Oliver Maclaren

Not necessarily my favourite and nothing on reinforcement learning but does have decision making under uncertainty github.com/omaclaren/op...

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

Gödel Escher BaNAch Spaces

Anecdotally lots of other researchers using LLMs to help with research also find themselves getting infuriated with them lol

So quite useful I think but also at times infuriating and messy. Definitely can’t just ‘vibe research’ without things getting messy and or completely misleading

Also hard to keep track of messy context when exploring different avenues and having to backtrack or discard ideas. Still need to actively manage all the different threads and context yourself and sometimes just start fresh or combine summaries of results so far etc

Good: Definitely help explore and test ideas quicker. Make implementing quick test code pretty easy. Bad: infuriating tendency to claim success based on clear failures (‘Perfect! These results clearly demonstrate the theory working exactly as predicted ✅’ — garbage results).

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

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,...

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

Evolutionary algorithms are back maybe?

You’re getting a lot of hate for this on this site even with all the caveats lol but fwiw I think your basic statement is pretty obviously true. For better or worse, GenAI is not going away and better we adapt and learn how to best use this new tool than pretend this isn’t the case.

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

Working a treat so far! 🙏

Nice, thanks!

Pros and cons of them vs other similar providers? You’ve had good experiences with them?

I’ve got some students who wanna play around with eg 70B open llms. Can’t run locally obvs. What’re their best options in terms of ease of use and cost? Mainly just running for now but also maybe fine tuning at some point

This choice clearly shaped by a falling out over me being unhappy with another paper due to a) again not properly citing prior work (by my past PhD student) and b) glaring technical issues. Instead of fixing the resolution was to remove me as author. See attached

Yeah essentially. Was discussed as collaborative follow up work building on first paper. Now presented as standalone alternative with a citation to original work buried in a list of many.

No matter how much beef you have with someone you don’t need to do this shit. Crediting people appropriately is allowed

In my mind the original paper is pretty much just an attempt at ‘what would David Cox do if he was analysing simple population growth models?’ (at least my contribution was to try make it that). It’s all pretty straightforward stuff tbh but people found it useful

Invariant Image Reparameterisation: A Unified Approach to Structural and Practical Identifiability and Model Reduction Both structural and practical parameter non-identifiability present fundamental challenges when using mathematical models to interpret data. This issue is particularly acute in complex, applied areas ...

I mean, it should apply to them too! Models considered were for growth scenarios but no reason couldn’t apply to decay…Also, a closely related but more obscure application of ideas to a decay model here arxiv.org/abs/2502.04867

Here’s a thing. A bit bitter sweet for a few reasons, and also my name is spelled wrong lol but still www.sciencedirect.com/journal/jour...

Claude Code is very cool and helpful but also

great stuff ay

Similarly with basic lit review or even arithmetic questions. I use AI constantly at the moment and it’s incredibly useful but the hallucination issue — essentially a lack of understanding issue — is still massive for many real use cases

I mean I think AI is very impressive but it’s very frustrating in a technical setting that you can eg upload a manual on a piece of simulation software, ask an LLM a question clearly answered directly in the manual and get a hallucinated answer with high prob.

A Bear Case: My Predictions Regarding AI Progress — LessWrong This isn't really a "timeline", as such – I don't know the timings – but this is my current, fairly optimistic take on where we're heading. …

Smashing the endorse button as fast as I can
www.lesswrong.com/posts/oKAFFv...

Some material on statistical decision theory here tho mileage may vary! github.com/omaclaren/op...