quick testing, but Mistral free tier actually seems like a good option for teaching a module on llm agents? quick api set up, medium 3.5 seems capable enough, from what I can tell rate limits are decent??

Has anyone managed to use eg openrouter/free for introducing students to llm agents in teaching? What’s the best free yet passable and generally accessible option for students these days?

testing qwen3.6:27b in pi studio, now with file navigation/viewer, pane resizing, zen mode, keyboard shortcuts, REPL, simple text/code completion suggestions

added simple collaborative REPLs to pi studio...

Linearising consciousness by taking the neural tangent kernel

mature things tend to get pressured into being less weird over time, so if you're starting from scratch your advantage is that you can be very weird if you don't start weird you're probably trying to compete inside the local maximum of something more boring than you could be

Finally uploaded a revised version of this preprint on parameter identifiability arxiv.org/abs/2502.04867. Approach called 'invariant image reparameterisation' though 'image' used in mathematical sense not picture sense. Took me a while to revise for various reasons, but fairly happy with it now

made a low quality little demo of my pi studio gui thing for pi. see github.com/omaclaren/pi...

New agent manager: just use a browser with vertical tabs + tmux + ttyd

I doubt you can take these stats at face value but this a lot more than have cited my academic papers to date…

inspired by opencode folk showing love to the people of pi, I had a go at slopping up an opencode plug-in version of my pi extension pi-studio...seems to work! see github.com/omaclaren/pi...

for people of pi (and on mac since uses glimpse). instead of asking models side questions while it does stuff, you answer questions about the code in a side quiz: github.com/omaclaren/pi... not sure if any good yet but another attempt to encourage more active code engagement when using agents

another little pi extension -- pi-repl (github.com/omaclaren/pi...)

prettied up pi studio (github.com/omaclaren/pi...) with more syntax highlighting and stuff... no idea if actually useful lol but fun being able to make tools like this so quickly with pi.dev (thanks @mariozechner.at )

Me and pi made another pi.dev extension…this time ‘pi studio’: opens a (local) browser workspace for annotating model responses/files, running edited prompts, and requesting critiques www.npmjs.com/package/pi-s...

pi + codex + zed is the most peaceful workflow for me at the moment. Codex is less sneaky and makes less dumb mistakes than claude, pi lets you add your own little simple workflow tweaks and tools easily, and zed is nice and simple and fast for actually reading code

Got pi coding agent to make itself an extension to render markdown replies with equations in the terminal or browser www.npmjs.com/package/pi-m...

I missed gh copilot suggest after they deprecated it so here’s a simple bash function to use claude code etc to suggest executable shell commands like it used to gist.github.com/omaclaren/e1... Same deal as eg Simon Willison’s llm-cmd but uses your clis instead of apis

Minimal bash function (add to .bashrc) to generate executable shell commands from natural language. Uses claude code in non-interactive mode (with -p). Can be adapted to other clis easily

Minimal bash function (add to .bashrc) to generate executable shell commands from natural language. Uses claude code in non-interactive mode (with -p). Can be adapted to other clis easily - ccc.sh

gist.github.com

Reading the world beyond your head and it’s talking about my claude code usage

What's the best 'agentic CLI' tool that is easy to set up with a free tier (no cc details or initial credit req.) for students to use in a university course? Main path I can think of is github copilot student developer pack + e.g. opencode (also just agent mode in vscode...)?

Some lecture notes on spatial and temporal modelling, illustrated using R - darrenjw.github.io/spatio-tempo... - #rstats #rspatial #quarto

Olmo 3 is notable as a "fully open" LLM - all of the training data is published, plus complete details on how the training process was run. I tried out the 32B thinking model and the 7B instruct models, + thoughts on why transparent training data is so important simonwillison.net/2025/Nov/22/...

Olmo 3 is a fully open LLM

Olmo is the LLM series from Ai2—the Allen institute for AI. Unlike most open weight models these are notable for including the full training data, training process and checkpoints along …

simonwillison.net

Any really good actually insightful or helpful articles or resources on teaching/learning in the age of LLMs?

Some misc observations on trying to use LLMs to help on a current research project…

Can we infer chemical reaction networks from time series data? Can we observe concentrations of substances over time and infer which substances react with each other? Learn if (and how much) this is possible from Yong See Foo, @adrianazanca.bsky.social and JenniferFlegg arxiv.org/abs/2505.15653

Quantifying structural uncertainty in chemical reaction network inference

Dynamical systems in chemistry and biology are complex, and one often does not have comprehensive knowledge about the interactions involved. Chemical reaction network (CRN) inference aims to identify,...

arxiv.org

Evolutionary algorithms are back maybe?

Matej Balog@matejbalog.bsky.social · last yr.

Announcing AlphaEvolve, our new LLM coding agent that has - made new scientific discoveries - discovered algorithms that are now deployed at Google (in Gemini, Transformers, TPU hardware design & data centers) Blog: deepmind.google/discover/blo... White paper: storage.googleapis.com/deepmind-med...

I tried defining a simple ‘prompt script language’ that sort of works and is fun to play with gist.github.com/omaclaren/b5...

I love you bluesky but I admit I don’t totally understand the rabid and uncritical AI hate that predominates here

The voice mode for all the llms keeps trying to sound more human but I kinda just want my robot friends to sound like robots