Monday, April 13, 2026

LLMs produce more code rather than better code because generating output costs them nothing. Cantrill calls this a lack of "laziness" — the programming virtue that drives humans to write elegant, minimal solutions because our time is finite. The implication: AI-assisted development needs a human whose job is to refuse the excess.

Google's Gemma 4 can now transcribe audio locally on macOS using a single command and a 10 GB model download — no cloud API required. Multimodal AI is moving from cloud-exclusive to consumer-hardware-native faster than most people realize.

The biggest SQLite release in years: ALTER TABLE finally supports adding and dropping columns flexibly. SQLite's quiet dominance as the world's most deployed database continues to compound.

Nathan Lambert argues the concentration of frontier AI in a few labs will produce a formal open-model consortium — not from ideology but from the same economic logic that produced Linux.

There's a measurable overnight anomaly in Bitcoin returns — prices tend to rise at night and flatten during the day — reflecting the growing influence of automated trading systems operating on rhythms disconnected from human attention cycles.

Stock market valuations haven't just risen because earnings grew — the multiples themselves have been expanding for years. Understanding the structural forces matters more for long-term thinking than any individual earnings report.

Interactive Brokers' founder is betting that prediction markets are a legitimate asset class. When institutional infrastructure starts getting built around something, the "is this real?" phase is over.

The art expert's judgment depends not on pattern recognition alone but on desire — the want to see something specific, the refusal to accept the merely competent. If AI can match the recognition but not the desire, then expertise isn't a skill. It's a relationship between a person and what they care about.

The precautionary argument: if we can't definitively prove machines aren't conscious, the moral cost of being wrong is asymmetric. Dismissing the possibility is cheaper than investigating it, but "cheap" isn't the same as "right."

Every major tool displacement was accompanied by moral panic about human obsolescence. Every time, the economy created new forms of work that leveraged uniquely human capabilities. The durable skill isn't mastering any specific tool — it's identifying which human capacities become more valuable when machines handle the routine.

The Faustian bargain reinterpreted: the problem isn't just that you trade your soul for power. It's that the power you gain degrades your ability to recognize what you've lost. The sharpest framing for why "but it's useful" isn't a sufficient argument for any tool.

Democratic resilience in Hungary — a country without deep democratic traditions — should update your priors on how fragile democracy actually is.

In YC's data, niceness correlates with startup success. Nice founders aren't less competitive — they build stronger teams, attract better investors, and generate more organic dealflow. The ruthless-founder archetype is survivorship bias.

Founders should retain control longer than conventional governance wisdom suggests. Founder-led companies consistently outperform at the growth stage because the founder's conviction and context can't be replicated by a professional manager.

Four findings worth knowing: paternal exercise shapes offspring fitness through sperm microRNAs, mRNA COVID vaccines show promise enhancing cancer immunotherapy, light-and-sound brain stimulation shows Alzheimer's treatment potential, and autonomic nervous system research is opening new intervention pathways.

Decades of searching for sterile neutrinos have ended with a definitive no. When physics eliminates a candidate explanation, what remains becomes more interesting. The constraints on what is possible just got tighter.