AI
- ‘Skills Explained: How Skills Compares to Prompts, Projects, McP, and Subagents’: Claude Skills are portable folders of instructions/scripts that auto-load via progressive disclosure to deliver repeatable expertise. Use prompts for one-offs, Projects for persistent shared context, MCP to connect tools/data, and subagents for independent, tool-scoped task execution. Prefer Skills when knowledge is reused across chats or agents; combine them so subagents leverage Skills, with MCP for access and prompts for situational guidance.
- ‘You Should Write an Agent’: Argues you should build your own LLM agent: a simple looped model with tools and saved context. Multi-turn conversation is an illusion over a stateless model; tool calls are easy, sub-agents are just separate contexts. Skip MCP/plugins; use APIs and write your own. Security and context engineering matter: segment tools, manage tokens, ground truth. Open problems: balance randomness vs structure, reliable orchestration and interchange formats, and controlling cost.
- ‘Nano Banana Can Be Prompt Engineered for Extremely Nuanced AI Image Generation’: Max Woolf reviews Google’s “Nano Banana” (Gemini 2.5 Flash Image), an autoregressive T2I model with a strong long-context text encoder, excellent prompt adherence, nuanced editing, text-in-image, and solid subject consistency—often without LoRAs. He built a gemimg Python wrapper and uses the API to avoid UI quirks/watermarks. Weak spot: style transfer. Moderation and IP filters are lenient. He shares prompts/notebooks to prove reproducible, high-quality results.
- ‘Meta’s Top AI Scientist Is Quitting as Zuckerberg’s Spending Spree Sputters’: Meta faces a shake-up: chief AI scientist Yann LeCun plans to quit for a startup. The Turing-winning AI pioneer, a skeptic of LLMs, pushed ‘world models’ while Mark Zuckerberg pivoted from open research to commercial LLM ‘superintelligence’ led by Alexandr Wang after a big Scale AI bet and Llama 4’s letdown. Despite huge offers, investors balked; shares fell 11% and nearly 3% on the news.
- ‘Giving Your AI a Job Interview’: Mollick says standard AI benchmarks are leaky, noisy, and miss what matters, though they show rising ability. Instead, interview models: beyond playful vibe tests, run realistic, expert-scored tasks (e.g., GDPval) to reveal strengths, weaknesses, and judgment styles. Models differ by domain and risk appetite; at scale these gaps compound. Orgs should test on their own use cases, repeat, and compare models.
- ‘Introducing Kimi K2 Thinking’: Moonshot introduces Kimi K2 Thinking, an open-source thinking agent that scales test-time thinking tokens and tool steps. It executes 200–300 sequential tool calls, solving long-horizon tasks (e.g., PhD-level math). Scores: HLE 44.9%, BrowseComp 60.2% (vs 29.2% human), SWE-Bench Verified 71.3%. Strong in reasoning, agentic search, coding, and writing with empathetic tone. Live on kimi.com and via API; full agentic mode coming.
Technology
- ‘We Built a Vector Search Engine That Lets You Choose Precision at Query Time’: ClickHouse introduces QBit, a bit-plane transposed vector type that lets you pick precision at query time by reading only high-order bits. It speeds vector search without duplicating data, avoids HNSW’s RAM-heavy indexes, yet stays O(n). SIMD untranspose and selective casting optimize compute. Benchmarks on 29M comments show nearly 2x faster searches with strong recall, enabling a tunable speed-accuracy tradeoff alongside HNSW/quantization.
Economics
- ‘Amazon Upheaval: With Morale Shaken, Jassy Looks for Next Big Play After Mass Layoffs’: After layoffs of about 14,000 with more to come, Amazon CEO Andy Jassy is flattening layers, enforcing office returns, and pushing a startup-like culture to speed decisions, saying cuts target bureaucracy, not finances. Morale is strained as work remains while staff shrink. AWS faces Microsoft/Google pressure; an OpenAI cloud deal and rising AI spend aim to catch up. Staff are urged and tracked on AI use, but some doubt gains. Jassy hunts the next growth pillar beyond retail, cloud, and Prime.
- ‘EU Rolls Out $1.1 Billion Plan to Ramp Up AI in Key Industries Amid Sovereignty Drive’: The EU launched a 1 billion euros Apply AI plan to accelerate AI in key industries and reduce reliance on US and Chinese tech. It eases startup compliance and targets healthcare, pharma, energy, mobility, manufacturing, construction, agri-food, defence, communications and culture. Measures include AI screening centers and agentic AI. Funded via Horizon Europe and Digital Europe, with possible matching funds from countries and private sector.
Management
- ‘“Good Engineering Management” Is a Fad’: The essay argues that “good” engineering management is a fad driven by business cycles, not morality: 2010s hypergrowth prized hiring/retention, while post-2022 favors hands-on execution amid AI optimism and tighter capital. To stay effective, build durable skills—execution, team, ownership, alignment, taste, clarity, navigating ambiguity, and timescales—use self-assessment questions, and make deliberate, energizing tradeoffs across pace, people, prestige, profit, and learning.
- ‘How to Get Feedback’: Deb Liu says that as you get more senior, people stop telling you the truth due to power distance. Make it safe with specific, low-risk prompts and rituals like magic wand dinners. Treat feedback as collaboration, act on it visibly, and see it as a mirror, not a weapon. Leaders must shift from craft to culture, expand their lens beyond function, and proactively seek the voices that make them better.