Sunny

← All weeks

What I’ve been reading (W33, 2025)

More AI things — GPT5, AI at Meta/Facebook, AI search reflections and thinking in China vs US

  1. Meta is improving ad performance with AI, they use Andromeda, a retrieval system co-designed with custom chips (Grace Hopper + MTIA) that looks at millions of ads to thousands in real time and AdLlama, a Llama 2–based generator fine-tuned first on curated ad text and then via reinforcement learning from performance feedback (CTR as the reward signal with policy gradients + KL control), creating a closed-loop, population-scale optimization engine that only Meta can run because of its unique data, compute infrastructure, and end-to-end control. These systems helped deliver 22% year-over-year revenue growth to $47.5 billion in Q2 2025, with about 5% higher conversions on Instagram and 3% on Facebook. The quarter showed Meta can still generate a lot of cash from ads, and that consistency is the permission Zuck needed to spend freely on superintelligence and LLaMA 4.x. https://stratechery.com/2025/meta-earnings-meta-turns-the-dial-social-network-r-i-p, https://arxiv.org/pdf/2507.21983, https://www.facebook.com/business/news/ai-innovation-in-metas-ads-ranking-driving-advertiser-performance
  2. OpenAI just launched GPT-5. It can keep track of way more info at once, switch between fast and deep-thinking modes automatically, handle text, images, and even email/calendar, help you build apps or debug code, and you can even tweak its personality and tone—so it’s a big step up and worth trying https://openai.com/index/introducing-gpt-5/, https://simonwillison.net/2025/Aug/7/gpt-5/
  3. LLMs are good at generating code and solving problems when the requirements are clear and well-scoped, but they’re bad at maintaining mental models, they forget context, get confused about whether to fix code or tests, hallucinate details, and have recency bias in the context window. A software engineer, by contrast, keeps a consistent mental model, tests deliberately, decides whether to change the code or the requirements, and knows when to step back or start over with a clearer understanding https://zed.dev/blog/why-llms-cant-build-software
  4. China is an engineer’s country, focused on building railways, factories, and tech companies. The US is more of a lawyer’s country good at writing rules but often stuck in red tape. That balance made sense decades ago, but now it’s slowing America down. China’s building drive delivers both big successes (TikTok, BYD, high-speed trains) and big failures (harsh Covid lockdowns, the one-child policy). China isn’t guaranteed to overtake the US, the real risk is America forgetting how to build real things, not just apps https://archive.is/MioJK#selection-1601.0-1601.2
  5. People are searching more and asking longer, more complex questions. In Google’s case, AI Overviews show more links, creating more chances for websites to surface and get clicks. For simple fact-based queries, people may stop at the initial answer, but for exploratory or purchase-related ones, they still click through to dig deeper. These clicks are higher quality, since the AI provides context upfront, and those who continue are more intentional and engaged https://blog.google/products/search/ai-search-driving-more-queries-higher-quality-clicks/