Join our webinar to explore the power of our AI Engineer! Learn how to build an AI Brain for your org, connect to all your systems, automate pipelines, and work with all kinds of data.
RSVP today!
You can transform a simple PDF earnings report into an interactive dashboard!
Play around with different LLMs to see which works but we found that LLama-400B and Sonnet 3.5 does this task the best.
LLMs now have step-by-step reasoning! Smaller models like Llama-70B and Smaug can tackle complex problems by breaking them down. This is only the beginning – they will be able to handle even more complicated tasks with ease!
Architecture For An AI Brain for Organizations
A real-time permission-aware index that connects to all your services
An AI brain that can access the index and can build agents, chatbots and models
Gemini's Real Superpower - It's 10x Cheaper Than o1!
The new Gemini is live on ChatLLM teams if you want to play with it.
Predictably, the new version is better than the old version, and Gemini now trails behind o1, Sonnet, and gpt-4o but is significantly better than
This survey paper (link provided in alt/image description) is a great read on GraphRAG, improving standard RAG systems' accuracy.
Graph RAG uses relationships between enttiies to improve information retrieval, leading to more accurate and context-aware responses!
Realistically AI does not contribute more than 10% to a code base of medium complexity today
By the end of next year it will be almost 30%
That’s a very big leap forward! 🚀
Autonomous agents are automatically performing these tasks:
•Reading and responding to email
•Generating reports and posting them on Slack
•Generating legal and HR documents
•Basic code reviews
•Tackling tech and support desk requests
AI is beginning to do more and more
To CoT or Not CoT?
Great paper that shows chain-of-thought mainly helps math and symbolic reasoning
It only has marginal benefit on other tasks
On MMLU, directly generating the answer without CoT leads to almost identical accuracy as CoT unless the question or model's
Today’s AI is missing thinking at different layers of abstraction and memory
It’s insanely hard to actually use AI to fix bugs in a complex code base
Good human SWEs can do it with ease
LLMs can’t explore, think, and remember at different levels of abstraction
AI can get
Qwen 2.5 72B Is The Best Open-Source Model In The World
The Qwen 2.5 models dropped today, and they have some excellent scores.
It beat Llama-405b on Livebench AI (August questions). We will soon release its performance on the September challenge.
With an excellent score of
With our Matrix Agent, you can perform complex analysis. Here’s what you can do:
•Portfolio analysis
•Stock analysis
•Customer and leads analysis
•Decision science
The best way to apply o1 models is through an LLM router that automatically routes your query based on task and difficulty level.
Here is the best LLM router that you can design
sonnet - writing, coding
gpt4 - web search, doc generation
gemini - video understanding
o1 - complex
Google has dropped this, and several CoT papers in the past, and this one speaks to o1's performance.
This proves that transformers can solve most problems by generating intermediate reasoning tokens during inference time.
It is remarkable that they do A++ research and are
LLMs can write code to automate business processes or create chatbots/AI agents, but how would you run this code?
Our AI engineer can host code, run pipelines, connect to hundreds of apps, and understand structured and unstructured data.