Machine Learning Engineer, NLP Engineer, AI Engineer, AI Researcher, Director of AI/ML
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by TheSequence.