1 of 20

INTELLIGENT SOFTWARE FOR FUTURE SPACE

2 of 20

WHOAMI

  • Software Engineer at Siemens
    • Focusing Electronic Design Automation and Artificial Intelligence
  • Masters from Dalhousie University
    • Focusing on understanding software bugs using neural language modeling
  • Public Speaker
    • Focusing software engineering and artificial intelligence
  • Hiker
    • Focusing uphill
  • Biker
    • Focusing downhill

2

3 of 20

A LITTLE HISTORY LESSON

4 of 20

THE FIRST TRANSFORMER

4

5 of 20

DECODING

5

<start>

<start>

One

<start>

One

Ring

<start>

One

Ring

To

<start>

One

Ring

To

Rule

<start>

One

Ring

To

Rule

Them

<start>

One

Ring

To

Rule

Them

All

<start>

One

Ring

To

Rule

Them

All

<end>

6 of 20

BEAM SEARCH

Image courtesy: vnopenai.github.io

6

7 of 20

WHERE IS IT HEADED?

8 of 20

IS IT COMING AFTER MY JOB?

Only if your job is not worth doing.

8

9 of 20

THEN IS IT JUST HYPE?

Yes, and No

9

10 of 20

HOW BETTER IT WILL GET?

My very personal opinion is it is near its plateau

11 of 20

TIME TO LEARN IT NOW

  • Satellite Image Analysis
  • Astronomical Data Processing
  • Space Weather Prediction
  • Educational Tools
  • Spacecraft Design & Maintenance
  • Mission Planning and Simulation
  • Communication Systems
  • Resource Management

12 of 20

MAKE IT PART OF YOUR APP

13 of 20

RETRIEVAL AUGMENTED GENERATION

  • Trained models are outdated too soon
  • Retraining is costly
  • After training, hard to maintain privacy
  • Retrieval Augmented Generation (RAG) for rescue

13

14 of 20

AGENTS

  • An AI agent is a system or program designed to autonomously perform tasks on behalf of a user or another system.
  • From post processing response to driving your car

14

15 of 20

TEMPERATURE 🥵

HIGH TEMPERATURE (>1)

Flattens the distribution, making less likely tokens more probable.

Sharpens the distribution, making the model more likely to pick the most probable token.

LOW TEMPERATURE (<1)

15

16 of 20

TOKENS TO CONSIDER

Number of tokens to consider

TOP K

TOP P

Total probability to consider

16

17 of 20

PLAYING WITH TEMPERATURE

17

Top P = 0.5

18 of 20

PREVENT REPETITION

PRESENCE PENALTY

Penalize the model if a token is reused

Increases the penalty with each reuse

FREQUENCY PENALTY

18

19 of 20

AND A LOT MORE TO EXPLORE

19

Audio

Image

Code Execution

Function Calling

System Instruction

Yours to Build

20 of 20

THANK YOU

Questions?