Augmentation of Data Governance with ChatGPT and Large LLMs

Summary

在这期《数据之光》节目中，我们邀请到了RecordPoint的联合创始人兼CEO安东尼·伍德沃德，深入探讨了数据治理如何通过ChatGPT和其他大型语言模型得到增强。安东尼分享了他对于大型语言模型定义的见解，以及它们如何帮助企业更有效地管理数据，降低风险，并提高合规性。此外，他还讨论了大型语言模型在数据分类、总结、问题回答、实体识别和情感分析等方面的应用，以及使用这些模型时可能遇到的风险和挑战。通过此次对话，观众能够对如何利用这些前沿技术优化数据治理过程获得更深入的理解。

Takeaways

• 😀 Anthony Woodward是Record Point的联合创始人和CEO，这是一个快速发展的软件服务解决方案，旨在帮助组织发现、治理和控制他们的数据，以便更严格的合规、更高效和更低的风险。
• 📊 大型语言模型（LLMs），如ChatGPT和其他，被讨论用于数据治理增强，它们通过神经网络，能够生成文本，进行对话，写作文，等等。
• ☕ Anthony分享了他对咖啡的热情，以及搬到西雅图后，如何继续他的咖啡爱好和学习雪板作为新爱好。
• 🧠 LLMs是基于数学模型，能够模拟人脑的连接神经元方式来处理和生成语言，这让它们在处理语言方面表现出色。
• 🔍 LLMs在数据治理中的应用包括数据分类、数据摘要、问答系统，以及实体识别和情感分析，这对于管理风险和合规性至关重要。
• 🤖 尽管LLMs表现强大，但它们也容易受到偏见的影响，需要管理和调整以确保准确性和公正性。
• 💡 LLMs能够在不需要预先定义的本体论的情况下生成分类和本体论，但通过细化可以提高它们的准确性。
• 🚀 Record Point使用LLMs来提高数据分类和标记的效率，同时也提供了一个反馈循环，允许人工校验和改善分类结果。
• 📈 LLMs预计将极大加速数据治理和其他商业流程的能力，预示着人工智能技术将成为未来技术的一个核心组成部分。
• 🔑 Anthony提醒，使用免费版本的LLMs（如ChatGPT）可能会与数据隐私政策冲突，推荐在商业应用中考虑使用付费版本或API。
• 🌐 LLMs的未来发展将进一步深化它们在数据治理、隐私保护和业务决策支持中的作用，但同时也需要注意数据的安全性和隐私保护。

Q & A

安东尼·伍德沃德是谁？

安东尼·伍德沃德是Record Point的联合创始人兼首席执行官，这是一家快速增长的软件即服务解决方案公司，专注于帮助组织发现、管理和控制他们的数据，以实现更严格的合规性、更高效率和更低风险。

什么是大型语言模型（LLMs）？

大型语言模型是一种人工智能或机器学习的变种，特别是使用神经网络，它们是非常擅长生成文本的数学模型。这意味着它们能够与人进行对话、生成论文，并且能以与问题上下文相关的方式进行，而且语法正确。

大型语言模型如何帮助数据治理？

大型语言模型可以帮助数据治理通过：1. 分类数据，2. 汇总数据，3. 提供问答服务，4. 实体识别，5. 情感分析。这些功能可以用于标记数据、简化信息，并根据风险和治理要求采取相应的行动。

大型语言模型能否自动生成本体论？

大型语言模型位于监督学习和无监督学习之间，它们可以在一定程度上自动生成本体论，但通过微调模型可以提高准确性。尽管它们能够生成英语基础的本体论，但提高特定本体论的精确度需要进行定制化训练。

大型语言模型在处理个人数据时可能面临的风险是什么？

大型语言模型在处理个人数据时可能面临的风险包括偏见、数学幻觉、误导信息等。这些模型可能会反映它们所代表的数据的偏见，或者因为模型混淆而产生错误的联系和输出。

如何确保使用大型语言模型时考虑到批判性思维？

使用大型语言模型时，需要培养用户的批判性思维能力，以判断模型提供的结果是否可信。这包括理解模型的局限性、识别可能的偏见，并在决策过程中综合考虑多个因素。

Record Point是如何利用大型语言模型改进数据治理的？

Record Point利用大型语言模型通过分类和标记数据，理解数据的隐私背景，以及提供数据的控制和处理能力来改进数据治理。他们专注于数据清理、分类和企业内不同系统的数据控制。

大型语言模型能否直接应用GDPR本体论来构建政策和规则？

大型语言模型本身不能直接应用GDPR本体论来构建政策和规则。它们可以提供关于如何应用GDPR的建议，但实际上实施特定的数据保护和隐私管理措施需要更专业的工具和流程。

大型语言模型的未来发展方向是什么？

大型语言模型的未来发展方向预计将深入整合到各种技术和业务流程中，提高信息汇总、决策支持和数据处理的效率。它们将在未来的技术解决方案中扮演关键角色，影响从数据治理到客户服务等广泛领域。

安东尼·伍德沃德平时更倾向于使用哪个大型语言模型？

安东尼·伍德沃德在使用大型语言模型时没有特定偏好，他在不同的情况下会根据需要选择使用Chat GPT或Bard。他认为这两个模型各有优缺点，并且随着模型的不断更新和进化，他的选择也可能会改变。

Outlines

00:00

🎉 引言与嘉宾介绍

本视频节目专注于数据治理与大型语言模型（如ChatGPT）的增强。特邀嘉宾Anthony Woodward，RecordPoint的联合创始人兼CEO，深入探讨了数据治理的现状及其与隐私保护的交集。Anthony分享了自己的背景，包括他从澳大利亚搬到西雅图后的生活变化，以及他对咖啡的热爱和对冲浪与滑雪的兴趣。

05:00

🔍 大型语言模型简介

Anthony对大型语言模型（LLMs），如ChatGPT和Google Bard进行了简介，强调它们在文本生成方面的强大能力，能够进行对话、撰写文章等。他解释了这些模型如何通过学习互联网上的大量数据集来识别语言模式，并在回应中保持语法正确性。尽管这些模型表现出色，但Anthony指出，它们目前还不能理解或反映人类的复杂思维过程。

10:01

🛠️ 数据治理中LLM的应用

讨论了LLMs在数据治理中的潜在用途，如分类数据、总结数据和问题回答，以及识别实体和情感分析。Anthony解释了LLMs如何帮助标记数据，提高数据治理的效率，并通过分析和总结数据，支持决策过程。他还探讨了LLMs在隐私和控制方面的应用，如识别与个人相关的细节（如种族或健康信息）并分析数据的情绪倾向。

15:02

🚀 LLM在实践中的挑战

Anthony讨论了LLMs实际应用中的挑战，特别是它们倾向于反映并放大现有数据中的偏见。他强调了LLMs可能"幻觉"或产生误导信息的风险，这通常是由于模型训练数据的不完整性或偏差。此外，他提到了用户在使用LLMs结果时需要批判性思维的重要性，以及专家在纠正模型偏差和验证信息准确性方面的不可替代性。

20:02

🤖 LLM与商业数据的桥梁

探索了LLMs如何作为商业和数据之间的桥梁，尤其是在数据驱动决策和数据故事讲述方面。通过将LLMs应用于企业内部数据集，可以改进信息搜索和总结，支持新产品研发和市场分析。Anthony同时提醒，使用免费版LLM服务时要注意数据隐私问题，建议通过付费服务来保护企业数据。

25:03

📊 RecordPoint的角色与贡献

介绍了RecordPoint如何利用LLMs增强数据治理，通过自动化数据分类、管理和控制来提高效率。公司开发了大量连接器，使客户能够轻松地将LLMs应用于不同的数据源，如Salesforce和Twitter。Anthony强调了人工检查和模型训练的重要性，确保数据处理的准确性和符合规范性。

30:04

🌟 数据治理的未来趋势

讨论了数据治理和LLMs应用的未来发展，预见LLMs将如何继续革新数据管理和分析过程。Anthony预测，随着技术的进步，LLMs将被广泛集成到各种业务流程中，从而提高决策效率和数据利用率。他将LLMs的兴起比作互联网的革命性影响，预示着数据治理领域即将发生的深远变革。

35:05

🔑 LLM工具选择与应用实例

Anthony分享了他个人在使用LLMs工具时的偏好，比较了ChatGPT和Bard的优势和用途。他说明了在不同场景下，根据所需的信息类型和处理能力，会选择不同的LLM工具来辅助工作。通过具体例子，展示了LLMs在撰写商业计划和报告中的实际应用，体现了其在提高工作效率和信息获取能力方面的价值。

Keywords

💡数据治理

数据治理是指对数据的管理和控制过程，确保数据的质量和安全。在视频中，数据治理与大型语言模型（LLMs）如ChatGPT的结合被讨论，强调了利用这些模型对数据进行分类、总结和实现合规性的潜力，提高效率，降低风险。例如，通过应用大型语言模型对数据进行自动分类，有助于实现更严密的数据治理。

💡大型语言模型（LLMs）

大型语言模型，如ChatGPT，是基于深度学习的模型，能够生成文本、理解和回答问题。视频中将其解释为通过数学模型近似人脑的方式，利用神经网络对大量数据进行处理，以此生成与输入相关的、语法正确的文本。这些模型在数据治理中的应用，如自动化数据分类和提供总结，展示了它们如何辅助处理大量数据。

💡数据分类

数据分类是将数据根据其特征、用途或其他标准分成不同类别的过程。视频强调了LLMs在数据分类上的能力，如能基于内容自动生成标签，对数据治理而言，这能大大提高对数据的理解和管理效率。例如，通过自动识别和标记数据的敏感度级别，帮助企业遵守相关的数据保护法规。

💡合规性

合规性涉及确保企业活动符合相关法律、政策和标准。视频中提到，利用LLMs可以帮助组织更有效地发现、治理和控制数据，以满足更紧密的合规要求。通过自动分类和处理数据，LLMs能帮助企业更好地理解其数据，确保遵守数据保护法律如GDPR。

💡风险降低

风险降低是指通过各种措施减少潜在损害的可能性。视频中，通过使用大型语言模型来增强数据治理，可以帮助组织更有效地识别和管理数据相关的风险，比如通过识别敏感数据并确保其合理处理，从而降低数据泄露或不合规的风险。

💡效率提高

效率提高是指用更少的资源或时间完成更多的工作。视频讨论了LLMs如何通过自动化处理大量数据任务，如分类和总结，来提高数据治理的效率。例如，使用LLMs自动化生成报告或提供数据洞察，可以节省时间并加快决策过程。

💡总结能力

总结能力是指将大量信息压缩成简短、易于理解的格式的能力。在视频中，强调了LLMs在数据治理中的应用，特别是它们在自动总结数据方面的能力，这有助于快速提供数据的关键洞察，支持更有效的决策和报告。

💡隐私

隐私是指保护个人信息不被未授权访问或披露的权利。视频中讨论了LLMs在处理数据时如何识别和处理个人敏感信息，以及它们在帮助组织遵守数据隐私法规方面的潜力，例如通过识别包含个人信息的数据并适当处理以保护隐私。

💡实体识别

实体识别是从文本中自动识别和分类具有特定意义的信息（如人名、地点、组织）的过程。视频中提到，LLMs能够执行实体识别，这对于数据治理尤其重要，因为它可以帮助识别和标记涉及特定个人或实体的数据，从而更好地管理和保护这些信息。

💡情感分析

情感分析是指确定文本的情感倾向（如正面、负面或中性）的过程。视频中讨论了LLMs在识别数据情绪方面的应用，这可以帮助组织理解数据背后的情绪，如客户反馈，从而对策略或产品进行相应调整。

💡偏见与误导

偏见与误导指的是由于数据或模型的不完善而产生的不准确或歧视性的信息。视频中提到，虽然LLMs在数据治理中有很多优点，但它们也可能因为训练数据的偏见而产生误导性的结果，强调了监控和纠正这些偏见的重要性。

Highlights

Introduction of Anthony Woodward, CEO of Record Point, and discussion on augmenting data governance with ChatGPT and other large language models.

Anthony Woodward shares his hobbies, including a passion for coffee and snowboarding, illustrating his personal connection to the topics of discussion.

Explanation of large language models (LLMs) as AI that excels in generating text, holding conversations, and producing grammatically correct essays.

The potential of LLMs in classifying data, summarizing information, and assisting in data governance by applying classifications and ontologies.

Discussion on the role of LLMs in enhancing data governance, including their ability to improve data classification and summarize data points.

Exploration of how LLMs can contribute to data governance by recognizing entities and analyzing sentiment, particularly in privacy-sensitive contexts.

Addressing the capabilities of LLMs in generating ontologies and classifying data based on both predefined ontologies and content-derived ones.

The conversation highlights the risks associated with LLMs, including biases and the potential for generating inaccurate or misleading information.

Insights into using LLMs as a bridge between business and data, enhancing storytelling with data, and providing a more nuanced understanding of business data.

A word of caution on the different privacy implications between the free and paid versions of OpenAI's services.

Discussion on Record Point's approach to leveraging LLM technology for data governance, emphasizing the importance of a feedback loop for verifying classification.

Insight into how Record Point's unique application of LLMs enables tailored data governance solutions that respect customer privacy and data integrity.

The potential of LLMs to build policies and rules based on specific ontologies, such as GDPR, and apply these to classified data for enhanced governance.

A look into the future of data governance with LLMs, suggesting a significant acceleration in capability and potential applications across various business processes.

Closing remarks on the evolution of LLMs and their impact on data governance, with Anthony Woodward sharing his preference between ChatGPT and Bard based on their respective strengths.

Transcripts

[Music]

hello everybody everyone welcome on the

lights on data show today we're going to

talk about augmentation of data

governance with cha gbt and other large

language models

I love it it's a great great topic and

very very relevant today our wonderful

guest today is Anthony Woodward Anthony

Woodward is the co-founder and CEO of

record point of fast growing software

service solution focused on helping

organizations discover govern and

control their data for tighter

compliance more efficiency and less risk

Anthony is regarded as one of the

leading thinkers on the intersection of

data and privacy welcome Anthony it's a

wonderful to have you on the show

thank you thanks for having me

now before we go into

um this great topic today I wanted to

ask you if you have some interesting

hobbies that you would like to share

with our listeners and if maybe you've

discovered a new hobby since you moved

all the way from Australia to Seattle

yeah I know uh look I I bought one of my

hobbies um with me and I certainly

discovered a hobby as well

um I share a a passion for coffee I

wouldn't be very Australian if I wasn't

feeding everyone over their head about

how much better coffee is in Australia

and how I think most Australians think

we've discovered coffee which I know is

very untrue but um you know certainly

there's a view that we've perfected it

so um I know I followed George on

LinkedIn and see your your weekly posts

on coffee and um if you head over to my

LinkedIn I try to one-up you on that um

with some some different uh you know

siphon coffee or some other pieces so um

certainly coffee is a strong passion and

um I don't know how anybody survives a

day without something funny

we're in great company

absolutely and the good news you know is

um uh really having moved to Seattle

which is the U.S capital of coffee they

had some work to do I think they catch

up on Melbourne uh but um I was able to

at least a transplant

um that that particular passion and

continue it uh in Seattle and look you

know I'm coming from Australia I I

certainly um surfed a lot as a child and

spent a lot of time out on the waves

which is a lot harder in the Pacific

Northwest

um but at least being how cold it is in

the ocean

um so I'm a bit of an avid snowboarder

these days it certainly transferred my

um my board uh you know uh water but

water in the water into the board in the

snow so that that's uh that's certainly

a piece out how to spend my weekends but

so yeah I've been going for hours about

that but but

um those are might kind of be too

obvious

we love it it sounds very very exciting

maybe you can go to Whistler sometime if

you haven't already and to Tofino for

surfing certainly certainly it may not

always though and and planning to be out

there again when the season rolls back

around so

wonderful

well we're happy to to have you closer

um so let's get into the topic and

today's topic so our first question uh

just to ease us into it

um what are or what is a large language

learning model

yeah it's a great question you know

there's a lot of hype out there uh how

chat GPT and you know Bard and a bunch

of other uh large language models but

um if I can really break it down

um to the level of Simplicity you know

large language models are uh it's AI or

you know A variation of machine learning

in particular using a neural network and

we'll get to that uh in a little bit in

a second but they're just really

um mathematical models that are

astonishingly good at generating text so

you know that what does that mean well

it means that they can hold a

conversation with you uh they can

generate an essay they can

um do that

um in a way that is

um you know really coherent to the

context of the question you're asking

them

um and they can do that you know that is

grammatically correct particularly in

English I mean it's not talking about

other languages at this stage they're

not not as good

um but what

um were there really

um are at the end of the day is a

mathematical model that

um really approximates the human brain

in the sense of having interconnected

neurons so if you think about a brain

and how a brain works where there are

neurons and there's a firing to connect

different thoughts and processes

together as a sort of conceptual model

uh you know large language models are a

way to use a neural network

to transform

weighted mathematical models into

desired output as language and what that

really comes down to is a really really

large set of mathematical models that

are able to generate and transform data

from the statistical numbers it has for

each of the words so all of that's kind

of a lot of a lot of pieces but to

really simplify that at the end of the

day if you think about

chat GPT or Google bard what they've

done is being said large data sets out

of the internet they've broken that down

to look at what are the patterns of

language or words that go together that

understood all right if I I'm asked a

question about

um a particular topic what is the kind

of language I've seen on that topic and

then give it back to you in a structured

grammatically correct English way

they're not able to think they're not

able to

um really go beyond

um you know the recognition of entity so

recognition of what are the things

occurring around these words

um and they're just able to then

regurgitate these things within concept

so you know the hype cycle around

um them being able to you know turn into

Skynet or something of that else is is

very far off they are really just a very

complex but

um large set of data in a mathematical

model

okay that's good news about about uh not

being there yet that's kind of level so

we have a question here I love the way

you explained it it was really very you

know in a very easy to understand way

more complex thing you know very easy to

understand why thank you yeah I think we

have a question here from Kate strasny

and she's wondering then what problems

are llms solving for data governance or

you know what how can we use llm to

augment our data governance efforts

great question Kate and it's something

that certainly

um I think still going to evolve right

we probably don't have all the answers

today by

um there are some really exciting areas

that this opens up so

you know really considering that that

definition of large language models

there are some things that they're

really good at you know firstly

classifying data so if it's able to

understand

different strings of text or potentially

even different strings of data rather

than just hacks then

um you know we're able to quite

accurately and reliably apply a

classification or a series of different

ontologies uh to that data and that gets

super interesting because now we can tag

the data up for different aspects of

governance slash risk if I can think of

both sides of the coin governance is as

a result of risk that we've identified

um you know to do things and really

that's where things get interesting

right because if we're able to

understand those tags we can use a

um statistical model that isn't

pre-trained to look over a lot of data

set that my needs and governance and

apply that labeling

but that's sort of job one job two is

being able to summarize that data you

know one of the problems I think we have

in the data governance field is not just

understanding

you know what is the data how should we

action the data what is the risk of that

data but also how do we give it back to

those that need it in forms that's

consumable so you know

um uh we we've traditionally had search

engines and other processors to allow us

to search and control through find

things that that come up as hits and

then go through that data the great

thing about Live Language models is it

is it's really great at summarizing

um different data points but potentially

even summarizing multiple data points

within a relationship and that's where

it gets really interesting so if you can

imagine

I have data it's classified within

um you know an ontology of even risk um

you know that allows us to look at you

know a good classic example um depending

on the jurisdiction you're in

occupational health and safety OSHA in

the US OHS and other parts of the world

um and we can start to break down you

know a summarization of all the risks

inside the organization based on using a

large language model and then output

another ontology that can apply to

classification this stuff all kind of

Builds on itself I'm so it's really good

at that it's also really good at

question and answering so when you think

again about the application of these

tools you know going and asking

um can you quantify uh how much of a

thing do I have how much of information

I have to provide for lawyers in a legal

case is very very good at that and then

then lastly you know the last two points

um which I think really come to the area

of privacy

um and control it's really good at

entity recognition but not even the sort

of traditional entity recognition that

we think about of people places

organizations companies you know those

are all pretty straightforward and

although you know llms are really good

at that entity recognition as you

recognition to things like ethnicity

um you know uh you know health issues

um that might relate to a person so

really getting more granular and

thinking that about the risk of that

data and how is your government you know

that that's it within that and then

connected to that is sentiment analysis

so being able to look at the sentiment

that wraps around it you know I was

talking to a

um a large customer today in the

healthcare sector and they were talking

about look I want to find all of the

doctor reports that

um are to do with a particular type of

cancer that

um have had a negative outcome for the

patient

um and we use a large language model to

break that data down and blur through it

obviously you know you've got to be very

careful when you get get all the

doctor's reports in a particular

um Health in a particular hospital

because there's got to be a bunch of

these airlines are really good at

getting away from those false positives

understanding the sentiment

understanding the connectivity of those

words and then bringing that data back

um to be used in a meaningful way so you

know very good at that kind of data

mining content categorization and then

summarizing a data right okay well we

have some follow-up questions to this

and uh the first one comes from Dan

Everett and specifically on what you're

mentioning on the that the llms can

really help with that classification

he's wondering then let me just bring it

up here

is it doing the classification based on

ontologies that you have to build or

deontologies are being derived from the

content itself

what a great question then

um the beauty of llms is

um in in the world of AI there are a

couple of different AI tools so all AI

methods so there's supervised and

unsupervised so supervised means that

we've got a set ontology and we're

training it we're going to train it to

do a thing um it could be an ontology it

could be some other uh Factor the thing

about llms is they are somewhere in the

middle of the supervisors and and

supervised in that they're kind of

semi-supervised so you do need to give

them some tokens to fine-tune the models

but

um it's it's fair the models by

themselves are pretty good at uh

building an unsupervised set of

processes so what does that mean it

means that they are really able to

generate ontologies by themselves but

they are better if you fine-tune them

like if you want to get down to the

higher levels of accuracy then you can

fine tune them by training them and and

manipulating

um the models to be more effective but

they will generate an ontology without

that they'll you know they do have an

understanding of English in particular

and I'll keep saying that because

they're not necessarily so good today at

non-english and that will evolve and get

better over time by

um you know they are very good at

building English based autonomy

ontologies

um and being able to do that

classification themselves but you can

make them better by doing the sort of

what AI experts would call transfer

learning where you're able to give you

more specificity uh to the model

um on the ontology you're trying to

create

and and uh my other question was you

you're mentioning that you can really

figure out different entities

um from from that data so you mentioned

ethnicity as an example so it doesn't

mean that it's it's really extrapolating

figure out the different attributes that

uh would be based on everything else

that you kind of have in there and

determine oh you know you are you know

this and that

yeah so it's some really interesting

um uh academic papers around

um you know looking at even the use of

language and what that indicates around

uh your your ethnicity and background

um and the types of words that you use

you know I'm I grew up Australian so

um we do use some funny

um uh phrases and phraseology

um that when we write and and talk about

things

um yeah I confuse the uh my American

friends all the time with uh with things

I say

um and I think you know that will come

through in my emails and my speech and

and so the models are able to use that

to then match where they've seen that

you know because these large language

models are large they've got you know

English language from many places again

they can attribute that back into a

particular

um Regional rationale or driver based on

the language and the use of that

language it does go further though where

um you know you can also look at the um

data about you particularly if you're

feeding it with

um privacy data or data around

um the individuals that that are

concerned and it can use those factoids

to build a a picture of your ethnicity

and background also

well those are all incredible benefits

as well

yeah

but of course there's risks uh I I would

think I mean that last example that you

mentioned it can definitely turn into a

big risk do you are there any that come

to mind that we should watch out for

plus you're uh we talked about the good

things about large language models in

reality there there have already been

some bad things

um so you know and I think

um it is worth saying that the you know

I started off by saying it's not Sky now

and and some of the proof points that

it's not skymet are our large language

models are really prone to biases if not

manage properly

um you know they can effectively

mathematically hallucinate

um you know and lie and really that's

just because it's a reflection of the

data that they represent but they get

confused

um and they get confused because they

see a a group of things and I say they

see it because again I'm trying to refer

them almost as if they're sentient and

they're not

um but but the model sees

um you know a a mathematical set of

Concepts that are connected those

neurons again that we spoke about

um but but in the same way that your

mind can make things up and and give you

you know

um experience the concept you haven't

actually experienced they you know the

model can do the same and that's

collected a whole bunch of facts so it

crew it creates out of those facts or it

doesn't actually it sees in those facts

that aren't really joined a relationship

and from that relationship it will give

you a bias or it will give you a

hallucination or maybe even the data

itself that it was fed had a bias or a

hallucination in it but when you look at

it across the abstraction you can't see

it as a set of individuals but when it's

connected it's there so

um you know that's something we really

have to be very careful about like how

factually accurate are they

um you know how how

um how are those models going to

um be used and and what is the

prevalence behind the model so that you

know you know how it's constructed and

and what biases are going to lie inside

that

and also how can we maybe educate the

users

to you know in critical thinking when it

comes to the results the chat TPT brings

up and this is in general also for

Google searches right

um You need to use your own critical

mind to establish if those things that

come up are actually trustworthy

but I think the chargpt this makes it

even more complex and more difficult

yeah absolutely and it's you know it's

why you know

um we're always going to need experts

because you know

um they're still going to need to be

some level of human triage of those

things like the models are not good at

understanding their own biases or

self-reflecting you know being able to

do that self-reflection they're just a

mathematical model so of course they

can't self-aver play that seems

self-evident but you know we do I mean

even myself you're Amazed by tap jpt you

go ask it a question it gives you this

fantastic answer it appears sentient but

but it's really not it's just this

mathematical model that has no ability

to self-reflect

um so we can't think about whereas

biases are and correct for that and

let's just talk to so again the good

thing about these models is they can be

trained out of some of those behaviors

and you know we're only at the beginning

of the arm to do that well

so

um Petra here has has a question uh for

you Anthony and he's wondering

how do you see llms especially Chad gbt

but not necessarily can be a bridge

between business and data both in terms

of being business slash data driven but

also in telling stories about data and

stories with data

yeah I think what's really great about

the Live Language models is you can use

them as a seed

um to match against your own business

data so you know I talked before around

the different ways that we can use them

from a data governance perspective but

um it also can apply just as well to

data sets

each other inside the firewall inside

the business so the

um you know there are some really great

scenarios of using them as a much better

search engine again because they're much

better at that

um summarization of information and much

better at understanding the context of

information and much better understand

the context of your question

um so you know you can you know go ahead

and use them to bring that data set

together you know uh when you're asked

for you know in a business context of

being asked to research a new product or

research a new capability you know being

able to look at all of the data already

inside your inside your business

summarize that data and you know uh

develop that new product for instance

they're very very good at that

um so you you can already using open AI

using some of the apis from Google and

and others Facebook and others

um you know do that today I will say

just click one word of caution

um if you are going to go use open OA

and chat GPT please go pay for it not

that I'm not recommending you pay them

money but a really big different privacy

implications if you use the free version

we probably don't have time to cover it

all in the episode today for the whole

other Episode by itself uh but just a

word of warning the privacy policy for

the free version versus the paid version

is very different effectively sharing

your data for free with open AI if you

don't pay for it

um and that can have some really big

implications from a business perspective

as well

very interesting so and and yes like you

said we're not going to get into it but

just to clarify is it um if you pay for

it so if you uh upgrade yours you have

the P subscription or if you're using

their API to then connect is it either

or or just the API version that would

not would respect your privacy a little

bit more

yeah unless it's changed like I believe

in order to use their API in any

meaningful way you actually have to pay

for it anyway so they can't do the same

rate

um there are some free elements for the

API by um in order to properly use the

API of business data you would you would

need to pay right right and is that ever

it is you know reminding us if it's a

free service then your data is the price

absolutely absolutely and look I think

you know as we've already seen again um

with these large language models you

know the very large companies behind

them are just chomping through our data

and and using it to perfect their models

and without attribution and you know

there's a whole a bunch of law to be

written you know seeing what's happening

in Italy and other parts of Europe

um where there are some real caveats

around what they're doing uh you know in

the world and that's yet to play out um

let's see how that case will

um develops over time

so so Anthony

um let's talk a little bit about Breaker

Point too where where um where does it

come in how how can it help our data

governance efforts

uh we're at a point of being um doing

data governance for some time where a

SAS solution uh you know we work with

people like New York City Seattle City

and

um there in British Columbia where you

are absorbed you know with um the the

liquor and cannabis board for instance

um the

um what what we've done for some time is

classify data to apply retentioning but

to also understand its privacy context

um how that data should be treated and

handled over time

where

um we've been doing a lot of work and

then where we've been adding to our tool

and where

um we'd be really looking at large

language models is really improving that

classification that tagging so what

we've done

um that we think's rather unique is

um we build about a thousand odd

connectors to Upstream systems so you

can feed these models

um you know be that Salesforce be that

you know Twitter and and

um like data Lakes like snowflake or

just traditional places like file shares

and so we've really focused I guess on

three points one is how do you clean

your data up so the very basic things of

looking at redundant obsolete trivial

data because we want to you know in

order for these large language models to

work we actually need better cleanse

data that has less biases and all those

sorts of things

um so absolutely that that data

cleansing capability or we called Data

inventory

um two then looking at the

categorization of data what um

ontologies does it apply to so from a

data governance perspective that could

well be a risk ontology it could be a

record-keeping file plan ontology it

could be um you know looking at

different dimensions of privacy you know

who are the entities in this document

who does it talk to what are the

relationships or or data um you know

they use the word document and data

interchangeably here

um and then providing that control so

that the Enterprise is able to do that

in all of those systems

um within a large language model so

that's very much where we've been

focused at record point is really

building out

um very simple ways for a customer to

leverage these Technologies

um but point and click connect to a

Salesforce classify that Salesforce

manage that Salesforce and Anthony I'm

very curious if if as you're leveraging

this technology do you have some sort of

a feedback loop that your somebody is

sort of verifying that classification

for example as a first round and you

have somebody from the business and

they're like yeah yeah you know spot

checking and uh it's good to go

yeah some of the things that we've

always we've really felt um is important

um I don't think our customers agree

um you can operate our tool again it's

it's semi-supervised

um so you can operate a completely and

unsupervised fashion and it will just go

and do a thing and the accuracy comes

down but there are scenarios where that

makes sense I've got a large set of data

I just wanted to go and discover right I

just wanted to go classified it can go

do those things

but we also recognize and and I think

you know um one of the things that's

quite yeah that we've really focused on

and done a lot of

um development r b on is the ability to

um override or have a human train the

model because it's it still is the case

these large language models are very

good

um but they still as we said earlier I

have biases and problems and we need

humans to be able to ride back things

we've done an outsole that relates to

give you that capability go it's not one

of these or it is one of these

um in a much stronger way and then

adjust the model

um based on your data set and that's

that's a really important

um governance piece right because you

need to be able to govern the model not

just govern the data

and what record point is doing to help

one client would it be able to benefit

another as well

yeah again you get into some some

interesting legal issues so yes in terms

of process we we learn

um where people are tweaking the facts

in the models we learn

um what are the various seeds we should

provide we don't share in our case every

customer has a model to themselves so

their data never leaves their own

tenancy within our SAS solution

um and is confined to that so that way

we're not sharing data across sort of

boundaries because even though the data

isn't trapped in a model there is waste

reverse engineer that or you know you

may end up with outputs that are raised

based on another another business if if

you shared these models so we've

developed a way to create a effectively

a seed model that um doesn't have any of

the businesses own data in it but gets

you started

um and then on that you'll then build

with your own data over time and it but

it is completely um separated from every

other

um customer in in in our platform

Dan is wondering about this uh very

specific application if if you have the

llm that's um we have an ontology based

on gdpr

then can you use the LMS to build

policies and rules based on that

ontology classify that data based on it

you know and apply property policies and

rules to the classified entities

um llms by themselves won't do those

things

um they'll go close to giving you

and on you know if you've given them the

rules of uh of gdpr and then you've

looked at uh asking if the question

around how duty power applies

um they will give you answers but they

won't they won't

um do those things uh where you take the

specific policies of say data subject

access requests in dsat and then track

the time periodicity of supplying that

back um to a customer they won't do that

out of the box

um yeah and I thank you very much for

the tool because it does does lead back

a little bit to what we do at record

Point that's how we're augmenting those

models is looking at those workflows and

those specific processes in gdpr or CCPA

or whatever the legislative outcome is

um and and giving you the actions that

come from that classification so that

that that's very much the world we live

in is how do you you apply a set of

regulatory actions to an AI outcome

so I mean I can

I understand all these benefits and it's

fantastic it really makes my heart jump

uh in a positive way it's just looking

at all the possibilities and how the

world is advancing in the series is

purely fantastic do you think I mean

it's definitely removing a lot of the

hassle from organizations on how to deal

with some of these data governance

aspects do you feel it will still be

able to or it should still involve

stakeholders from across the business

and their data governance efforts now

with you know Chad GPT and LMS and

record Point really helping a lot of the

groundwork

absolutely I I mean you have to bring

the

um yeah the business owners and and the

stakeholders along

um in this journey you know we you know

the thing about AI is it is a big leap

like it is really accelerating our

ability to do things at a really great

Pace that previously took a lot longer

um and was much more difficult to do

but you can't just go take that leap in

the corner because if you take the leap

in the corner

um you've kind of you know jumped a

dimension if you will

um but nobody else came with you

um and so I think you know there's so

many benefits here for the business and

how we're organizing data and thinking

about

um they've been making data more

effective that you'd want to share it

you'd want to bring your stakeholders

along you'd want to integrate them into

that and I think what's really great

about what's happened in the public

discourse around open Ai and chat GPT

and the rest of it is there's a real

awareness of wanting to have that

conversation and seeing this as a as a

real mechanism to make their lives and

their jobs easier you know what we've

really always focused on our record

point is how do we abstract as much of

the data governance and and risk

decisions from the user so it doesn't

burden them the really great things

about llms is we've really taken about

now that Leap Forward and I think that's

the starting point of that conversation

of you know I'm not here to come and add

something more to your day I'm actually

here to come and take something away

from you today

um and make it more effective and better

for you and that's really

um you know when you have that

conversation it's really productive

right yeah and it will frees up the mind

of the stakeholders so they can focus on

the Strategic decisions which is awesome

so that's great to bring them in because

now they you know they have more

capacity to to focus on that

um I wanted to ask Anthony how do you

see that um you know this Evolution will

help data governance so where where does

this all go towards you you mentioned

acceleration

um just now so where are we heading

yeah

um you know uh it's it's the it's the

million dollar question

um look uh we really really are at a

very early stage for what we've seen

from our chat GPT and and and you know

the the really large language models

that are out there that there is you

know whilst they are going to

revolutionize

um the processes we have you know I

think in data governments they're

actually a really easy application like

we're one of the first areas that we'll

be able to take advantage of them very

quickly

but

um there is a plethora of additional

workloads and capabilities that we're

going to be able to use them for right

because they're really good at

understanding uh context you know and

um summarization and

um capturing

um you know different

um decisions is a strong word but

supporting evidence to decisions

um you know you're really going to see

them more and more pop up into different

business processes that we all have so

um you know you're being able to augment

a workflow you know onboarding a patient

to a hospital but bringing that data set

summarizing the kind of problems they've

had in the past so really bringing facts

and making them

um Bubble to the top in a timely manner

of that context and that's a very simple

example but that really applies to

everything so you know uh we will really

see them roll out in in systems

everywhere I think in in 10 15 years

time

every technology will have some AI llm

component to it you know I I certainly

see llms and what's happened with AI

here is just as revolutionary as the

internet if not more Revolution internet

yeah

yeah the future is very exciting lastly

I want to take this question from Ravid

he's wondering what do you use most

charging PT or Bart

um I'm really oscillating between the

two

um they both have some strengths and

weaknesses as a as you as you probably

have experienced

um look if I asked to say advice to take

today as an example I use Bard a little

bit more today

um I I found bad uh I I just a touch

better and

um I've today I was working on a new

business plan and some board reports

um and I was looking for some data to

bring into that but a little better at

bringing in statistical information and

and those sorts of things so

um I've I've leaned in that direction

but these models are evolving so quickly

you know open AI put out gpt4

um there are others coming in the

pipeline that we'll see come out and and

so you asked me that same question again

tomorrow and I'll probably give you a

different answer but it's a great

question

thank you Anthony

well thank you so much for being on the

show and putting the lights on uh

augmentation of the governance with Chad

GPT

[Music]

not having me it was a lot of fun

Anthony where can people find you and

also find more about record point

I'm usually on a plane over the Pacific

somewhere that um

uh so so look uh yeah Anthony Woodward

on LinkedIn and many other places

um you can find me in you know any any

social network

um I'm relatively active but

um if you if you hit recordpoint.com

um you know we have a whole bunch more

description our website we have a number

of blog articles on large language

models and chat GPT and how we're using

Ai and those processes

um and you know certainly from there

um if you uh hit um one of the contact

Arsenal or some of those processes there

um you'll be able to talk to myself or

one of the team and

um always happy to have a reach out on

Twitter so I'm at woodwas w o o d w a on

Twitter or hit me up on LinkedIn um

always hoping to have a chat with

anybody that's interested in this topic

oh thank you

all right well thank you so much

everybody for joining in on this

exciting topic and uh looking forward to

see you next week

and thank you very much for the

questions everyone really really good

questions and very good engagement

thank you and have a wonderful day thank

you