LLMs

HEPI, a UK higher ed thinktank, released the results from their 2025 survey of AI use by students. There’s a lot in the report, so here’s a few quick highlights. The big headline is that basically all students are using genAI: In 2025, we find that the student use of AI has surged in the last year, with almost all students (92%) now using AI in some form, up from 66% in 2024, and some 88% having used GenAI for assessments, up from 53% in 2024. The main uses of GenAI are explaining concepts, summarising articles and suggesting research ideas, but a significant number of students – 18% – have included AI-generated text directly in their work. ...

Piotr Mazurek and Felix Gabriel have an amazing post up on LLM Inference Economics from First Principles, which I found on Bluesky. They go into a huge amount of detail about how inference works and how that affects processing speed. But I saw the graph above and thought “we can get energy from that”. And so I asked chatgpt-o3: “Looking at these graphs of throughput at different batch sizes on a 4xH100 80gb cluster, what ranges of power per token do they equate to?” ...

Notes that I wrote for work on how to get started with Claude Code, Anthropic’s agentic code tool. Anthropic’s own documentation is good, but I wanted a single page that I could share with people to help them get rapidly up to speed. Once past this stage, their Claude code best practices page is excellent. Installation You’ll need to setup an API key to use it. Put this in $ANTHROPIC_API_KEY then run the installer: ...

A very late set of weekly links! This time it’s heavy on AI and it’s impact on research and teaching. Plus cyborg cicadas. First up: A NYMag article on LLM use at Universities went viral (at least in my education and AI heavy feeds!) last week. So that’s not the link as you’ve probably already read it. Instead, here’s Benjamin Breen on how AI makes the Humanities more important and also wierder and Will the Humanities Survive Artificial Intelligence? from the New Yorker. ...

We have some actual numbers on tokens! In their commentary on the latest earnings report, Microsoft report: We processed over 100 trillion tokens this quarter, up 5X year-over-year – including a record 50 trillion tokens last month alone So we’ve got a baseline, and can estimate a growth rate. Assuming exponential growth and a three month token split of 19/31/50 we get 62% month on month growth, or, if it keeps going 100 T → 424 T → 1.8 P → 7.6 P, which is lower than the 5x yoy growth MS quote. ...

As Andy Massick explains well, if you are making a few requests per day to a simple non-reasoning text based AI model the carbon impact is, to use a technical term, bugger all 1. I probably ask ChatGPT and Claude around 8 questions a day on average. Over the course of a year of using them, this uses up the same energy as running a single space heater in my room for 2 hours in total. Not 2 hours per day, just a one-off use of a single space heater for 2 hours […] By being vegan, I have as much climate impact as not prompting ChatGPT 400,000 times each year ...

Finally a week with more than two links! Beatrice de Graaf is a Professor at Utrecht University. Her talk at Stanford on The Quest for a Third Rome describes using text mining techniques to analyse Putin’s speeches since 1999. The paper, written with her student Neils Drost, is open access: Putin and the Third Rome:: Imperial-Eschatological Motives as a Usable Past “In this corpus, we ask how Putin used the combination of history and Russian Orthodox theological, biblical tropes as a usable past, to create his powerbase, to understand how he built up to defending and legitimizing the invasion in Ukraine” A similar area was covered by Teitelbaum in his book War for Eternity about the influence of Alexsandr Dugin and mysticism on both Putin and Steve Bannon in the US. Does Putin believe it? Maybe. Is he using mysticism and religion for his own ends? For sure. ...

Some AI development resources this week… How to Build an AI Code Agent Thorsten Ball has published how to build your own AI code agent. He creates a toy example in Rust that’s <400 lines but still includes tool use. Geoffrey Huntley does the same. It’s worth reading both Claude Code Best Practices Anthropic has released a detailed guide on Claude Code best practices. Of note: Claude can (and should!) use the architect/plan then code model that Aider makes explicit. Telling Claude to read code/docs, then plan, then code, then commit gives better results than just telling it to write code Ask Claude to review things (e.g. the CLAUDE.md hint files) and suggest improvements. Repeat. (As an aside, the Anthropic workbench can improve general prompts) Telling Claude to ’think’, ’think harder’ etc results in more thinking tokens being used Clear context frequently (/clear) - usable context window is smaller than the advertised max The model of asking Claude to review and improve its work (or human starting points) repeatedly was not obvious to me ...

I’m trying not to be perfectionist about this blogging thing, and concentrate on getting stuff out so in that spirit so this week there’s just a few links, all on AI. First up - how to use AI. I see so many people type some random question into ChatGPT/Claude, get a mediocre answer back and move on. But how to get good results from these systems is not at all obvious. For a complex example, see my post on Experimenting with Deep Research or an even more extensive example of using one model to generate a prompt for another ...

Today’s fun is setting up the new Zapier MCP integration. This creates a MCP server that allows Claude Desktop or Cursor or anything else that supports MCP to call any Zaps Setting Up Zapier MCP Step 1: Install mcp-remote First, install the mcp-remote package using npm: npm install mcp-remote Step 2: Configure MCP Endpoint in Claude Desktop In Claude Desktop, you’ll need to add the Zapier MCP endpoint to your settings. Add the following configuration to ~/Library/Application Support/Claude/claude_desktop_config.json ...