Seonglae Cho영화 리뷰
음악 리뷰 몇년간 진행중 가수들 팔로우 몇명
Hi everyone, I’m Seonglae, an AI engineer and researcher. I joined Holistic last year after winning the Holistic AI hackathon, mainly to work on reversing agent traces. I’m particularly interested in Mechanistic Interpretability and AI Control, and I’m looking forward to any collaboration.
Hi everyone, I’m Seonglae, an AI researcher and engineer at Holistic AI. I joined Holistic last year after winning the Holistic AI hackathon by controlling AI by steering vector idea. Currently here I am mainly working on reversing agent traces to predict next step using current state and early failure prediction by building state machine from the agent trace. In general, I’m particularly interested in Mechanistic Interpretability and AI Control. If you want me to introduce my research journey brielfy, I can do that.
First I started from interpretable AI at Yonsei university at Korea during bachelery to understand LLM deeply. After that I face the great monsemanticity paper that changed my route completely. For mech interp reserach I went to UCL where has many good ereserachers like Neel and Arther courney etc. (anthropic 한명추가 여자 steering vector) . During UCLmaster, I was obseeed to the AI control with steering vector. I worked as a thesis based on correlation and reinforcement learning applied to control and made papers for each. Now I felt a limitation on steering vector and looking more broad area on AI safety such as agent trace based failiture prediction and measureing AI confidence and AI optimism whcih I can talk for each more detail.
limitation 뭘 느꼇냐 질문에
production 효과는 모르겠지만 일단 steerign 만으로는 research 검증에서는 training 에 비해 월등히 보여줄수 있다는 한계가 있다는 걸 느꼇음. 그래서 internal probe 가 요즘 아주 핫하고 결국 회귀하게 되더라. 그걸 통한 internal understanding 을 통해 결국 training method 에 기여하고 싶다는게 요즘 생각
Engineer or Researcher
Interconnectec
Advice
- 바게트 맛남
- Storage requests are often difficult to get approved initially, but persistence pays off. While GPU usage has a poor reputation, it can be quite beneficial if used properly.
- I published a workshop paper from the NLP module at term 2, get teammates very carefully but you have to move quickly to get a good teammate at the same time, even from the term 1
- hackathon appears a lot in london. Make idea before the hackathon is recommended
- There are two career paths: researcher or engineer. Focus on one direction and complete projects aimed at either publishing papers or creating demos, hackathon. I recommend working on a single practical project during Term 1 and trying to publish it, while dedicating Terms 2 and 3 to focusing on your thesis.
- send email as much as you could to reserachers with same interest
- for the paper acceptance, feedback is the most important thing
- do not expect too much from the classes; the real value of this program lies in the connections you make and being in London
- Don't take a deep dive on every topic from the modules since we are graduate students and each person should have their own focus
- For AI Safety, MATS programme and there are a lot of programs and communities in London.
- Checkout LessWrong, LASR, LISA, Safe AI London's MARS AISC etc.
Hello, I was also an AI4SD student last year, and now I am working at Holistic
Mindset Advice
Take a deep dive. Find your true interest. Prove yourself. It takes time. don't worry or be anxious too much.
Overall, don't worry or be anxious too much; instead, create some events within the program
논문쓸때 너무 거창하게 하려고 하지말고 안될거같으면 내가 그냥 원하는거면 빠르게 접어라
