Idea diary | Notion

I try to write ideas down when they pop up and this is a running collection. If any of them sounds interesting to you, please reach out :D

Last updated Sept 2025.

Products/Applications

*Haven’t thought about these in a while, mostly doing research now!

LLM for peer reviewing because it’s so broken in ML…
- Turns out my professor dropped a paper on this! Check it out: https://arxiv.org/abs/2504.09737
- Some kind of ‘reproducibility agent’ that goes and fixes code base releases.
Finetune a vision model to describe in real time the items most relevant for situational awareness (some inference optimization) and attach a camera to glasses/some other eyewear. Sort of allow blind people to ‘see’
TAs don’t get paid enough to grade manually written homework. What is we use OCR+LLM to make an autograder and then leave the regrade requests for TAs?
An AI/ML journal/conference/workshop that publishes negative results
- In addition to current flaws of the incentive structure of academia. it inherently encourages capabilities research that will get many references because they’re immediately applicable. Incentive for documenting negative results can help alignment because we also need to know what DOESN’T WORK
An browser extension over LLM interfaces to collect hallucinations/bad responses in general. Companies need bad examples to RLHF their LLM continuously and improve then, not sure how they collect it at the moment
Ok this one is goofy but what if we made a place where people can rant about their jobs and problems in their life (in detail) and people looking to solve these problems (us) can go contact them to get more info and feedback
An agency to do deep tech diligence for investors for VCs
- wtf happened with Rabbit R1
Research the correlation between capability at like $1k worth of training and capability when training is complete, could use neuroevolution to discover better architectures for LLMs.
Community notes + reddit native on every site, all over the internet
- Already done, but hard to grow socially hahaha
RAG + Anki automation for books & papers
Gamify and socialize productivity - points app to see visible progress, “we make your efforts seen”
- Ok I know how cringe this sounds but lemme cook
Robot powered kitchen on campuses selling weekly meal prep
Help yt and podcasters reach international audiences with AI voice and language translation
LLMs as stream/chat/server moderators
Alignment Hackathons
Power up LLM content detectors and put them ALL over the internet, pitch to chrome/edge, this is a BIG issue rn and will only get worse
Webapp for career fair and networking events - attendees can share their resume to a database accessible by search by employers, employers can star resumes etc, should be transparent to students too
- Feature: contact sharing for networking: attendees fill out all forms of contact/profiles they’re comfortable with sharing, ability to star people etc
Train/API a classification cv model and put them inside low cost cameras on frequently used bins, government contract
Producing AI tools review and under the hood info content, too many useless wrappers
Reinvent search engine, complete LLM engine with multiple interfaces (Screen, voice, or even neural in the future) companies pay to rank higher
Use AI to detect deepfakes
- Tbh all firms developing production models should be required to produce a tool that can reliably identify content their models created
Mitigate social misalignment on the integration of AI systems into society
Start a autopilot clout optimizer twitter acc
LLM system for assessing the rationality of argument and award bayes points, twitter acc
Birthday reminder app with ads for gifts

Research Ideas

Recent thoughts in the cooperative AI research agenda: https://docs.google.com/document/d/1sw3tXEdHs66ZDyswV1nt-I2mZlfkrqmyjOF3STHwXPk/edit?tab=t.0
A mega project / an organizations dedicated to tracking important metrics/methods over time, across models. Lots of important research/evals work happen without releasing large benchmarks that are continually used (especially for behaviors like scheming and techniques designed to identify them) and they are never used again after the paper is published. We need to track how well our techniques work over time.
Increasingly concerned about gradual disempowerment in worlds where we have “gpt5 level AI that’s aligned enough to be a product, but also not AGI/ASI” for prolonged periods of time.
- This seems to be the vision of all major tech companies/investors who are not ‘AGI pilled’ or whatever that means. The world of self-fulfilling at an extent.
- I don’t think a feasible technical solution exists. This is like a collective choice problem where we need to show that resources should be directed towards places where AI can readily improve the human condition - build ‘tools’ not ‘agents’. We need founders!
Study scaling laws of activation engineering techniques
- If mechanistic interpretability works out in the long term, alignment/steering techniques would likely involve modifying activations at inference time. This would be valuable knowledge.
- Published a paper on this and it turns out that OOTB activation engineering techniques fail to generalize to off-distribution but semantically consistent contexts… making its application a systems level challenge.
  - https://arxiv.org/abs/2505.03189
Improving in-the-open reasoning in natural language
- Systems like language model agents that reason in the open are probably more interpretable and therefore controllable. If we want this to be used, we need to make it good enough to beat black box reasoning methods
Pre training scoping: we do the HF part of RLHF to decide which parts to include in the data. Cheaper than AIF as the reward model should be way smaller!
Research agenda focusing on black-box psychology style research on LLMs/agents. Focus on reproducing the important studies in human psychology on this new form of intelligence. Based on this.