Generative AI Weekly - July 25th, 2024
New models, releases, and milestones. Commentary on the AI bubble and years ahead.
Quick news: LLMs are allowing us to think about solving problems using orders of magnitude more scale than we could have accomplished in the past. For solutions where the bottleneck is time over a repeated set of similar judgment cases - LLMs open up capabilities that did not feasibly exist before.
We've developed a little internal application to help solve two problems we see when dealing with these types of projects:
- Judging the efficacy of a prompt or model versus alternatives is difficult - we don't yet have many tools to rank and establish improvements for everyday use cases. We needed a way to quickly form opinions about these outputs with higher degrees of certainty.
- When dealing with situations where hundreds or thousands of opinions need to be made independently, LLMs can help provide a base opinion or hypothesis that a human can then accept or reject - versus needing to form a judgment and decision simultaneously.
If you would like to see a pilot demo - please reach out.
This week:
- Meta releases an open-source GPT-4 level model. Mark Zuckerberg releases an open-source manifesto.
- Google Deepmind achieves silver-medal standard solving International Mathematical Olympiad problems.
- Benedict Evans on "The AI summer." Comparisons to the dotcom bubble and some great insights on development and sales cycles that are relevant to the LLM bubble we find ourselves in.
- Bain's AI survey - show me the money!
- OpenAI announces SearchGPT.
- A nice primer on eight types of data analysis - useful when thinking through how you are working with LLMs for data interpretation.
- Gumloop - a favorite tool of ours - announces its seed round of funding.
- Wolfram LLM Benchmarking Project. "we've been continually tracking the performance of LLMs.."
- The three things LLMs actually do. Expansion - compression - translation.
- And for this week's paper - The Adoption of ChatGPT.