Is Terence Tao actually using LLMs for math or just for the

Terence Tao uses large language models not as mathematical oracles but as tools to manage information density, summarizing technical sections and writing verification code, according to a practical guide for high-level research. The article warns that LLMs fail at pure mathematical reasoning due to probabilistic patterns, logical gaps, notation confusion, and lack of formal rigor, advising researchers to automate tedious tasks rather than delegate logic.

Is Terence Tao actually using LLMs for math or just for the Tao doesn't treat AI as an oracle. Instead, he treats it as a tool for managing information density. When you are working at the frontier of mathematics, the sheer amount of existing papers, theorems, and lemmas is overwhelming. He uses these models to help parse through complex documentation and to summarize specific technical sections, which essentially acts as a force multiplier for his reading speed. The practical AI workflow for high-level research If you want to move beyond basic chat prompts and actually use an LLM agent for deep research, you have to change your mental model. Here is how a top-tier researcher approaches the technology: 1. Information Synthesis over Generation: Rather than asking "Prove this theorem," the prompt engineering focus should be on "Summarize the core logical steps of this specific argument" or "Identify the connections between these two disparate lemmas." The goal is to reduce the cognitive load required to scan a paper. 2. Code-Assisted Verification: For mathematicians working with computational aspects, using AI to write small scripts for testing conjectures is highly effective. It’s not about the AI knowing the math; it’s about the AI knowing how to write the Python or Lean code that verifies the math. 3. Explaining Complexity: One of the best uses for a large language model is to act as a "rubber duck" for complex concepts. If a particular definition in a paper is phrased in an incredibly dense way, asking the model to "rephrase this concept using simpler terminology" can help bridge the gap before diving back into the formal notation. Why LLMs fail at pure mathematical reasoning We have to be honest about the limitations. If you try to use a standard LLM for high-level symbolic reasoning without a formal verification system, you are asking for hallucinations. The model operates on probabilistic patterns, not logical necessity. Logic Gaps: LLMs are prone to "logical leaps" where they skip a step that looks intuitively correct but is mathematically invalid. Notation Confusion: In highly specialized fields, the notation can be non-standard, and the model often reverts to the most common but incorrect interpretation it saw during training. Lack of Formal Rigor: An LLM cannot "feel" the weight of a proof. It doesn't understand truth; it understands the likelihood of the next token. For anyone looking for a real-world deployment of AI in technical fields, the takeaway is clear: don't look for a replacement for your brain, look for a way to automate the tedious parts of your research pipeline. Use it to organize, summarize, and code, but never let it hold the pen when the actual logic is being written. Next South Korea is making AI a public utility for every single → /en/news/8009/ a practical ChatGPT prompt guide https://tanyan888.com/ , with plenty of directly applicable cases.