Social Media Briefing — April 13, 2026

Mollick observes that the current state of AI agent development involves debating which markdown files (skills, memory, tool instructions) and in what order to feed AI for best output, and argues this is likely a temporary phase in agent development.

Responding to yesterday's Reddit revelations about hidden reasoning effort tags, Boris Cherny (Claude Code team) refutes claims that switching Claude Code's default to 'medium' reasoning effort was sneaky. Explains it was based on user feedback about token usage, included in changelog, and shown via opt-out dialog.

Burkov coins 'the AI Fermi Paradox' - if AI is so transformative, where are the multi-billion dollar businesses that should have been started in the 3+ years since LLMs became available? Calls industry claims lies or incompetence.

Mollick evaluates thinking trace UX across major AI providers: ChatGPT has the best display (short summary + detailed sidebar audit), Claude is close but more summarized, and Gemini's thinking trace display is notably weak.

Mollick argues that truly interesting/outlier ideas will become increasingly valuable as AI reduces the cost of executing ideas. Notes research shows AI is good at generating interesting ideas but not at generating exceptional outlier ideas.

Following yesterday's Reddit report on Alibaba's open-source retreat, Nathan Lambert commentary on open model companies starting to worry about money. States that making frontier-level open models with free use licenses is unsustainable for everyone except maybe Nvidia.

Following yesterday's Reddit Gemma 4 benchmarks, Svpino reports running Gemma 4 locally via Ollama. Says it's unusable with Claude Code (can't load/execute skills) but decent as a chatbot. Cross-posting questions between Claude and Gemma 4 shows Gemma's answers are usable. Wishes for better UI harness with projects/memory.

Burkov argues the LLM industry is a bubble, comparing it to big data: big data at least had the famous Target pregnancy prediction story as a success case, but three years into the 'LLM bubble' there's no equivalent success story.

Anthropic's trq212 announces there's now a TurboTax connector in Claude Code, joking about procrastinating on taxes.

vLLM announces day-0 support for MiniMax M2.7, highlighting its agentic-first design with multi-agent orchestration, strong coding capabilities, and office automation features.