🤔 Can you depend on the answers AI gives?

“I have a bad feeling about this.” ― Luke Skywalker , Star Wars: Episode IV

LLMs lie about the facts all the time. But Google, decided to add it to search anyway. We’ll look at why someone should’ve had a bad feeling about that.

In today’s newsletter:

• Can you depend on the answers AI gives?

• Startups

• Papers

• Learn

Penses

LLMs are the current tech hype because they are very easy for a non-techy consumer to use. But that doesn’t free it up from having tons of pitfalls.

Google released a new feature called AI Overviews, and here’s some of the more comically entries that were shared by the people across the web

If you’re brand is a entrenched in everyday life as Google, maybe you don’t need to worry about it as much. But some businesses this could make the brand take a major hit.

Three ways to mitigate the effects of hallucinations. .

1. Don’t rely on external data. Instead of building a Large Language Model(LLM), you can build a Small Language Model ( SLM) tailored to your information.

2. There’s also a method of data augmentation called Retrieval Augmented Generation (RAG) . RAG is basically attaching a vector database of your information to a LLM to help it retrieve a better set of answers.

3. Fine-tuning the model is a bit more time consuming. It requires additional prompt/answer pairs to train the model on more information. It provides more control of the potential answers for common questions.

Startups

Vector Databases received a lot of funding as a result of the LLM boom. Each has a slightly different approach to vector database configuration. Here’s an interesting benchmark article on the topic.

Here’s a few you should probably watch:

Qdrant ($28M) - Open sourced option

Pinecone ($138M ) - Managed database solution with free version

Chroma ($18M) - Open source option

Weaviate ($66M) - Open source option

Papers

Here’s some of the insights I gathered from papers today:

• This paper has a very large graph of papers on the causes, risks, and mitigation of hallucinations. Common causes of hallucinations are architecture, attention glitches, samples and the list goes on . Read more here.

• LLMs are always going to hallucinate according to these researchers. Mostly because of time complexity. They get into a lot more of the math and theory here.

Learn

A intro to vector databases

A different take on vector databases