LookAside

如果说LLM带来的机会是如同互联网或者移动手机的级别的话，那LLM扮演的将会是一种全新处理信息的方式，而不是一种自动化手段。 因为传统的系统需要explicitly对事物逻辑进行建模，i.e. relational schema，而LLM则implicitly对事物逻辑进行建模，i.e. model weight。

如果说过去20年的搜广推解决了信息触达的问题，那么为什么人们还会有信息茧房呢？ 因为瓶颈是在信息的处理和整理上, which is why we build lookaside。

具体而言，LookAside解决几个核心问题：

1. Capture
2. Organize
3. Output

Why, Who and How?

User persona: knowledge workers with doing research as the center of the worklow.

Current Offerings

1. Pocket
2. Cubox
3. Protolyst
4. Elicit: Find research paper to cite and define reserach directions.
5. Infranodus

Alternative solutions:

1. Note-taking Apps
2. Bookmark Apps
3. Read-it-later / Reading Apps

How to establish a scaling advantage?

How to make LookAside 5% better every week?

How to solve the cold-start problem?

Short Term Planning

1. Do a comprehensive research on existing researching tools
   1. what do they offer and what's their understanding of the problem?
2. Derisk the intelligence features, i.e. prompt engineering
   1. Evaluate relevant offerings
   2. Implement LlamaIndex with Vercel + Supabase + Modal
3. Derisk the objective decompose capability
   1. Study Relevant Projects' Codebase

Relevant Researching Tools:

1. Guides & Notes by Genei
2. Cubox
3. UpWord on Knowledge Works powered by AI
4. Elink on bookmarking for content creation
5. Elicit on AI assisted research

Relevant Projects:

1. DataBerry
2. LlamaIndex Citation Query Engine
3. MarkPrompt
4. Mendable
5. AskWise
6. Protolyst

Relevant Github Projects:

1. LlamaIndex -- How LlamaIndex chunk the docs?
2. AutoGPT -- How LLM decomposes an objective?
3. gpt-engineer -- How users interact with the system
4. embedchain -- How it load youtube videos?
5. aider -- UX logic with LLMs
6. Danswer -- QA query on private data
7. EasyEdit -- Modify LLMs to align with intent
8. AutoDoc -- Auto generate Docs based on codebase
9. khoj -- NBX for Emacs & Obsidian
10. llm_agents -- Study how LLM agent works
11. Customer Service GPT -- How to use process to drive GPT models
12. ArxivDigest -- How to integrate the pipeline from Arxiv
13. gpt-researcher -- How to orchestrate differnt agents

Intelligence Features

1. QA bot based on the content of a set of records (LlamaIndex)
2. Decompose an objective into smaller more actionable ones (AutoGPT)
3. Consolidate two objectives into one objective
4. Reason the object's viability based on the collected information
5. Reason the gap between the available information and the objective and ask for more information (gpt-engineer)
6. Topic modeling based on the content of the saved records

Process to learn a codebase

1. Identify a typical user workflow (E2E tests)
2. Identify the key subsystems and how they are composed together
3. Identify the key APIs & data structures of the system

Questions to answer

• Objective和Task的关系和区别是什么？
• Annotation/Capture, whether having it really matters or not?
• Record的多种形态：Webpage，Notes，Files

Vision

LookAside is an attempt to rethink how knowledge workers work in the LLM area.

1. Human should focus on collecting the right information, LLM would do the work automatically.
   1. Evidence: Elink create content automatically
2. User owns the objective, it cannot be delegated to the system

Power of Random Thoughts

Putting the right information in the right place is the key to remedy information anxiety.

How to find seed users?

1. VCs are idea customers but probably not the best for seed users, as they don't keep the secret.
2. Ideal seed users would be the people that won't raise legal issues with me, i.e. copyright
   1. Friends, but they care about their data and privacy
   2. Users who are interested in similar projects

How to formulate the feedback loop of objective modeling?

Say in feteching relevant objectives, what if the user knows a better candidate than the system, how do we capture such feedback as input?

Similar Projects Analysis

Protolyst

The key value provided by Protolyst is making the evolution of each idea or claim back-trackable.

Go To Market

Youtube Influencers

1. Matthew Berman
2. IndyDevDan