AI Discussions and Events

OpenAI Embeddings and Graph Embeddings #

• Discussion on utilizing OpenAI embeddings model to create graph embeddings
• Naive approach suggested: use “text-embedding-ada-002” to get node embeddings and compute edge embeddings using adjacency matrix
• Pinterest recommends starting with any base embedding and then updating them
• Article on HN shared for reference

GPT Model for Financial Data and Analysis #

• Query on availability of GPT model finetuned with financial data and analysis
• Mention of Bloomberg GPT, but not yet usable or publicly available
• Doubtful if it will ever be open sourced as it is trained on proprietary data

Generative AI Mixer Event #

• Invitation to a generative AI mixer event at Lightspeed office in Koramangala
• Interested participants asked to fill up a Google form

Deepfake Detection #

• Query on research happening in deepfake detection and any good app/software for it
• Mention of Spoofsense.ai working on it
• Microsoft’s effort on disinformation and deepfakes shared, but no working model for public
• Discussion on not releasing authenticators to avoid making it easier for next generation of cheaters to evolve/reverse engineer
• Mention of Yahoo’s nudity detection model from 7 years ago

API for Categorizing Questions #

• Query on API to categorize questions into HR, Finance, Marketing, etc.
• Offer to create one over the weekend if given 20 samples
• Discussion on using GPT4 in Google Sheets to classify questions

OpenAI Credits #

• Discussion on how to get more than $1K/month hard limit of OpenAI
• Mention of it being expensive, with 5K questions a day

Prompt Injection in GPT4 #

• Discussion on the status of prompt injection in GPT4
• Experimentation with prompt injection and simulation of another agent with different ideals in chatGPT and GPT 3, 3.5
• Mention of GPT4 being good with handling prompt injections and sticking to the system prompt

Miscellaneous #

• Mention of RISC-V International event in India
• Query on the largest OpenAI customer in India
• Sharing of article on Zero Knowledge Machine Learning and its use case in medical data inference
• Mention of HuggingFace launching a GPT4 alternative, Vicuna 13B

Links #

The description and link can be mismatched because of extraction errors.

• https://docs.google.com/forms/d/e/1FAIpQLSfRg1AOSSfSs8ZBbKzWd5ne_8JqzJrwFvy_NN0K-TUIdG2jYA/viewform - a form to be filled up by the recipients of the message, which is related to an AI event happening in Bangalore on April 14th.
• https://community.riscv.org/events/details/risc-v-international-risc-v-in-india-presents-nerds-talking-to-nerds-about-risc-v-hosted-by-tenstorrent/waitlist/151: The message in the same link as the URL asks if anyone is aware of research happening in Deepfake detection.
• https://blogs.microsoft.com/on-the-issues/2020/09/01/disinformation-deepfakes-newsguard-video-authenticator/ - Microsoft’s effort on disinformation and deepfakes, but doesn’t have a working model for public use. Mention of Spoofsense.ai working on the issue as well.
• Can use this https://kroop.ai/the-vizmantiz/ for cryptography, which is a constantly-evolving process/field. The message also asks if anyone has seen an API to categorize questions into HR, Finance, Marketing etc.
• https://www.buildt.ai/blog/incorrectusage - The blog post discusses incorrect usage of AI and mentions Pepper as one of the larger customers of OpenAI in India.
• https://analyticsindiamag.com/hugging-face-launches-gpt-4-alternative-vicuna-13b/ - A message discussing the potential of the newly launched Vicuna 13B as a knowledge distillation tool with human input for higher quality labels.