🚀 VivaTech 2024 : Keynote - More than a Model: The Gen AI Essentials for Business Innovation

Summary

このビデオスクリプトでは、生成的AI（generative AI）の未来とその応用について語られています。スピーカーはAmazonでのインターンシップを通じてクラウドコンピューティングの基礎を築き、その後の20年間で技術業界で最も革新的な技術に関与してきたと語ります。生成的AIは医療、マーケティング、ゲームなど様々な業界で革新をもたらし、カスタマイズされた言語モデルやタスク自動化、セキュリティとガバナンスを備えたエンタープライズアプリケーションに組み込まれています。Amazon BedrockとAmazon Qは開発者とビジネスユーザーが生成的AIを活用し、業務の効率化と新しいアプリケーションの開発を促進するツールとして紹介されています。

Takeaways

• 🧑‍💼 スピーカーは、ジェネラティブAIを活用して新しい製品やサービスを創造することに情熱を燃やしているビルダーであり、夢を追う人々と共通しています。
• 🌐 クラウドコンピューティングの初期段階から、スピーカーはテクノロジー業界で最も革新的な技術に関与してきました。
• 📈 ジェネラティブAIの時代には、ビッグデータ、機械学習、そして大規模な言語モデルがプログラム可能となり、ビジネスのイノベーションを促進しています。
• 🛍️ マーケティング業界では、ジェネラティブAIが自動化された広告キャンペーンやハイパーパーソナライズドされた顧客体験を可能にしています。
• 🏥 医療業界では、ジェネラティブAIは医師を支援し、患者との会話を聞き取り、診療ノートを要約、画像診断方法を提案して診断の正確性を高めています。
• 🎮 ゲーム業界では、個々のプレイヤーの行動に基づいてキャラクター、ミッション、インタラクションの無限のバリエーションを持つ個々の体験を作成できます。
• 🔧 ジェネラティブAIは顧客体験の向上や従業員の生産性の向上を通じて、会話検索、テキスト要約、コード生成など、多くの業界で共通のテーマとユースケースを形成しています。
• 🛠️ モデルのカスタマイズ、タスクの自動化、セキュリティとガバナンスの強化が、ジェネラティブAIアプリケーションを構築する上で重要な要素です。
• 📚 データはビジネスや顧客を理解するための重要な差別化要因であり、RAG（Retrieve Augmented Generation）やファインチューニング、継続的プリトレーニングを通じてモデルにデータを活用できます。
• 🤖 開発者がAIモデルを利用し、タスクを自動化し、組織内の仕事を行うには、多くの手動プログラミングが必要なため、時間と専門知識を要するプロセスです。
• 🛡️ セキュリティとガバナンスは、ジェネラティブAIアプリケーションの展開において重要な課題であり、データプライバシーとエンタープライズグレードのAIの制御が必要です。

Q & A

ジェネラティブAIがどのような時代を迎えていると話されていますか？

ジェネラティブAIは、現在、すべての業種がその力を借りて自己再発明を行っている時代を迎えています。AIは経営会議での議論の中心となっており、マーケティング、ヘルスケア、ゲームなどさまざまな分野でその能力を発揮しています。

ジェネラティブAIがビジネスに与える主な利点は何ですか？

ジェネラティブAIはビジネスにとって、顧客体験の向上、従業員の生産性の向上、会話型検索、テキスト要約、コード生成支援、ビジネスオペレーションの最適化など、多くの利点をもたらします。

ジェネラティブAIを活用するためにはどのような技術が必要ですか？

ジェネラティブAIを活用するためには、大規模言語モデルや他の基礎モデルへのアクセス、モデルのカスタマイズ能力、タスク自動化能力、セキュリティとガバナンスの組み込みなどが必要です。

Amazon Bedrockとはどのようなサービスですか？

Amazon Bedrockは、ジェネラティブAIアプリケーションを構築し、拡大するためのフルマネージドサービスです。さまざまなAI企業から提供される基礎モデルへのアクセスを提供し、カスタムモデルのインポートも可能です。

モデルのカスタマイズに必要なデータ管理とはどのようなものですか？

モデルのカスタマイズには、ビジネスや顧客に関するデータが不可欠です。データは、汎用的なジェネラティブAIモデルとビジネスに適ったアプリケーションを区別する要素であり、データ管理はデータの統制と統合を意味します。

タスク自動化とは何を意味していますか？

タスク自動化とは、ビジネスや顧客のために日常業務を自動化する自律的なエージェントを作成することを意味します。これには、多くの手動プログラミングが必要な場合があり、開発者がこれらのタスクを実行するためのステップを定義し、ワークフローを調整する必要があります。

セキュリティとガバナンスがジェネラティブAIになぜ重要ですか？

セキュリティとガバナンスは、ジェネラティブAIにおいては、モデルの出力に対する信頼性と責任を確立するために不可欠です。これには、データプライバシーの保護、モデルの品質管理、人間による判断の重要性などが含まれます。

Amazon Qはどのようなサービスですか？

Amazon Qは、セキュリティとプライバシーを念頭に置いて開発された、最も高度なジェネラティブAIアシスタントです。開発者やビジネスユーザに効率をもたらし、タスクの自動化や情報へのアクセスを支援します。

mral AIはどのような企業ですか？

mral AIは、AI分野において新しいプレイヤーとして登場し、独自のアプローチを提供することを目指す企業です。フランスを拠点としており、効率と多言語のサポートに重点を置いています。

mral AIが今後どのような方向性を貫くんですか？

mral AIは、今後も多様なタスクに対応するためのモデルの多様化と最適化に注力し、さらにはコードの垂直化や多感覚モデルの開発にも取り組む予定です。

Outlines

00:00

🤖 ジェネラティブAIの未来とその影響

スピーカーはジェネラティブAIの未来について話しており、自分自身がクラウドコンピューティングの初期段階から関わっており、その歴史に貢献してきたと紹介。AIの進化を振り返り、大規模言語モデルがビジネスや医療、ゲームなど様々な業界に与える影響について触れている。特に、マーケティング業界では顧客体験の向上、医療業界では医師の支援、ゲーム業界では個々のプレイヤーに基づく体験のカスタマイズが可能になるという具体的な例を挙げている。また、AIの進化が企業のイノベーションと製品開発に与える影響についても語られている。

05:01

🔧 ジェネラティブAIを活用するための技術的基盤

AWSが提供するジェネラティブAIの基盤について説明しており、多様なファウンデーションモデルを提供することで、顧客がさまざまなユースケースに対応する柔軟性を持つことができると強調。開発者がAIモデルを利用しやすくするために、モデルのカスタマイズやタスクの自動化、セキュリティとガバナンスの確保が重要であると語られている。さらに、モデルの選択以外にも、データの整理や管理が急速にジェネラティブAIを導入している企業にとって重要であると示唆している。

10:03

🛠️ モデルのカスタマイズとタスクの自動化

ジェネラティブAIを活用する際のモデルのカスタマイズ方法とタスクの自動化について詳述。開発者が企業のデータを使用してモデルをカスタマイズする方法として、リトリーバル・オーギュメントド・ジェネレーション(RAG)やファインチューニング、継続的プリトレインメントという3つの技術を紹介。これにより、企業のビジネスや顧客を理解するモデルを構築することができると説明。また、タスクの自動化においては、開発者が手動でコーディングを行っているプロセスが煩雑であることと、自動化を活用するための課題についても触れている。

15:04

🛡️ セキュリティとガバナンスの重要性

ジェネラティブAIアプリケーションを展開する際のセキュリティとガバナンスの重要性を語っている。言語モデルのパラメータ調整とシステムプロンプトを使用して、モデル出力の誤解釈（hallucination）を管理できると強調。また、企業グレードのAIにはデータプライバシーの新しいコントロールが必要であり、企業が自社のデータを第三者へ送信することはリスクが高いと警告。AIシステムの行動を監視し、AIの悪用を防止し、ワークフォースの再教育プログラムを実施することが、顧客との信頼を構築し、AIを責任を持って構築するために必要であると語られている。

20:05

🌐 Amazon BedrockとQによる開発者の支援

Amazonが開発者を支援するサービスであるBedrockとQについて紹介。Bedrockはフルマネージドサービスで、開発者が自分のジェネラティブAIアプリケーションを構築・拡大しやすく作成されており、多数の顧客がすでにその基盤として利用していると報告。BedrockはリーディングAI企業からのファウンデーションモデルへのアクセスを提供し、カスタムモデルのインポートも可能としている。Qはセキュリティとプライバシーを念頭に置いて開発された、最も高度なジェネラティブAIアシスタントであり、開発者やビジネスユーザーの生産性を高めるためのツールとして機能している。

25:05

🌟 ジェネラティブAIの未来とmral AIの展望

mral AIの共同創設者であるTimothy Lacroixとのインタビューが行われており、mral AIのビジョンと将来性について語られている。mral AIは独自のアプローチでAI分野に参入し、効率性と多言語のサポートに重点を置いている。彼らのモデルはコスト効率に優れており、特にヨーロッパの言語においてはパフォーマンスが優れていると強調。mral AIは今後、多様なモデルの開発とカスタマイズの煩雑さを軽減し、より多くのユーザーがAIアプリケーションを構築できるようにすることを目指していると語られている。

Keywords

💡生成的AI（Generative AI）

生成的AIとは、既存のデータから新しいものを作成することができる人工知能の分野です。このビデオでは、様々な業界において生成的AIがどのように利用され、顧客体験の向上や業務効率化に寄与しているかが語られます。例えば、マーケティング業界では広告キャンペーンの自動化や、ゲーム業界では個々のプレイヤーの行動に基づくキャラクターやミッションの無限のバリエーションの作成などが挙げられます。

💡クラウドコンピューティング

クラウドコンピューティングとは、インターネットを通じてコンピュータリソースを提供するサービスのことです。ビデオでは話者がAmazonでインターンシップを経験し、クラウドコンピューティングサーバーの開発に携わったと語ります。これは、コンピュートストレージがプログラマブルなリソースになることで、技術業界における大きな変革をもたらしたと説明されています。

💡ディープラーニング（Deep Learning）

ディープラーニングは、人工知能の分野の一つで、多层のニューラルネットワークを用いてデータから学ぶ技術です。ビデオではディープラーニングが生成的AIの基盤となる技術の一つとして紹介されており、従来のルールベースのシステムから発展し、ディープラーニングモデルがプログラマブルなリソースへと進化したと説明されています。

💡カスタマイズ（Customization）

カスタマイズとは、企業のデータを使ってモデルを調整し、ビジネスに合わせて最適化するプロセスです。ビデオでは、開発者が自身の企業データを使用してプロンプトに応じてより良い応答を作成する「リトリーバルアугメンテッドジェネレーション（RAG）」や、企業固有のデータを使ってモデルを微調整する「ファインチューニング」などの技術が紹介されています。

💡タスク自動化（Task Automation）

タスク自動化は、ビジネスプロセスの自動化を意味し、生成的AIを用いて時間のかかる手動プログラミングを削減することを目的としています。ビデオでは、オンライン小売業者のチャットボットが顧客のリクエストを処理する例として挙げられており、開発者がタスクを自動化するためのステップを説明しています。

💡セキュリティとガバナンス（Security and Governance）

セキュリティとガバナンスは、企業が生成的AIアプリケーションを展開する際に重要な要素です。ビデオでは、言語モデルのパラメータを管理する手法や、モデル出力を正当化するためのシステムプロンプトなど、セキュリティとガバナンスの取り組みが話されています。また、データプライバシーの新しいコントロールも触れられており、企業データが外部のモデルに漏洩するリスクを軽減する必要性があると強調されています。

💡モデルの選択（Model Selection）

モデルの選択とは、特定の用途に最適な生成的AIモデルを選ぶプロセスです。ビデオでは、企業が異なる用途に適したモデルを選択する必要があると語られており、例えば商品説明を生成する場合にアンソルピック社のCLAモデルを使い、商品画像の背景を生成する場合にはステイビリティAIのイメージ生成モデルを用いるといった具体的な例が紹介されています。

💡スケーラビリティ（Scalability）

スケーラビリティとは、システムやアプリケーションを拡大し、より多くのユーザーやデータを処理できるようにする能力を指します。ビデオでは、開発者がインフラを管理することなく、大規模なモデルを実行する能力をスケールアップする必要があると説明されており、これは生成的AIアプリケーションを効果的に展開する上で重要な要素とされています。

💡アマゾンベドロック（Amazon Bedrock）

アマゾンベドロックは、ビデオ内で紹介されているフルマネージドサービスで、開発者が生成的AIアプリケーションを構築しスケールアップするのに役立ちます。このサービスは、リーディングAI企業からの幅広い選択肢のモデルにアクセスできる利点を提供し、カスタムモデルのインポートも可能にしています。

💡アマゾンQ（Amazon Q）

アマゾンQは、ビデオ内で紹介されている生成的AIアシスタントで、セキュリティとプライバシーを念頭に置いて開発されています。Qは、開発者やビジネスユーザーがアプリ開発サイクル全体を効率化し、エンタープライズシステム全体にわたるデータを安全に分析し洞察を得るのを支援します。また、Qはビジネスユーザーが独自のアプリケーションを素早く簡単に作成できるよう支援する「Qアプリ」も提供しています。

Highlights

Generative AI is at the center of boardroom discussions and industries are reinventing themselves with it.

Generative AI can automate end-to-end advertising campaigns and provide hyper-personalized customer experiences in marketing.

In healthcare, generative AI supports clinicians by summarizing clinical notes and suggesting imaging modalities to increase diagnostic accuracy.

Generative AI in gaming creates individualized experiences with infinite variations based on player actions.

Common themes and use cases for generative AI include enhancing customer experiences, boosting employee productivity, and optimizing business operations.

Large language models (LLMs) have deep reasoning capabilities and can fundamentally change how we innovate and build new products.

Foundational models for generative AI are evolving rapidly, with new models being released frequently.

Customers will need flexibility to use a combination of models for different use cases in generative AI.

Model customization is key for businesses to differentiate their generative AI applications using their own data.

Techniques like retrieve augmented generation (RAG), fine-tuning, and continued pre-training can augment data with models.

Strong data foundations are crucial for companies that move fastest with generative AI.

Task automation with generative AI can save time in creating autonomous agents for everyday business tasks.

Security and governance are important when deploying generative AI applications to manage risks like hallucination.

Amazon Bedrock is a fully managed service that simplifies building and scaling generative AI applications.

Amazon Q is a gen-powered assistant designed for security and privacy, enhancing productivity across various roles.

Amazon Q Apps allows every employee to create apps on enterprise data securely and quickly.

Mol AI focuses on efficiency and accuracy for a given use case, including support for European languages.

Mol AI is working on multimodal models and aims to simplify the customization of models for a wider audience.

Transcripts

good afternoon everyone really excited

to be here talking about all things

about generative AI just a little bit

about

me first is I'm a builder and a dreamer

at heart just like many of you have been

super excited in being able to build

with generative Ai and if you look at my

own history I started as an intern in

Amazon more than um 18 20 years ago and

my internship project was to build um

what turned out to be one of the first

uh cloud computing

servers

actually

and over the past two decades I had the

opportunity to be part of some of the

most

Artful technology in the tech

industry sorry can we go back a slide I

think it's uh

yeah

and the beginning of cloud industry was

actually pathbreaking in a big way for

the first time in the history we had

compute storage became a programmable

resource then Big Data Systems like

database and analytics became

programmable and then we had machine

learning which started from like

traditional rule-based systems to uh

linear machine learning to gradient

boosting trees to deep learning they all

became programmable and then we entered

into the era of generative AI where

large language models took place and

with the era of generative AI disc

Discovery now we are entering a space

where AI is now at the center of every

boardroom discussion and every industry

is Reinventing it self with generative

AI to begin with in marketing industry

gen can automate endtoend advertising

campaigns and enable remarkably hyper

personalized customer

experiences in the healthcare industry

it can support clinicians by listening

to Patient conversations summarize

clinical notes and even suggesting

Imaging modalities to increase the

diagnostic accuracy

and then in gaming industry it can

create

individualized experience with infinite

variation of characters missions and

interactions based on the individual

player

actions and across all these industries

we are seeing a few common themes and

use cases emerge that simply wouldn't be

possible without generative AI from

significantly enhanced customer

experiences through highly intelligent

generative AI Assistance or personal

personal virtual

assistance to boost in employee

productivity but things like

conversational search text summarization

or being able to generate code and

coding

assistance and to finally highly

optimize business operations through

intelligent document processing or being

able to do Predictive

Analytics just a decade ago these

projects and ideas were just stuff of

dreams and now they are part of what I

call as a new gen powered reality a

reality that often feels like magic

because these llms have deep reasoning

capabilities and we can add them to our

Enterprise

applications and this technology will

fundamentally change the way we innovate

and build new products and if you look

at the presentation what you're seeing

right here on the vi Tech stage all

these images have been generated by

gen now with with all of these

possibilities many organizations are

asking themselves how do we actually get

started how do I innovate with

generative AI so before we get into it

let's look at uh what is a key look at

the technology that is powering every

gen application

ation the foundational models I'm sure

that many of you are already

experimenting and playing with many of

these foundational models these are

large language models or other

foundational models it seems like every

day we are learning about it another new

model that got released and new powerful

models from companies like anthropic

meta mrol Ai and many many more just

last month even amaz we introduced new

additions to Amazon's own foundational

model called Titan the pace of

innovation we are seeing in the space as

like

unprecedented at AWS we were the first

one to recognize no one model will rule

them all these Foundation models will

continue to evolve at an amazing speed

and customers will need the flexibility

to use a combination of models for

different use cases imagine if you are a

retailer you might want to use

anthropics CLA model to generate a

product description for a new shoe

launch but you may also want to use

stability ai's image generation stable

diffusion model to present a unique

background for your product

image now to enable more developers not

just machine learning scientists to take

advantage of gen it should be easy to

access and evaluate the AI models for

each of your use

case we believe that in the very near

future every developer need to build gen

applications because gen will be the

fundamental part of every experience we

create in the digital world so how can

we make this future a new reality what

would it take to create a seamless

endtoend customer customer experience

that is powered by generative

Ai and I talked about model choice but

in addition to model choice you need lot

more capabilities to build H

applications first you need the ability

to customize these models with your data

next you need the ability to be able to

automate various tasks in your

organization with Ani and finally you

also want the a to Leverage The built-in

security and governance controls to

mitigate any potential risk and finally

you need the ability to scale the

innovation without having to manage the

infrastructure to run these Large Scale

Models now let's double click on some of

these today starting with model

customization now when it comes to

generative AI that know your business

and your customers your data is your

differentiator data is the difference

between a generic gen model powered

application and an application that

knows your business and your customer

now you might be wondering what is the

best way to actually augment data with

your models there are a few different

ways and to how you can use your data to

maximize their

value the easiest place to start is with

a technique called retrieve augmented

generation or rag with rag the

developers need to retrieve their own

company data to add context or

information to their proms using which

they ultimately create better responses

to their customer questions or

requests now you don't need to just stop

at Rack to further improve the accuracy

of your model outputs you can also use

and update the outof the Box found ation

models to create a fine-tune model with

the process of fine-tuning you're taking

a small set of labeled data examples

that is very specific to your company

data to train the model with your

corporate lingo and business

policies and then to take it one step

further you can change the underlying

coree parameters of the model through a

process called continued pre-training

continued pre- training means you pick

up from where the model provider left

off and you train with unsupervised

training with actually totally

non-annotated data this means

essentially you are extending the model

to be an expert in your company's

knowledge base just like how you are

trained a custom model all together now

I talked about all these three

techniques not only do you need tools to

support all these three form of

customization but you also need to

organize your

data what we have seen is that the

companies that move fastest with Gen are

the companies that have the strong data

foundations these are the companies that

typically store various types of data

including their Vector data that can be

used for customizing models and they

have them integrated and across all the

use cases they also use all the

different variety of database tools Al

together with a well governed manner as

well now let's talk about also task

automation with generative

AI one of the most common time consuming

projects for AI developers is creating

autonomous agents that help perform

everyday task for your business and for

your customers for example let's say a

customer wants to exchange black shoes

for brown shoes that they purchase

through an online retailer they use the

sites customer service chatbot interface

to communicate their requests and

confirm that their order returns were

accepted and a new order was placed this

process seems incredibly simple right

but to program such a thing there is

actually a lot of manual programming

that is involved in creating even a gen

power chat

interface first to make this happen

developers need to go through series of

set of steps and all of these are time

consuming like you need to define the

instructions and orchestrate the set of

workflows that you need to go through

you need to write custom code and

finally you need to manage

infrastructure to run these agents all

of this process can take weeks and

require high level of machine learning

Lear expertise being able to tweak the

prompts

appropriately and that's what slows down

the ability to leverage gen to do

Automation and finally let's look at

security and

governance a common challenge when it

comes to deploying a gen application is

hallucination these large language

models have hyper parameters that can

help manage them but you can also use

safeguards like system prompts to site

the sources for the model outputs so

that you can actually get control of

hallucinations there are also new

controls for data privacy for Enterprise

grade geni your data should never be

exposed to the core foundational models

to train their base Foundation model

sending your company data to a third

party is a large risk for any

organization

and for higher risk applications quality

control is an absolute must AI can help

us make predictions but not decisions

which is why human judgment is

absolutely Paramount when it comes to

inference implementing all these best

practices in responsible manner is

Central to building trust with your

customers to build AI responsibly you

will also need need to consider how you

are going to monitor AI system Behavior

prevent AI ABS abuse and Implement

educational programs to rescale your

Workforce at AWS we are investing

heavily in responsible innovations that

enable guardrails for these models

evaluation support and watermarking and

many more now that we have covered a lot

about what does it take to build gen

applications let me start explaining by

how Amazon is meeting these needs for

developers in one easy to use place that

is Amazon

Bedrock Bedrock is a fully managed

service that makes it easy to build and

scale your gen applications and we

recently announced that it is available

in our Paris

region tens of thousands of customers

are already using Bedrock as the core

foundation for the Gen strategy because

it gives them access to the broadest

selection of foundational models from

leading AI companies and with the

addition of a new feature in Bedrock

called the custom model import companies

that are building their own foundational

models can import onto bedrock and

leverage the rest of

capability we know that model choice is

important but what about all the other

things that I talked about

in fact what we have learned is that

majority of Amazon Bedrock customers use

more than one model and this includes

customers like Al liquid which leverages

a variety of models on Bedrock to

quickly prototype digital

experiences now Beyond actually picking

the right model there are a bunch of

other things that you need to streamline

to build J applications faster

and with tools like Bedrock we are able

to do things like knowledge bases for

Bedrock or agents for Bedrock to do

agent Automation and builtin Enterprise

great security and privacy and with

support for various regulatory standards

and

gdpr now by enabling every developer to

build with Gen really easy every

business will be able to move faster and

innovate faster with

Gen but if you want to harness the power

of gen for everybody you also need to

make it easy for every employee not just

developers to leverage it for their

daily

task one way we can accomplish this is

by leveraging gen assistance that

integrate your Enterprise data with an

AI application enabling you to quickly

and easily take advantage of gen for

accelerating employee productivity these

assistants can streamline various set of

tasks just even today right from helping

a software developer to a data analyst

to data scientist we believe these

assistance will provide enormous

impact in fact the first set of area

that is going to get really

revolutionalized is automating

repetitive development Vel ER task these

assistance can remove the heavy lifting

associated with software development

like coding writing tests app upgrades

and security scanning they will also

help employees access their information

faster and get insights and take actions

based on their insights and these

assistants also have the potential to

help everyone build their own gen

application at AWS we have invested in

accelerating the productivity in each of

these areas with our own gen powered

assistant called Amazon

q q is the most capable gen assistant

available today and we built it with

security and privacy in mind right from

the get-go with Q you can get your job

done faster whether you are in it or in

finance let me share a few examples on

how you can put Q into

action with q for developers we are

helping developers become more efficient

across the entire app development cycle

right from planning to coding to testing

to create data engineering pipelines Q

even comes with built-in agent

capabilities to autonomously perform a

wide variety of tasks everything from

implementing features to performing

software

upgrades for example let's say you want

to ask Q to add a new checkout feature

to your e-commerce app it will analyze

your existing code base and map out the

implementation and execute the required

code changes and test in just

minutes I've been inspired by all of the

positive feedback we are receiving for Q

to date and it is the number one coding

assistant in thew bench

today but accelerating productivity

shouldn't just stop at

developers you want it to take it to

every employee in an organization and

that's what Q for business does it

connects to all your company data across

more than 40 plus Enterprise systems and

it provides summaries insights and you

can do analytics all in a secure

fashion but we wanted to take you one

step further and we asked ourselves how

do we Empower every business user to

build their own application so now let's

take a quick look at how we are bringing

this Vision to

life Amazon Q apps enables every

employee to securely create apps upon

Enterprise data in seconds like maryan

HR who needs to create an onboarding

plan for a new hire Amazon Q apps

identifies that this could be turned

into a useful app Mary likes the idea

and goes ahead to create the app with

just a single click and Nikki in sales

who is browsing the Amazon Q apps

library for an app to help her pitch the

company's wide range of products to her

customers she discovers an app created

by her coworker the app generates a

custom sales script tailored to her

request this is teamwork without

boundaries because with Amazon Q apps

everyone can

[Music]

build so there it is Amazon Q and Q apps

will radically change how our customers

and our employees can do their work

every day and now with this new era of

gen Discovery there has never been a

better time to be a

builder I believe that what you are

inventing today and tomorrow can lead to

a profound impact on the world changing

Industries and changing everyone's lives

so while you are here as you learn about

the tools and partners to invent your

next application and

experience please ask yourself what

magic will you build with

ji and now I'd like to invite another

Builder to the stage who is building a

magic of his own with latest

foundational models please welcome

Timothy lacroy co-founder of mral AI

[Applause]

[Music]

[Applause]

[Music]

so

Tim I know mol AI has been extremely

popular so let's talk about what

inspired mrol AI to start a new player

in the AI space in France rather than

let's say in the United States um so I

think there are two parts this question

why did we want to uh create a new

player I think we just wanted to do our

own thing uh we had been working at

large tech companies for a while and

wanted to to get the speed and control

over what we did French really never was

a question for us we're French uh all of

our life is there and there is um a lot

of talent there so it wasn't really a

question and I think it doesn't prevent

us from being a global company we have

offices in London and San Francisco so

okay all right let's talk about what is

the vision of meal Ai and uh how do and

your vision for the future fut of

artificial intelligence I mean I hadn't

seen your slides before but it's pretty

much the same as yours uh okay glad we

are alive I think what I see in the near

future is um a lot more integration uh

needs to be built so all of this uh

agent function calling Behavior will

develop rapidly on the application layer

and on our end we will build uh all of

the tools that are required to uh

actually act uh on these functions that

are available um this will enable a wide

variety of tasks and with this variety

comes the need for um models that

optimized for the complexity of these

tasks um so yeah more gentic behaviors

and more variety of models so when you

talk about variety of models you're

talking about like different sizes of

models or different vertical Industries

how do you think about that uh both so

anytime you verticalized uh or even fine

tuna model what you unlock is is better

capabilities for a smaller size smaller

size can mean either reduced cost or

reduced latency so better user

experience uh verticalization when it's

done for example by us is a bit more

generic but for the practitioner or

model builder uh it can enable some

applications that are maybe easy to

build with the larger models but

wouldn't have the right um latencies or

would be too expensive to put to product

yeah so that's where you're talking

about like different smaller sizes and

being able to customize it with the kind

of techniques we talked about like fine

tuning and Rag and so forth now being in

the middle of all these Innovation and

building your own foundational models

what are you seeing as like major Trends

in artificial intelligence are we going

to keep scaling uh forever this is going

to be all about computer and data

forever or are you going to see a

different breakthrough what are what do

you see in the coming months and coming

years in your opinion I mean people will

always try to scale um I personally tend

to be doubtful of infinite scaling I

don't think these things hold for a

while one limit uh that we run into

quite quickly is the amount of data um

Humanity has only been writing stuff for

so long uh once we've read all of it uh

there is going to be a limit um so some

breakthroughs will be needed but I don't

think we've seen the end of scaling yet

um so that's one uh Direction the the

other is um getting better at the lower

end of the scale I don't think we've

done quite enough work on getting small

models to do what they need um this is

because we don't really understand the

scope of what these models need uh if

you were to tell me okay I want a model

in Alexa what should it do I don't think

this has been built yet and this will

enable lots of research on smaller

models as well yeah and I know we

actually expanded our part parnership

and mrol AI is now available as part of

uh Bedrock how do you envision actually

our customers I mean millions of

customers that are innovating on AWS

today will benefit from mol what kind of

things are possible uh so I think

something that we've seen when building

M roll was that Enterprise customers

really want to uh build AI applications

in their cloud of choice they don't want

to uh move out of it so for customers

it's already uh an ability to get more

choice on uh their current platform I

think one benefit of using mistal models

is the care that we put into uh always

being on the efficiency FR Frontier all

of our models we strive to make them as

uh good as possible for their cost um

and especially in um European languages

with which we started so French uh I

mean English is wider than Europe uh

German uh Spanish Italian uh these

languages are really first class

citizens for us and so um performance

might be even better when using misi

models on this yeah I like that you're

focusing a lot on the efficiency and uh

uh add accuracy for a given use case and

also on languages that are not just

always English based uh it makes gen

more accessible one final teaser

question um what should customers expect

from your future of mral AI in the

coming months any spoilers um I mean I

don't know if it's much of a spoiler I

think we're uh we're working on

multimodal models uh so this will happen

uh I think reasonably soon uh we're

working on some verticalization on code

as well um it's also the easiest for us

because that's what we use every day uh

but more generally I think our goal in

the near future is to help with the

hassle of customizing models which I

think today is still very complex and

and we want to enable this for pretty

much anyone um to do ad scale and up to

prodct that's awesome hey thanks again

for actually coming here and uh joining

me today in this discussion and uh

thanks again for everyone in the

audience for taking good time