2023 LLMBC App in Hour

LLMBC 2023
LOGISTICS
Join the Discord!
https://fsdl.me/llmbc-discord
Slides will be shared.
Talks will be recorded

and posted online.
WiFi: llmbc-2023
password: SolidGoldMagikarp
1
Tensor Tier Sponsors
Matrix Tier Sponsors
Vector Tier Sponsors
Compute Credit Donors

LLM Bootcamp
2023
Launch an LLM App

in an Hour
Charles Frye
APRIL 21, 2023

LLMBC 2023
LLMBC 2023
Instructor Team
@charles_irl @sergey_karayev @josh_tobin_
Berkeley PhD alums with years of experience

building and teaching about ML-powered apps. 5
LLMBC 2023
Agenda
00 01 02 03
PROTOTYPING DEPLOYING NEXT STEPS
INTRODUCTION
& ITERATION AN MVP
Move fast No one can Why are there 10

Why are we here?
and write prompts. stop you. more hours of talks?
6
00
Introduction
LLMBC 2023
Imagine a world…
• Disturbingly simple and heuristic
computer programs are used to
mimic human cognitive processes.
• Humans are forming emotional

attachments to chatbot therapists.
• Computers can play chess, write

mathematical proofs, and even pass
some exams.
• You can even get investment advice

from an arti cial intelligence!
8
fi
LLMBC 2023
It happened in 1965.
9
LLMBC 2023
What’s di erent this time?
One tool can do it all:

a large language model.
ff
LLMBC 2023
Language Models model language.

Imagine a paragraph appearing word by word.
If you learned to guess the next word,

you’d get pretty smart.
Especially if you could predict things like:
2+2=4
def square(x): return x * x
The chemical formula for glucose is C₆H₁₂O₆
Large Language Models are really good at it. 11

LLMBC 2023
LLMs unlock LUIs: Language User Interfaces.
ELIZA, 1966 SHRDLU, 1970 Ask Jeeves, 1996
https://en.wikipedia.org/wiki/Natural-language_user_interface 12
LLMBC 2023
Language modeling is considered AI-Complete*.
*this concept is slipperier than NP-Complete etc.
https://www.researchgate.net/publication/251422306_AI-Complete_AI-Hard_or_AI-
Easy_Classi cation_of_Problems_in_Arti cial_Intelligence 13
fi
fi
LLMBC 2023
What happened last time?

LLMBC 2023
Over-selling and under-delivering led to an “AI Winter”.
https://www.cambridgewireless.co.uk/media/uploads/resources/
AI%20Group/AIMobility-11.05.17-Cambridge_Consultants-Monty_Barlow.pdf 15
LLMBC 2023
We failed to deliver products to match high expectations.

1972
“high hopes that are very far
from having been realized”
“[N]otorious disappointments in
the area of machine translation,
where enormous sums have been
spent with very little result”
“Suggestions from recent research …

augur badly for the future availability of
machine-translation programs versatile
enough to be commercially valuable.”
http://www.chilton-computing.org.uk/inf/literature/
reports/lighthill_report/p001.htm 16
LLMBC 2023
To avoid AI Winter, we need to

build products that people value.
LLMBC 2023
Good news: a lot more than just research!

1972
https://twitter.com/natfriedman/status/ 18
1646285680941862912
LLMBC 2023
Bad news: the gap from demo to product is big.
CES, 2017
https://www.youtube.com/watch?v=fmVWLr0X1Sk 19
LLMBC 2023
Like, REALLY big.
NeurIPS, 1988 20
LLMBC 2023
Good news: we’re building products ( nally!)

1972
https://www.reuters.com/technology/chatgpt-sets-record-
fastest-growing-user-base-analyst-note-2023-02-01/ 21
fi
LLMBC 2023
So let’s build an app!

01
Prototyping & Iteration:

The First Half Hour
LLMBC 2023
• “Foundation models” have

unblocked a lot of applications
• Prototype with a high-capability

Takeaways hosted model in chat rst
• Tinker with prompts and

build on open source frameworks
fi
02
Deploying an MVP:
The Second Half Hour
LLMBC 2023
• Cloud tooling makes it easy

to get started
Takeaways - But make sure to limit spend!
• Find a simple UI and start getting

feedback from users ASAP
LLMBC 2023
Let’s build a bot

to do sourced Q&A
on the FSDL Discord!
LLMBC 2023
LLM
ETL
Serverless
Backend
Data Vector
Storage Index Discord
Bot Server
User 28
LLMBC 2023
LLM
ETL
Serverless
Backend
Data Vector
Storage Index Discord
Bot Server
User 29
LLMBC 2023
Try it out!
#ask-fsdl-bot
in the Discord.
We’ll analyze its behavior

later today!
03
Next Steps:
The Rest of Time
LLMBC 2023
Doing it is easy,
knowing what you’re doing is hard.
LLMBC 2023
This morning: foundations.
https://fullstackdeeplearning.com/llm-bootcamp/#schedule 33
LLMBC 2023
This afternoon: core LM skills.
LLMBC 2023
Late afternoon: repo and invited talks.
LLMBC 2023
Tomorrow morning: user experience.
LLMBC 2023
Tomorrow afternoon: LLMOps and the future.
LLMBC 2023
The end: invited talks and a panel.
LLMBC 2023
Thanks!
/imagine green parrot perched on a

laptop, coder, #developer, pancake
@charles_irl @full_stack_dl logo, pixar style, big round eyes,

dorky glasses, 3d animation, Blender,
4k render, OpenGL - shaders, ray
tracing

2023 LLMBC App in Hour

Uploaded by

Copyright:

Available Formats

2023 LLMBC App in Hour

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

2023 LLMBC App in Hour

Uploaded by

Copyright:

Available Formats

LLMBC 2023

Slides will be shared.

Talks will be recorded

Matrix Tier Sponsors

Vector Tier Sponsors

Compute Credit Donors

Launch an LLM App

APRIL 21, 2023

@charles_irl @sergey_karayev @josh_tobin_

Berkeley PhD alums with years of experience

Move fast No one can Why are there 10

• Humans are forming emotional

• Computers can play chess, write

• You can even get investment advice

What’s di erent this time?

One tool can do it all:

Language Models model language.

If you learned to guess the next word,

Especially if you could predict things like:

def square(x): return x * x

The chemical formula for glucose is C₆H₁₂O₆

Large Language Models are really good at it. 11

LLMs unlock LUIs: Language User Interfaces.

ELIZA, 1966 SHRDLU, 1970 Ask Jeeves, 1996

Language modeling is considered AI-Complete*.

*this concept is slipperier than NP-Complete etc.

What happened last time?

Over-selling and under-delivering led to an “AI Winter”.

We failed to deliver products to match high expectations.

“Suggestions from recent research …

To avoid AI Winter, we need to

Good news: a lot more than just research!

Bad news: the gap from demo to product is big.

Like, REALLY big.

Good news: we’re building products ( nally!)

So let’s build an app!

Prototyping & Iteration:

• “Foundation models” have

• Prototype with a high-capability

• Tinker with prompts and

• Cloud tooling makes it easy

Takeaways - But make sure to limit spend!

• Find a simple UI and start getting

Let’s build a bot

We’ll analyze its behavior

This morning: foundations.

This afternoon: core LM skills.

Late afternoon: repo and invited talks.

Tomorrow morning: user experience.

Tomorrow afternoon: LLMOps and the future.

The end: invited talks and a panel.

/imagine green parrot perched on a

@charles_irl @full_stack_dl logo, pixar style, big round eyes,

You might also like