Slide 1

Slide 1 text

/ AWS Startup Lofts Anthropic keynote Maggie Vo Head of Technical Education & Enablement Alex Albert Head of Developer Relations

Slide 2

Slide 2 text

Frontier research that powers frontier models 2021 2023 2022 2024 Anthropic founded to prioritize AI safety at the frontier Claude 2 Claude 2.1 Claude 3 Claude 1 Frontier Research Mathematical Framework for Transformer Circuits Frontier Research Toy Models of Superposition (Interpretability) Frontier Policy Constitutional AI Frontier Research Dictionary Learning (Interpretability) Frontier Research Sleeper Agents: Deceptive LLMs Frontier Research Mapping The Mind of an LLM

Slide 3

Slide 3 text

Claude 3

Slide 4

Slide 4 text

Foundation models crafted for scaled applications Foundation models Claude 3 model family Claude 3 Opus Claude 3 Sonnet Claude 3 Haiku Top-tier model intelligence Balance of skill and speed Fastest, most compact Get started with Claude 3 models on: Anthropic API AWS Bedrock Powerful alone. Better together.

Slide 5

Slide 5 text

Analyze mountains of text Create content with depth of knowledge Automate workflows Documents, emails, transcripts, FAQs, records, code databases Summarize channels, function-calling and tool use, prioritize & assign action items Product support, business content, general advice “oracles” Core Capabilities of Claude

Slide 6

Slide 6 text

Improvements from previous Claude generations Faster More steerable Vision Faster models available in each intelligence class Better results out-of-the-box with less prompt optimization and fewer refusals The fastest vision model with comparable quality to other state-of-the-art models More accurate & trustworthy Twice as accurate as Claude 2.1 on difficult, open-ended questions

Slide 7

Slide 7 text


Slide 8

Slide 8 text

Fast & capable vision, trained for business use cases ● Understands enterprise content including charts, graphs, technical diagrams, reports, and more ● Faster than other multimodal models while achieving similar performance 1 ● Excels at use cases that require speed & intelligence • Extract data from documents, charts, graphs, … • Analyzing images for insurance claims, adjustments, … • Transcribe handwritten notes, diagrams, … • Generate product information & insights from images 1-Based on internal evaluations for Claude 3 Haiku. Summarize this report Recreate this graph in Python Describe the condition of this vehicle What’s the condition of this package?

Slide 9

Slide 9 text

Tips for prompting with images ● Put images before the task, instructions, or user query where feasible ● When you have multiple images, label each image, like “Image 1:” and “Image 2:” ● Increase performance by having Claude describe and extract details from the image(s) before doing the task User Image 1: [Image 1] Image 2: [Image 2] How are these images different? Claude response [Claude's response] User Image 3: [Image 3] Image 4: [Image 4] Are these images similar to the first two? Claude response [Claude's response] Example conversation:

Slide 10

Slide 10 text

No content

Slide 11

Slide 11 text

“Transcribe this whiteboard in JSON. Identify the main entities, attributes, or categories and use them as keys in the JSON object. Then, extract the relevant information from the text and populate the corresponding values in the JSON object.”

Slide 12

Slide 12 text

No content

Slide 13

Slide 13 text

“Write unit tests for these functions”

Slide 14

Slide 14 text

No content

Slide 15

Slide 15 text

Automated product metadata and comparisons “I am going to give you a schema describing a product and an image of that product, I would like you to suggest amendments by returning a corrected version of the schema, with inline comments on any fields you amended. ``` { “Color”: “Black”, “Maker”: “Nike”, “Purpose”: “weighlifting”, “Ideal_distance”: “None”, } ```”

Slide 16

Slide 16 text

Automated product metadata and comparisons

Slide 17

Slide 17 text

Automated product metadata and comparisons “Write out a much more detailed schema with all the attributes you can confidently infer from the image”

Slide 18

Slide 18 text

Automated product metadata and comparisons “Write out a much more detailed schema with all the attributes you can confidently infer from the image”

Slide 19

Slide 19 text

Agents + tool use

Slide 20

Slide 20 text

Claude 3 has improved agentic capabilities “I need to be reimbursed for my blood pressure medicine” User Request Determine Goal(s) Complete Tasks Take Action ● Determine if user should be reimbursed ● If yes, send funds ● If no, politely explain reason ● Pull customer record ● Pull reimbursement policy ● Run drug interaction safety check ● Escalate to human if concerned or unsure ● Write draft copy ● Review answer ● Affirmative chat response ● Reimbursement initiated

Slide 21

Slide 21 text

How do you create an agent? With tool use! User prompt What was the final score of the Yomiuri Giants’ game on June 16, 2024? List of tools name: get_score description: Get the score of a baseball game required parameters: team, date name: top_song description: Get the most popular song played on a radio station required parameters: title, artist Chosen tool & inputs name: get_score parameters: team=Yomiuri Giants date=2024-06-16 Tool use is a way to expand Claude’s abilities with external tools & functions

Slide 22

Slide 22 text


Slide 23

Slide 23 text

No content

Slide 24

Slide 24 text

No content

Slide 25

Slide 25 text

No content

Slide 26

Slide 26 text

Web browsing + multimodal + interactive coding (REPL) + subagents

Slide 27

Slide 27 text

Tool use resources Tool use documentation (Anthropic official) Tool use & prompt engineering course (Anthropic official on AWS)

Slide 28

Slide 28 text

Useful resources

Slide 29

Slide 29 text

Claude prompt generator Claude can write prompt templates on the your behalf, given a topic or task details

Slide 30

Slide 30 text

Claude prompt generator on AWS

Slide 31

Slide 31 text

Claude prompt generator - considerations ● The generated prompt is meant as a starting point to solve the “blank page” issue by outputting a well performing, decently engineered prompt ● The prompt generating tool does not guarantee that the prompt it creates will be 100% optimized or ideal for your use case - run more tests & continue refining! If you’re curious about the underlying structure of the prompt generator, you can read more about it in our documentation at under the prompt engineering section

Slide 32

Slide 32 text Anthropic prompt library

Slide 33

Slide 33 text

General resources ● Anthropic’s Claude 3 model card ( ○ Detailed technical information about evaluations, model capabilities, safety training, and more ● Anthropic cookbook ( ○ Code & implementation examples for a variety of capabilities, use cases, integrations, and architectures ● Claude user guide documentation ( ○ Prompt engineering tips, production guides, vision guide, model comparison tables, capabilities overviews, and more

Slide 34

Slide 34 text

Thank you