TL;DR
-
Research: Exploring and implementing the application of generative AI for visual inspection using multimodal Large Language Models (LLMs).
-
Life: Fully enjoying the Christmas season. Savoring new encounters and interactions.
0. San Fransisco
People from all over the world have settled here, drawn by the beautiful cityscape visible from the top of the hill and the perfect climate. However, the public safety situation is steadily worsening. There's a noticeable presence of homelessness, and incidents of vehicle break-ins are frequent. The decrease in daytime population due to remote work, along with the decriminalization of shoplifting items worth less than $950, has led to a surge in group thefts, causing many retailers to withdraw. Numerous properties now lie vacant. It's intriguing to consider the future of a city known for its IT industry, liberalism, and diversity.
1. Research
(12/10-16 NeurIPS)
12/11 MeetMakers @D.School,Stanford
12/13 re:InventRecap @AWS,SF
The lab is extremely busy with NeurIPS. As soon as it concludes, we'll be heading into winter break.
Progress
- PoC: Developing a basic app based on GPT4V for visual inspection services. Currently exploring prompting techniques.
- Although it's still under consideration, achieving accuracy comparable to supervised learning using the current LLM's Zero-shot approach seems challenging.
- Expecting that it can contribute to an "Early Deploy Strategy" for system and inspection specification revisions, as it shows a certain level of performance as an intermediate-stage AI.
- Research and Paper Review:
- AGI
- Thinking System: Creating a "thinking" unit that converts inference time into accuracy. (ref. Tree of Thought)
- Self-Improving AI: Comparable to LLM version of Alpha Go
- Superalignment: Monitoring an AI student whose performance far exceeds its human teachers
- LLM Architecture
- Mamba: Alternative to Attention mechanism
- Hyena: Alternative to Attention mechanism
- Monarch Mixer: Alternative to Attention mechanism
- Material Science
- GNoME: Exploring new material candidates using GNN and DFT (Deepmind)
- MatterGen: Generating material candidates with StableDiffusion (Microsoft)
- AGI
Future Plan
- Experimenting with multimodal prompting (Few-shot Examples, Visual Reference Prompting, etc.)
- Implementing and comparing performance of multimodal LLM engines other than GPT4V (LLaVA, Gemini)
- Exploring edge inference implementation and assessing its necessity (Gemini nano, Ollama)
- Interested in expanding into defect image generation as well.
MeetMakers
Attended a results presentation meeting for a Design Thinking lecture at D.School.
re:InventRecap
Traveled an hour from home to AWS San Francisco to catch up on the latest releases.
- Bedrock: A LLMOps service for generative AI. Utilizes FM models, fine-tuning, RAG, Agents, etc.
- Q: A chatbot specialized in the AWS domain, similar to Google's DuetAI.
- Sagemaker Hyperpod: A training environment for CustomFM.
- One: Palm recognition is intriguing (identity authentication using palmistry).
- Monitron: An industrial anomaly detection service.
2. Life
12/04 🎄Celebration @Downtown,MV
12/05 ESL: 🇺🇸Mary(Are friendship forever?)
12/08 🍽️StanfordCS
12/09 🎄Concert @MemorialChurch
12/10 🍵🇺🇸Liz
12/14 🎄Party @🇺🇸Tom&Steffi🏠
12/16 🎄Party @My🏠
12/17 🍽️🇺🇸Liz
12/25 🎄Party @🇺🇸Liz🏠
12/27 🍵🇺🇸Arden
12/31 ✈️ -> 🇲🇽MexicoCity
12/01,08,15 GlobalChef 🍽️
Christmas Events
🎄 Season. Participating in many events.