Applied Language Understanding

@simonw @awnihannun More precisely, it implements an efficient LLM token selector driven by a JSON schema, with application examples on MLX. This is better that GNBF in that mapping to GNBF can produce unwieldy rule sets for things like min and max number of items in an array . @awnihannun @simonw

@simonw @awnihannun Check out this repo Simon, it implements json schema on mlx github.com/otriscon/llm-s…

Made a minimal example of tool calling with an LLM in mlx-lm. Works quite well with Qwen 2.5:

@thesephist For Mac (MLX) there’s github.com/otriscon/llm-s… which works with any model.

Is there a way (local or hosted) to do structured JSON Schema sampling from a non-instruct base model that's Llama 3 8B quality or higher?

✅ Against GPT-3.5-Turbo, we saw a 17% reduction in hallucination and similarly improved recall, at a 10x cost and 2x latency. blog.alu.ai/evaluating-gpt…

We evaluated GPT-4-Turbo in our pipeline and found it a very useful addition to the palette of LLMs available, although not a complete game-changer. ✅ Against GPT-4, we saw 66% reduction in latency and cost with an estimated 5% decrease in quality (F1 score). (1/2)

More... blog.alu.ai/fine-tuning/

From prompt whack-a-mole to fine tuning. We've learned a couple of lessons fine-tuning GPT-3.5.

Of course, to write the story for the submission, we enlisted AI help. blog.alu.ai/participating-…

For Atlassian’s codegeist hackathon, we sprinted on querying Confluence tables:

What’s the point of AI if you can’t rely on its output? We are building a trustworthy assistant for Confluence that tells you exactly how it got its answers. Not just sources, but every reasoning step. Read more in our blog: blog.alu.ai/introducing-x-…

Atlassian has just released the public beta of Atlassian Intelligence. We had a first look and found it still has some way to go before being trustworthy. Link to our report below.

Read about it in our blog: blog.alu.ai/tables-and-str…

Connie AI can now answer sophisticated questions about tabular data. We translate natural language queries to a pseudo-SQL language and execute it over your Confluence tables. Support for the new Confluence Databases coming soon.

New post in our blog: streaming OpenAI function calls. blog.alu.ai/streaming-open…

@typesfast Connie AI (connie.alu.ai). We’re ex-Atlassians building an AI assistant for Confluence. It’s miles ahead of Confluence search already.

Confluence is a corporate wiki without a working search function. What is the point of a wiki with no search? What are better alternatives out there?

Our internal evaluation tool, Gaucho, allows us to keep tabs on the quality of our system: retrieval, LLM prompts, etc. We wrote about it in the blog: blog.alu.ai/gaucho-our-eva…

If your team uses Confluence, chances are it suffers from this common condition 🤧😅 confluenzza.com

A bit of fun with Confluence, hope our Atlassian friends don't mind: confluenzza.com

A short demo of Connie AI: youtube.com/watch?v=QY_wNR…