In this issue: Inference scaling for Deep Research, Mountain climbing, podcasts about AI progress and Consciousness, scientific funding opportunities and misc links about travel and and RL.
Previous month: Links For September 2025
Projects
- In the first week of October I wrote Test Time Diffusion Deep Research tool for a competition. Given a question, it creates a research plan, searches API for answers, chunks results, ranks them using LLM ranker, and does several iteration of these during inference time to improve before writing a final report. The code is not that difficult, the biggest problem was to fit two local Qwen3 models into a 24gb VRAM docker container and to not go out of memory. Check out a git repo for more detailed description!
- Not exactly a project, but I went multi pitch climbing to the mountains around Arco, Italy! It was very scary, but very fun, here’s a good guide on what one needs to learn before climbing multi pitch, and here’s an example route
Links
- I really liked these videos about macroscopic movements
- Analogue is an R&D ecosystem with a fund at its core. They want to back foundational insight before it’s legible to institutions.
- I like to travel light and this post with the best gear is useful
- I love this post about why big (tech) companies are may be slower on purpose, matches with my observations
- There’s an amazing podcast episode from Dwarkesh Patel with Andrej Karpathy — AGI is still a decade away. This also matches with my experience, especially about LLMs for coding and AI agents, which fail hard if your work is out of distribution
- Another amazing video: Towards Consciousness Engineering by Max Hodak which I liked so much I want to write down my own thoughts on this topic, but it’s very hard!
- I worked with Reinforcement Learning and I saw lots of parallels with real life, here’s a good post that describes some of my thoughts on it
- And lastly, here’s a Bio Protocol Researcher Referral Program
Before you start prioritizing tasks, understand who the end user is — the person who’ll actually benefit from what you’re creating
Funny that another quote
Make a list. Check it twice. Write down all the tasks that could possibly be done within the project. Then set priorities. Put at the top of the list those tasks that have the highest value and the lowest risk.
may be exactly the opposite for research projects, where one has to start with a highest risk/uncertainty to figure out what’s the biggest unknown