HN user
If you’re interested in an easy explainer for backprop, then highly recommend math for deep learning by Kneusel. Finished it recently and you can apply the chain rule backwards by hand with a tiny NN, and code one too.
Saw Empire of AI in a bookshop recently but held off buying as wasn’t sure if it was going to be surface level. You’d recommend?
I struggle with this a bit because while my network isn’t bad - I really can’t stand social network like threads/X etc - I’m not on any social media bar here and LinkedIn.
Do you think investing in bluesky is worth it? I’m in industry but have a PhD ongoing in TTI models so I should probably get on it :/
Nice read - I can’t seem to subscribe? Getting an error with Buttondown.
i’m sorry this really made me laugh - I hope she’s ok!
I think the docs are great to read, but implementing was a completely different story for me, ie, the Ask AI recommended solution for implementing Claude just didn’t work for me.
They do have GitHub discussions where you can raise things, but I also encountered some issues with installation that just made me want to roll the dice on another provider.
They do have a new release coming in a few weeks so I’ll try it again then for sure.
Edit: I think I’m coming across as negative and do want to recommend that it is worth trying out langfuse for sure if you’re looking at observability!
I do understand why it’s a product - it feels a bit like what databricks has with model artifacts. Ie having a repo of prompts so you can track performance changes against is good. Especially if say you have users other than engineers touching them (ie product manager wants to AB).
Having said that, I struggled a lot with actually implementing langfuse due to numerous bugs/confusing AI driven documentation. So I’m amazed that it’s being bought to be really frank. I was just on the free version in order to look at it and make a broader recommendation, I wasn’t particularly impressed. Mileage may vary though, perhaps it’s a me issue.
I’ve been slowing crunching through Math for Deep Learning, so spent a fair amount of time looking at Hessian matrices + second order optimisation. I’ve been slowly reading this book for a year, so stopping to do most of the math by hand each time. One chapter to go!
Then I was sick all last week, so ended up down a rabbit hole about the current card collecting bubble (right word?). Super interesting.
Amazing! Thanks will join.
This is fantastic. I think it’s nailed in the substack what was missing from a lot of these LLM driven NPCs that did not feel authentic. I have a couple of follow-up questions on specifics relating to analysis of behaviour with LLMs (in game-dev myself). Would it be possible to speak to you directly on them?
GAMM in Rome is quite fun (Italy):
Shout-out to the museum of Ancient Greek technology, with their wine automaton (Athens, Greece): https://kotsanas.com/
Bletchley is great! I need to go again
Haven’t seen Wiby before but it’s fantastic, thank you.
The response to the Sal Khan op-ed resonated with me, along with other parts of this article. Something I’ve been digging more into is some of the figures around proposed job losses from AI. I think I even posted a simulation paper last week.
After posting that, I came across numerous papers which critique Frey & Osborne’s approach, who are some of the forefathers for the AI job losses figures we see banded around commonly these days. One such paper is here but i can dig out others: https://melbourneinstitute.unimelb.edu.au/__data/assets/pdf_...
It has made me very cautious around bold statements on AI - and I was already at the cautious end.
Thank you for putting it together. I looked at a couple of the references and they look like they point to your blog. Do you have a view at all of popular patterns in terms of citations? Might be useful
I thought Gas Town was a satire until I saw the GitHub. Maybe it’s a very involved satire?
It is right? “ Do not use Gas Town.”
Catch 22 is just fantastic isn’t it. I might re-read in 2026!
Something I’ve been thinking about is how as end stage users (eg building our own “thing” on top of an LLM) we can broadly verify it’s doing what we need without benchmarks. Does a set of custom evals built out over time solve this? Is there more we can do?
Completely agree. I was working on something with TensorRT LLM and threw Nemotron in there more on a whim. It completely mopped the floor with other models for my task (text style transfer), following joint moderation with another LLM & humans. Really impressed.
Complete tangent but “in lockstep” really seems to be doing the rounds again as corpo speak. It honestly annoys me but I can’t explain why - maybe because they’re not a military but an ad production company.
Amazing, thank you! Will go through these.
Season of the Witch is honestly such a great book. Nice to see it mentioned here.
Wow, thanks for link to Paged Out. Any ideas how to discover any more tech focused zines? We have a zine culture in UK but afaik it’s more culture / music focused (v happy to be wrong here)
Genuinely VSCode has been broken for me for with copying due to it desperately trying to vibe code for me. You’ve reminded me to fix that.
If you’re in the UK this is quite an easy solution. The library is right by two large train stations and you can sign up same day for a card.
Oh great flag that it’s open access. Will give this a read.
Agreed. NotebookLM is one of my most used tools for my research work. I’m able to give the audio overview to friends and family who don’t want to read a paper but am interested in what I’m doing. And the “live” talk allows me to interrogate my own work and identify gaps where I haven’t explained something well.
Great post, thanks to you I’ve realised I use the “how much am I reading” to how burnt out am I proxy.
Quite tempted by the phone, but predominately a physical book reader.
I completely agree with this. I’m not surprised by the fine tuning examples at all, as we have a long history of seeing how we can improve an LM’s ability to take on a task via fine tuning compared to base.
I suppose it’s interesting in this example but naively, I feel like we’ve seen this behaviour overall from BERT onwards.