Snowflake Ventures is investing in @Omni Analytics to supercharge self-service BI and data modeling! Omni expands the AI Data Cloud’s analytics power, making it easier to extract value from data, faster.

We are excited to share that Snowflake has signed an agreement to acquire the TruEra AI Observability platform to bring LLM and ML Observability to its AI Data Cloud. We are looking forward to this next phase in our journey with the Snowflake team with whom we share a

On a tear at Snowflake ...
* Text analytics on your data: Cortex functions in GA ✅
* Better semantic search: Arctic embed in Cortex, vector functions in PuPr ✅
* Higher performing LLMs in Cortex: reka-core, Llama3, Arctic ✅
* Enhanced AI safety: LlamaGuard2 + Arctic in

Good read, highly recommended

We just released my interview with sridhar and Baris Gultekin on #SnowflakeArctic . Interesting to hear their perspective on the future of generative AI in the enterprise.

Congrats Sharon Zhou and team!

Academic benchmarks are losing their potency. Moving forward, there’re 3 types of LLM evaluations that matter:

1. Privately held test set but publicly reported scores, by a trusted 3rd party who doesn’t have their own LLM to promote. Scale AI’s latest GSM1k is a great example.

3rd week in a row, 3rd LLM from Snowflake ...

Arctic-TILT is a 800M model that has GPT-4 quality performance on information extraction tasks, as measured by the DocVQA benchmark.

And it fits in an A10!

Interesting, and very good work making new evals.

- I think we should look at responses for phi and mistral to see if its some other failure mode or plainly wrong.

Evals push the field forward.

🔥 another win for open source AI

SNOWFLAKE ARCTIC:

1] Mixture of Experts model with 128 (!) experts

2] open weights, open code

3] Company is sharing 'cookbook' and data and teaching people how to build their own world class MoE models.

4] very cheap model to train (16x

Interesting Wes Roth video on Snowflake's #SnowflakeArctic model exploring the implications of large, capable, and truly open models.

model = learn(data)

Synthetic data is great, but it’s not data. It’s an intermediate quantity created by learn(). Data is created by people and has privacy and copyright considerations. Synthetic “data” does not - it’s internal to learn().

Often hear companies that want alignment as a service.

In other words, configurable bar on what to refrain from answering.

For example, my chatbot should only answer questions relating to my domain and not general purpose questions like [who is the president of the United