Sunny

← All weeks

What I’ve been reading (W41, 2025)

Nondeterminism in LLMs, AI wealth, Oracle’s rise, and nerding out on computers

  1. Defeating nondeterminism in LLM inference https://thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/
    • All text and data is converted into vectors through embeddings
    • An LLM prediction is a massive sequence of floating-point operations such as matrix multiplies, reductions, and non-linearities on those vectors
    • Floating-point operations are approximate, and changing the order of additions or multiplications can cause rounding to shift results slightly
    • On modern hardware, operations are run in parallel and often batched together with other requests, which changes the shapes going into kernels
    • A kernel is like the cashier that processes your math, and many kernels today are not batch-invariant, so their internal math order depends on batch size or layout, causing two identical inputs to take slightly different computational paths depending on who else is batched with you
    • These tiny shifts usually do not matter, but when two candidate tokens are very close in probability, that minuscule difference can flip the winner
    • The nondeterminism is not just random floats, but also system-level batching and kernels that respond differently to different batch sizes, making LLM inference nondeterministic in practice
    • The fix is to design batch-invariant kernels that always give the same result for the same input, no matter the batch size, so server batching stays efficient and outputs stay deterministic
  2. AI will not make you rich https://joincolossus.com/article/ai-will-not-make-you-rich
    • AI growth looks less like the internet and more like containerisation: a big cost-saving technology that standardizes infrastructure, consolidates quickly, and mostly benefits downstream industries and consumers
    • The internet era made investors who bet on “the new thing” rich because experimentation was cheap and permissionless and anyone could launch a startup from a dorm room or garage
    • That freedom to tinker created surprising new markets and gave rise to entirely new giants (Google, Amazon, Facebook), not just extensions of old ones
    • By contrast, AI tinkering is constrained: building frontier models is too expensive, access is controlled by incumbents, and most app value can be captured by the model providers.
    • So fortunes in AI won’t come from owning the new tech itself, but from companies that reimagine industries with AI-driven cost savings and pass benefits on to consumers.
  3. Why language models hallucinate https://openai.com/index/why-language-models-hallucinate/
    • Hallucinations are plausible but false statements generated by language models
    • Language models hallucinate because standard training and evaluations measure model performance in a way that encourages guessing rather than honesty about uncertainty
    • So the fix is to penalize confident errors more than you penalize uncertainty, and give partial credit for appropriate expressions of uncertainty during evaluations
  4. Oracle https://www.fabricatedknowledge.com/p/oracle-and-animal-spirits
    • Oracle is going all-in on building the biggest GPU-first cloud, mainly for OpenAI. Microsoft slowed spend, Oracle stepped in, and became OpenAI’s core infra partner.
    • Their database still runs the world. For ERP, CRM, and warehousing it’s the default. Postgres and Mongo get the hype, but they’re not truly substitutes yet.
    • Oracle’s bet is to bring vectors and inference right into the database. They already sit inside every enterprise, and now they’re also the GPU heavyweight.
    • If they tie their database franchise to inference, they could become the default AI runtime for enterprises everywhere.
  5. Processes and threads https://planetscale.com/blog/processes-and-threads