AI Updates on 2025-11-30

AI data centers are consuming massive amounts of RAM, with hyperscalers buying enormous quantities of server DDR5, HBM, and LPDDR for AI clusters. Analysts expect server DRAM prices to roughly double over 2025-2026 due to AI demand, with this demand at the top of the supply chain reducing availability for consumer RAM @AskPerplexity
AI agents are enabling companies to scale work that was previously too expensive or impractical, making scarce services near infinitely available. Examples include continuous code review, system audits, contract analysis, and automated issue response - work that small companies lack resources for and even large enterprises can only do partially @levie
Startup founders report that fundraising challenges have changed significantly, with one founder feeling demoralized about fundraising despite having already raised a million dollars, suggesting shifting expectations in the funding landscape @paulg
A useful heuristic for identifying AI opportunities: if work in a domain already appears as slop when done by humans, current AI can likely do it well enough to be competitive @paulg

Economists need to think through how to help reduce the labor impacts of AI. While historically new technologies lead to more jobs, living through the transition can be difficult, and this time may be different, requiring more work on mitigation strategies @emollick
GenAI represents the fastest adoption of an economically consequential technology in human history. The speed and breadth of adoption, combined with continued exponential improvement, means society has no clear handle on what it all means @emollick
When powerful AI is in the hands of a billion people, many things happen simultaneously. Few aspects of society are not seeing early impact from AI, unlike previous transformational technologies which had much slower adoption and required expensive complementary assets @emollick

Developer successfully used Perplexity Comet for API endpoint testing using Postman after development, with the AI generating multiple test payloads and a final report after the full run @ai_for_success
A System Design Visualizer was built using Google Antigravity IDE that transforms static architecture diagrams into interactive visualizations. The tool uploads system design images, converts them to Mermaid diagrams using AI, and creates interactive graphs where users can click components for details @mehdiyarix
LLMs are making coding on mobile phones more feasible, enabling developers to build small but useful pieces of software entirely on their phones without needing full attention, compatible with being out in the world @simonw
A startup is using AI to compress information to fit into LLM context windows, discovering that compression is understanding - in the compressed form, the information can be used for other new purposes @paulg
OpenAI released ChatGPT for Teachers resource @gdb

Papers testing AI capabilities should test the strongest case as well as defaults. While it's acceptable to report that Llama 2 failed, researchers should also attempt serious tests using advanced models like GPT-5.1 Thinking in agentic harnesses to better map the frontier @emollick
The null hypothesis is that AI fails at tasks. To falsify this, researchers need the strongest attempts to prove AI success, not the weakest. This requires severe tests in the Popper/Mayo sense - best attempts to make AI work, rather than bad prompts or models that defend the null @emollick
All major AI models struggle with a specific creative task: creating an updated version of "The Fighting Temeraire" with the same style but different subject. While models understand retiring technology, they miss the symbolism of what is being retired and how, failing to capture the contrast between old and modern forms with nostalgia for the original @emollick
Three years of the Lem Test tracked progress from ChatGPT-3.5's release to Claude Opus 4.5 last week @emollick