OpenAI

If you (can) example it,

they will come

(& eat your lunch)

T&Cs

  • The demos you are about to see have not yet (nor will they ever...) been approved for launch by OpenAI.
  • This session must not be recorded.
  • Everyone watching this stream works for the same organisation.

The

OpenAI API

is a service provided by

OPENAI

which is powered by A MODEL CALLED

GPT3

which was fine-tuned to produce

CODEX

which is the model powering

Copilot

which is a service provided by

GitHub

which is owned by

Microsoft

who are significant investors in

OpenAI

The

OpenAI API

is a service provided by

OPENAI

which is powered by A MODEL CALLED

GPT3

which was fine-tuned to produce

CODEX

which is the model powering

Copilot

which is a service provided by

GitHub

which is owned by

Microsoft

who are significant investors in

OpenAI

AND ARE GROWING
INTENTIONS TO ACQUIRE THEM

  1. Background to using OpenAI
     

  2. Obligatory demos
     

  3. Building a terrible, but revealing, version of Copilot

TOKENS

OpenAI API Tokeniser

OpenAI API Tokeniser

Demo

TRANSFORMERS

tokenise("Once upon a")
tokenise(" time")
tokenise("Once upon a" + " time")
tokenise(",")

Response length

Run for n cycles

Stop Sequences

Run until the last n tokens match one of the stop sequences

DEMO

AI Language Models & Transformers - Computerphile

Controlling creativity

tokenise("Once upon a")
tokenise("time")
tokenise("Once upon a")
(tokenise(" time"), logprob("96.45%"))
tokenise("Once upon a")
[(tokenise(" time"), logprob("96.45%")),
(tokenise(" Time"), logprob("0.67%")),
(tokenise(" midnight"), logprob("0.31%")), ...]

FREQUENCY PENALTY

The extent to which each repeated occurrence of a token, lowers the future probability of that token.

 

 

PRESENSE PENALTY

The extent to which any previous occurrence of a token, lowers the future probability of that token.

Demo

TEMPERATURE

The likelihood that a lower probability token will have its relative probability increased.

Top P

Only consider the top n most likely tokens, having a cumulative probability of p.

0 = The best guess

0.5 = A good guess

1 = Hold my beer...


N.B. It only makes sense to use one of these, the other should be set to 1.

Demo

Best of

Generate m sequences of n tokens, then select the sequence of tokens with the highest probability.

 

Expensive...

Inject start text

Before submission, concatenate this string onto the prompt.

 

E.g. "\nA:"

Inject restart text

After receiving a response, concatenate this string onto the completion.

 

E.g. "\n\nQ:"

No Demo

(It's a bit dull)

One SHOT
FEW SHOT
FINE TUNE

One Shot

  • E.g. "Once upon a"
  • Unambiguous high probability intent

Few Shot

  • E.g. "A: Apple, B: Banana, C:" versus "A: Apple, B: Broadcom, C:"
  • Select from highly probable intents

Fine Tune

  • E.g. Copilot*
  • Requires extensive, accurate and non-biased datasets of examples

CheapPilot

The Challenge

~ 25 seconds

~ 50 seconds

~ 15 seconds

~ 15 seconds

Demo

Conclusions

  • Applying AI to business problems is becoming more accessible

  • AI is going to change the way we work as developers

  • It might not be be revolutionary but things will change

Conclusions

  • Applying AI to business problems is becoming more accessible

  • AI is going to change the way we work as developers

  • It might not be be revolutionary but things will change

  • When should you start to look into it? Tomorrow. It would be foolish not too