[AINews] OpenAI GPT-next disproves 80 year old Erdős planar unit distance problem for under $1000
Summary
OpenAI's GPT-next, speculated to be GPT 5.6, disproved the 80-year-old Erdős planar unit distance problem for under \$1000, running for less than 32 hours. This general-purpose LLM produced 125 pages of output, marking a significant milestone in extended reasoning beyond domain-specific math systems. Concurrently, Cohere released Command A+ as Apache 2.0 open weights, a 218B MoE / 25B active multimodal model supporting 48 languages, optimized for low hardware requirements like 2x H100s at W4A4. Benchmarks place Command A+ at 37 on Artificial Analysis's Intelligence Index, comparable to Claude 4.5 Haiku, noted for strong non-hallucination but weaker scientific reasoning. Other updates include new agent benchmarks like InferenceBench and MINTEval revealing struggles with system-level engineering and long-context memory, Google's Gemini 3.5 Flash and Omni rollouts, and various advancements in agent infrastructure, retrieval, and developer tooling. Alibaba's Qwen3.7 Preview also showed strong benchmark performance, with anticipation for open-weight 27B/35B models.
Key takeaway
For AI Scientists and Machine Learning Engineers evaluating frontier model capabilities, you should recognize that general-purpose LLMs are now capable of solving complex, long-standing mathematical problems, suggesting a shift in how scientific discovery can be accelerated. Prioritize exploring models like Cohere Command A+ for enterprise deployments, given its Apache 2.0 license and hardware efficiency. Additionally, focus your agent development efforts on robust infrastructure and dedicated memory subsystems, as these are critical failure points for current agentic systems.
Key insights
General-purpose LLMs are achieving significant scientific breakthroughs and driving architectural innovation.
Principles
- Inference-time scaling drives frontier reasoning progress.
- Memory systems need dedicated learned subsystems.
- Agents often fail on infrastructure reality first.
Method
Context compression systems can cut tokens by up to 70% while improving answer quality, using query-aware and citation-preserving techniques.
In practice
- Consider Apache 2.0 models like Cohere Command A+ for enterprise.
- Evaluate agent performance on system-level engineering tasks.
- Utilize MMR reranking for diverse vector retrieval in RAG.
Topics
- Large Language Models
- Mathematical Reasoning
- Open-Weight Models
- AI Agents
- Inference Optimization
- Retrieval-Augmented Generation
Best for: Research Scientist, CTO, VP of Engineering/Data, AI Scientist, Machine Learning Engineer, Director of AI/ML
Related on AIssential
Editorial summary, takeaway, and curation by AIssential. Original article published by Latent.Space - Www.latent.space.