Brazil's PL 2338/2023 establishes a four-tier AI risk framework, banning systems that violate fundamental rights and mandating impact assessments for high-risk applications. The bill's 199 Senate amendments reveal where compliance obligations remain contested, from oversight authority to generative
Human-AI pairs beat either working alone, but the decisive factor is process, not raw capability. Drawing on Kasparov's centaur chess model, this post explains how comparative advantage between human judgment and AI execution creates results neither could achieve solo.
An AI-driven iteration loop made the Snes9x emulator 61% faster on x86-64 and 52% faster on arm64 in three days for under $30, with every frame and audio sample bit-exact against the original. The argument: the cheapest performance gain is the compute already wasted in mature code.
From t-shirts that defeat thermal cameras to instructions injected into tool outputs, adversarial AI attacks now run directly through human bodies and input channels, undermining the assumption that humans permanently control the upstream position.
Agentic flows can make LLM bills snowball: every tool call adds context that gets sent again on the next turn. Prompt caching can keep those costs under control, but cache minimums, write surcharges, and TTLs mean you need to design your agent around the cache—not just hope it saves money.
Most of what you pay an AI coder per turn goes to overhead, not actual work. Real instrumentation shows tool schemas alone consume 68% of per-turn costs, and straightforward fixes like lazy-loading cut total session spend by 47%.
Qwen 3.8 27B may be Alibaba's most important release—not because it's the biggest, but because it brings frontier-adjacent AI to a single consumer GPU. As local models improve, owning AI is becoming a realistic alternative to renting it.
AI systems were built to compete, and that adversarial dynamic has now spilled out of the lab and into the open information commons. When machines fight over content, meaning becomes collateral damage, and the only durable human advantage is staying outside the loop as a verifiable source.
Distillation isn't a cyberattack, and violating an AI lab's Terms of Service isn't a crime. But as the Anthropic vs. Alibaba dispute reveals, the broad "do not compete" clauses in AI user ToS could turn everyday software development into a legal minefield.
When AI commoditizes expertise, the content artifact is no longer the moat. Publish the method generously, demonstrate judgment through specific real cases, and build the direct audience relationships that no algorithm can replicate or replace.
Building hard systems with AI coding agents doesn't transfer finished knowledge into your head. It exposes you to decisions under uncertainty at high speed, forging durable judgment that survives when the tool is gone.
News Update: Alibaba has announced Qwen 3.8 as an open-weight model. Following Kimi K3, this is another strong signal that frontier open-weight models are becoming more common—and that the gap with the best closed models may be shrinking faster than many expected.