Erdos #397 marked solved, Somani credits ChatGPT, Lean-verified
Somani says ChatGPT produced a solution; he checked it and formalized it with Harmonic, and the ErdosProblems site lists the result as Lean-verified.
Read full story→Back to archive
Issue #11··20 min read·10 stories
AI memory sold out; Nvidia resets AI factory economics
AI memory prices surge. Also, Qwen's new multimodal model and an open-source pen testing agent.
AI memory sold out yesterday, causing prices to surge and adding new considerations for infrastructure planning. Nvidia also outlined new economics for AI factories, further shaping compute costs. Builders can also check out Qwen's new multimodal model and an open-source tool for AI agent-based penetration testing.
▲NEWS
3 stories2
Nvidia pitches an AI factory stack that shifts TCO and lock-in
At CES 2026, Nvidia CEO Jensen Huang announced new reference architectures for AI factories, tighter integration of its software stack (CUDA, Triton), and a shift towards subscription-based access for some core services. If you are budgeting 2026 capacity, model lock-in costs as software plus networking plus memory, not just GPUs.
3
HBM and server DRAM stay tight, pricing pressure into 2026
High Bandwidth Memory (HBM) demand is crowding capacity, causing server DRAM Average Selling Prices (ASPs) to rise, with spillover price pressure into consumer segments. Analysts forecast a 15-20% increase in DRAM ASPs into 2026, driven by ongoing AI infrastructure buildouts.
⚙TECHNICAL
1 story1
Code-Free Libraries: Specs Drive AI Agent Generation
Drew Breunig proposed "whenwords", a software library where the repo contains only a precise specification (AGENTS.md) and language-independent YAML conformance tests. The implementation is generated per-language by an AI agent, with AGENTS.md constraining the agent's behavior to the spec.
◈ANALYSIS
3 stories1
US vs. China: Two AI Strategies Diverge
An essay argues the US prioritizes developing superior, proprietary AI models, assuming intelligence remains a scarce resource. China, conversely, is commoditizing intelligence via open-weight models, aiming to shift economic value towards coordination, execution, and energy infrastructure.
2
Antirez: LLMs change the default way we program
Salvatore Sanfilippo (antirez) argues that LLMs have changed the default way we program. For glue code, scripts, prototypes, and one-off tooling, LLMs compress timelines. For long-lived systems, you still need specs, tests, and review. The shift is where effort moves.
3
The "10:10 Problem": Why AI Can't Do Clocks
AI models consistently fail at generating and reading clocks, often misplacing numbers and hands. This happens because AI relies on pattern recognition, compounded by the "10:10 problem" from skewed training data, leading to a significant accuracy gap for reading analogue clocks. Builders should not assume visual numeracy from models.
⚒TOOLS
3 stories1
Awesome AI Apps: RAG & Agent Code Examples
The "awesome-ai-apps" GitHub repository collects curated examples and starter projects demonstrating RAG, AI agents, and complex AI workflows. It offers working code examples for modern AI use cases.
2
Strix: Open-Source AI Agents for Pen Testing
Strix released an open-source project offering AI agents built for penetration testing. This Python-based tool automates vulnerability discovery and recon, designed to be deployed in a sandboxed environment with no outbound access by default and strict logging.
3
Qwen2.5-Omni: Multimodal with Real-Time Speech
Alibaba Cloud's Qwen team launched Qwen2.5-Omni, an end-to-end multimodal model. It processes text, audio, vision, and video, and generates real-time speech. The model is available under an open license, with weights provided for local deployment via a Python package, requiring specific GPU configurations.