1 of 9

Intelligent Agents 2024

Microsoft BuildAI Day 2024 - Garry Tiscovschi

2 of 9

3 of 9

Field Examples: Constriction, Focus, Auto Regression

Tree of Optimal or ‘Correct’ Outputs

Tree of all possible token sequences

  1. Figure: (2023) Yann LeCun: https://drive.google.com/file/d/1wzHohvoSgKGZvzOWqZybjm4M4veKR6t3/view

Each chain of tokens stochastically generated influences the next. - many different paths can be generated. - Not all are desirable.

​

​

​

Hallucinated Competitor Offerings

Invented Policy cocktails

Not helpful answers

Focused & informed answers

So how do we push outputs into the optimal tree?

4 of 9

5 of 9

6 of 9

Example GenAI Impact

2.5 hr /section→ 15 min/ section

​

40 min → 10 min

​

40% Less Regulatory Enquiries

7 of 9

Data

  • Multiple agents passing critical data to each other.�
  • Data Sovereignty/Residency - GDPR / DPA�
  • Azure is a big help here - Hosting flagship models directly within the cloud tenant

8 of 9

Grounding & Orchestration

  • Multi Agent Chains
    • Using more efficient models to distil down context
  • Enhanced RAG (Vector DB)
  • User Intent Extraction
  • Pre-Tagging, Filtering:
    • Matching Intent to tagged data.
    • Using intent to filter away unnecessary context.
  • API Calls/ Browser Agent
  • Adding Metadata and Weights to stored context.
  • Follow up questions to user.
  • Generating SQL
  • API Calls - Toolformer Agent

​

9 of 9

Thank you!

Microsoft BuildAI Day 2024- Garry Tiscovschi