vault backup: 2024-03-12 - 3 files

Affected files:
Monthly Notes/Feb 2024 notes.md
Monthly Notes/Mar 2024 notes.md
Resources/AI-hackathon-stack.md

Author: swyx

Monthly Notes/Feb 2024 notes.md

@@ -46,6 +46,7 @@
 		- [fixed with reason](https://twitter.com/E0M/status/1760476148763644166)
 	- [due to lazy/extensive prompt?](https://twitter.com/dylan522p/status/1755086111397863777)
 - Misc
+	- [OpenAI Forum](https://twitter.com/officiallogank/status/1758520604754333867?s=12&t=90xQ8sGy63D2OtiaoGJuww) - a place to discuss, learn, and shape AI. The forum features online and in-person events along with paid activities that directly impact OpenAI models.
 	- misc: [OpenAI Feather](https://www.semafor.com/article/01/27/2023/openai-has-hired-an-army-of-contractors-to-make-basic-coding-obsolete) - probably just inhouse scale ai [per patent](https://x.com/ugo_alves/status/1761857718812315838?s=46&t=90xQ8sGy63D2OtiaoGJuww)
 	- shut down [State-affiliated Threat Actors](https://openai.com/blog/disrupting-malicious-uses-of-ai-by-state-affiliated-threat-actors)
 	- [Logan on Lenny's pod](https://www.lennyspodcast.com/inside-openai-logan-kilpatrick-head-of-developer-relations/) 
@@ -71,6 +72,7 @@
 		- **Gemini 1.5 Pro comes with a standard 128,000 token context window. But starting today, a limited group of developers and enterprise customers can try it with a context window of up to 1 million tokens via [AI Studio](https://aistudio.google.com/) and [Vertex AI](https://cloud.google.com/vertex-ai) in private preview.**
 		- Through a series of machine learning innovations, we’ve increased 1.5 Pro’s context window capacity far beyond the original 32,000 tokens for Gemini 1.0. We can now run up to 1 million tokens in production.
 		- when tested on long code prompts, [HVM author agrees gemini retrieval is better than GPT4](https://old.reddit.com/r/singularity/comments/1atjz9v/ive_put_a_complex_codebase_into_a_single/)
+		- very good audio and video recall: https://twitter.com/mattshumer_/status/1758526890174751160?s=12&t=90xQ8sGy63D2OtiaoGJuww
 	-  Sparse MoE multimodal model - Can handle 3 hours of video, 22 hours of audio or 10M tokens with almost perfect recall! - Better than Gemini 1 Pro, better than Ultra for text, worse for audio and vision - Sadly not much info regarding model size, # of experts, architecture explorations, etc
 	- useful for [scanning codebase and implementing fixes for](https://x.com/sullyomarr/status/1760066335898513655?s=46&t=90xQ8sGy63D2OtiaoGJuww) and 
 	- [Gemini rollout in Google Workspace (Gmail, docs, sheets)](https://blog.google/products/google-one/google-one-gemini-ai-gmail-docs-sheets/)
@@ -158,6 +160,16 @@
 - Google OSS: [Magika: AI powered fast and efficient file type identification](https://opensource.googleblog.com/2024/02/magika-ai-powered-fast-and-efficient-file-type-identification.html)
 	- Internally, Magika is used at scale to help improve Google users’ safety by routing Gmail, Drive, and Safe Browsing files to the proper security and content policy scanners. Looking at a weekly average of hundreds of billions of files reveals that Magika improves file type identification accuracy by 50% compared to our previous system that relied on handcrafted rules. In particular, this increase in accuracy allows us to scan 11% more files with our specialized malicious AI document scanners and reduce the number of unidentified files to 3%.
 	- I ran this on some web crawl data I have locally, so: all files you'd find on regular websites; HTML, CSS, JavaScript, fonts etc. It identified some simple HTML files (html, head, title, body, p tags and not much else) as "MS Visual Basic source (VBA)", "ASP source (code)", and "Generic text document" where the `file` utility correctly identified all such examples as "HTML document text". [HN](https://news.ycombinator.com/item?id=39391688)
+- [NotesGPT](https://github.com/Nutlope/notesGPT): generate action items from your notes in seconds. Powered by Convex, Together.ai, and Whisper.
+	- Convex for the database and cloud functions
+	- Next.js App Router for the framework
+	- Together Inference for the LLM (Mixtral)
+	- Together Embeddings for the embeddings for search
+	- Convex File Storage for storing voice notes
+	- Convex Vector search for vector search
+	- Replicate for Whisper transcriptions
+	- Clerk for user authentication
+	- Tailwind CSS for styling
 
 ## product launches
 

Monthly Notes/Mar 2024 notes.md

@@ -1,6 +1,7 @@
 
 ## openai
 
+- jan leike open sourced [transformer debugger](https://twitter.com/janleike/status/1767347608065106387) - It combines both automated interpretability and sparse autoencoders, and it allows rapid exploration of models without writing code. It supports both neurons and attention heads. You can intervene on the forward pass by ablating individual neurons and see what changes. In short, it's a quick and easy way to discover circuits manually.
 - nontechnical
 	- [elon sues openai](https://www.washingtonpost.com/business/2024/03/01/musk-openai-lawsuit/), [openai responds](https://twitter.com/openai/status/1765201089366773913?t=6FDPaNxZcbSsELal6Sv7Ug)
 	- [openai board reappoints sama](https://news.ycombinator.com/item?id=39647105), [NYT wrote some stuff about Mira](https://archive.is/uroRV)

Resources/AI-hackathon-stack.md

@@ -105,6 +105,7 @@ Where the focus is building AI tools for other developers.
   - HouseGPT generates raw MIDI data directly from few-shot prompted GPT-3 to create 🎶 house music 🎶 🔊 ([tweet](https://twitter.com/russelljkaplan/status/1616997544307089408?s=20&t=frXEVPqaJUjMPJOhbD9AUg))
   - [Rap Battle](https://twitter.com/russelljkaplan/status/1617070021406265345?s=20&t=frXEVPqaJUjMPJOhbD9AUg) - Pick any two people and it will generate a rap battle on the fly, using GPT-3 for lyrics, wavenet for vocals, and stable diffusion for the avatars. 
   - Game of Life, where each alive cell is a whimsical happy Stable Diffusion image and each dead cell is an eerie, dark Stable Diffusion image, all of which evolve over time. ([tweet](https://twitter.com/russelljkaplan/status/1616955356189687810?s=20&t=KIszRKntkT4Y-I-WwKI8Mg))
+  - [NotesGPT](https://github.com/Nutlope/notesGPT): generate action items from your notes in seconds. Powered by Convex, Together.ai, and Whisper.
 
 ## AI Infra and Tooling