Notes on what I’m studying 📝 * RubyOnRails * Professionalism * AI projects in Ruby * Building with LLMs
England
Joined December 2021
- Tweets3.5K
- Following999
- Followers592
- Likes15.2K
Session spent doing a bit more work on mario rl ready to kick off world 3 training run. Have changed how we distribute levels so we focus on those we're currently failing at, rather than the previous even distribution. Also read some more of
1 / 2
1
21
Session spent reading a bit of hundred page deep rl book (have ordered a physical copy). Good so far, looks like there's a fair bit of notation so will have to put some effort in! Mario RL is still on going, making a bit of progress on levels 2
1 / 2
1
29
Session spent on Mario rl (and grinding some math academy). Model now comfortably beats world 2 - 1. Skipping 2 - 2 for the minute as it's a water world. Going to train on world 2 - 3 and 2 - 4 at the same time, see how that goes.
1
101
Session spent continuing on mario rl project. Had to use up 15 mins of my session on math academy to get my xp for the day. Was lucky, mostly questions where I didn't need pen and paper! Mario continues to improve world 2 - 1 clear rate. Having
1 / 3
1
49
claude pull our a C extension which should take an 8 hour run down to 6.5 hours. It's a nice piece to pull out as it's just frame processing so doesn't really impact my ability to read the ML code. Do need to clean up the ruby a bit as there's
2 / 3
1
32
Session spent continuing on mario rl project. Last training run improved things, but needed more steps, have kicked another off. Did some discussion with GPT to get some concepts straight and had a dig throught the code. Cleaned up my
1 / 2
1
60
Session spent doing a bit more on mario rl and then reading this v interesting article goodfire.com/research/mechan… - will finish it tomorrow. Made progress on world 2 - 1. About 25% of runs now clear 2 - 1. Think it
1 / 2
1
39
Session spent reading the art of innovation. Needed something to do while waiting on my Mario rl training run to finish! World 2 - 1 looking promising
43
Session spent mostly reading through my mario rl project's notes I have claude write for me. IDK why but I find reading LLM generated text quite tiring / hard to get into. Also mario rl is now beating all of world 1 again (having switched from
1 / 2
1
64
Wow! I should see if Claude can just write the weights for my mario rl project! Bit less fun though
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
41
Session spent watching a video on how llms inference is served. Waiting on latest Mario training run to finish. New PPO model beats level 1-1, 1-2, 1-4, but isn’t beating level 1-3 yet, hoping this run sorts it
1
1
95
study every day 💻📝📚 retweeted
This is so well said.
Ed McLaren (@Anglican_Gonzo) destroys Matt Goodwin for his politically correct celebration of ‘model minorities’ on the latest episode of The Forge with @Con_Tomlinson and yours truly 🔥
Link to the full episode below.