Generally Intelligent

Generally Intelligent provides concise updates on AI advancements, focusing on large language models (LLMs), their evaluation, competitive landscape, and usage. It discusses OpenAI's developments, the rise of competitors, fine-tuning techniques, and the implications of the AI talent market on technology deployment and business strategies.

AI Industry Updates LLM Evaluation and Usage AI Model Development and Fine-Tuning Competitive Landscape in AI AI Talent and Market Trends

The hottest Substack posts of Generally Intelligent

And their main takeaways

OpenAI's API Problem

157 implied HN points • 25 Jul 23
  1. OpenAI extended support for older model versions due to user feedback
  2. LLM endpoints are underdocumented APIs, making upgrades challenging
  3. Migrating to new API endpoints without proper documentation can cause issues

LLM Evaluation Primer

137 implied HN points • 01 Aug 23
  1. Evaluation is a major challenge for teams using LLM-based products due to the complex input space and unstructured output of LLMs.
  2. When building an LLM application, key design variables to consider are the prompt, model, and information retrieval strategy.
  3. Teams use four main approaches to evaluate LLM applications: offline human evaluation, offline deterministic evaluation, offline model-driven evaluation, and online evaluation.

6 Ways to Choose a Language Model

98 implied HN points • 18 Jul 23
  1. Most common way to use language models is through large model providers like OpenAI.
  2. Fine-tuning models using OpenAI's endpoint has decreased popularity due to lack of support for GPT-4.
  3. Consider trying out open-source language models like LLaMA 2 or fine-tuning an open-source model for your specific task.

Claude 2, and the LLM Competitive Landscape

78 implied HN points • 14 Jul 23
  1. Claude 2 is a strong competitor to GPT-4, offering similar capabilities at a cheaper price.
  2. When choosing a language model, consider factors like model size, cost, and use-case dependent needs.
  3. Besides performance, factors like steerability, compliance, security, and privacy are important considerations in selecting a model.

MosaicML, and the cost of LLM Talent

39 implied HN points • 27 Jun 23
  1. Databricks acquired MosaicML for $1.3B, highlighting the high cost of training language models.
  2. Training language models well requires scarce and expensive talent, with estimates of median compensation packages for software engineers at around $900k.
  3. The real value in the acquisition of MosaicML lies in their talented ML engineers, showcasing the importance of investing in AI talent for successful business ventures.
Get a weekly roundup of the best Substack posts, by hacker news affinity:

Llama 2: Time to Fine Tune

39 implied HN points • 20 Jul 23
  1. Meta AI released Llama 2, a model comparable to GPT-3.5
  2. Fine-tuning Llama 2 could lead to more efficient models than GPT-4
  3. The rise of fine-tuning is expected with the open-source Llama 2 model

AMD is usable!

39 implied HN points • 10 Jul 23
  1. AMD GPUs are being used for modern LLM training
  2. AMD GPUs are slightly behind NVIDIA in performance
  3. Using AMD GPUs may help alleviate the current GPU shortage and reduce costs

Hearsay GPT-4 Architecture Details

19 implied HN points • 21 Jun 23
  1. GPT-4 architecture may consist of specialist models instead of one massive model
  2. Scaling up large language models (LLMs) is limited by the availability of high-quality training data
  3. Future of AI models may rely on domain-specific, smaller models to overcome data limitations and achieve higher quality

Let's Talk about LangChain

3 HN points • 12 Jul 23
  1. LangChain provides useful templates for developers working with language models
  2. LangChain struggles to provide static abstractions due to the rapidly changing AI landscape
  3. Developers may need more time to fully grasp and implement abstractions in the AI space

OpenAI API Updates

0 implied HN points • 14 Jun 23
  1. OpenAI added function calling capability to the Chat API.
  2. Introduced a 16k context-length version of `gpt-3.5-turbo`.
  3. Updates emphasize the importance of fine-tuning models for better performance.

Mistral, Inflection, and Challenging OpenAI

0 implied HN points • 23 Jun 23
  1. Mistral AI and Inflection AI are working on models to compete with OpenAI's GPT-4.
  2. Despite many recent model releases, OpenAI still dominates the field.
  3. There are key factors like cost, data, and model size that impact the ability of competitors to challenge OpenAI.