Slide deck from the BrightGen webinar to help organisations start thinking about what they need to do to get ready for AI Cloud, or generative AI in general.
by Generative AI • 51% are aware of the latest AI trends • 49% are not concerned about the potential for fake news • 34% are worried about use in phishing attacks • 33% worried about copyright issues • 27% worried it will be used to copy competitors designs 19/07/2023 6 Capgemini Research Institute - Why consumers love generative AI : https://www.capgemini.com/insights/research-library/creative-and-generative-ai/
the next 18 months • 33% consider it a top priority • 33% believe it is over-hyped • 80% who believe it is over-hyped still believe it will help • 71% believe it will introduce new security risks • 66% believe their employees don't have the skills • 99% believe their business must take measures to be successful 19/07/2023 7 Salesforce – IT Leaders Call Generative AI a ‘Game Changer’ but Seek Progress on Ethics and Trust : https://www.salesforce.com/uk/news/stories/generative-ai-research/
nobody is in charge, everybody is in charge • Who is accountable/responsible? Publish guidelines • Data controls • Approval • Acceptable use cases • Training Encourage collaboration 19/07/2023 9
model, unable to detect nuance • Overfitting – finely tuned to training data, memorizes rather than learns Too much data • Concept drift – outdated view • Curse of dimensionality – fields that aren't relevant skewing the model • Data saturation – all values present Bias • We'll cover this later. 19/07/2023 12
are guilty until proven innocent o How did it arrive at that answer? o Do citations/references exist, are they relevant • Create a clear explanation • Share this with your stakeholders – leadership/users/customers • Encourage engagement, so interested parties can: o Understand how it works o Understand how it benefits them o Provide input and guidance to improve 19/07/2023 15
about how you will use people's data o Allow them to opt out • Explain how your algorithms work o Identify limitations o Call out potential bias 19/07/2023 16 "Humans need to be aware that they are interacting with an AI system, and must be informed of the system’s capabilities and limitations." https://digital-strategy.ec.europa.eu/en/library/ethics-guidelines-trustworthy-ai
an individual or group based on their characteristics" https://towardsdatascience.com/understanding-bias-and-fairness-in-ai-systems-6f7fbfe267f3 Define what fair means to your organisation Weed out bias • Bias – certain elements are overweighted or overrepresented Bias leads to • Skewed outcomes • Low accuracy • Systematic prejudice 19/07/2023 17
bias Reflects social inequities Amplified by machine learning 19/07/2023 18 Example - Amazon CV screening project, 2014 • 10 years of historic data • Most employees male • Almost all technical employees male • Men flagged as suitable candidates • Scrapped in 2015
the population you intend to serve Sample set not randomised Limited set of volunteers (self-selection) 19/07/2023 19 Example – Nottingham NHS maternity enquiry • Letters sent in English only • Technical legal language • 250 white respondents • 20 Black/Asian respondents
confirm pre-existing hypothesis (confirmation bias) Keep tuning the model until desired outcome achieved (experimenter's bias) 19/07/2023 20 Example – training model for positive sentiment • Thumbs up considered positive • True for large parts of the world • Offensive in parts of West Africa and Middle East
are too broad 19/07/2023 21 Example – aggregated salary versus years of experience • True for IT, Legal, Finance, Consulting – large cohort • False for professional sport – small cohort • Aggregation biased against pro sports
schools, location Give all potential participants an equal chance to respond Look for outliers that will have an outsized impact Identify use of default values Internal and external review Don't oversanitise • Stops reflecting original values 19/07/2023 22
to humans! Impact depends on context • Blog on rugby tactics – low • Self-driving car – high • Medical diagnosis - catastrophic Mitigation: • Verify! • Incorporate context into prompts • Restrict the response – simple yes/no 19/07/2023 23 Lawyer Steven Schwartz, lawsuit vs Avianca airlines • Used ChatGPT for research • 6 cases cited • Included LexisNexis/Weslaw references • Did not exist
Even AI powered! • Secure data, model and inputs At least as reliable as existing systems: • Scalable • Performant Dependable • Consistent output, even under stress/attack • Fail gracefully and predictably 19/07/2023 24
of fresh water • 4,000 – 200,000 Kg CO2 GPT-3 Query • 10-20 ml water • 0.002 - 0.004 KWh electricity 19/07/2023 25 https://interestingengineering.com/innovation/training-chatgpt-consumes-water https://towardsdatascience.com/chatgpts-electricity-consumption-7873483feac4 https://shrinkthatfootprint.com/carbon-footprint-of-training-gpt-3-and-large-language-models/ "By 2025, without sustainable AI practices, AI will consume more energy than the human workforce, significantly offsetting carbon-zero gains." https://www.gartner.com/en/articles/gartner-top-10-strategic-predictions-for-2023-and-beyond
of AI • Audit use of data and AI • Create training program • Handle issues or complaints raised • Document how AI is being used Diverse membership Executive sponsorship 19/07/2023 26
by whom the arrangements necessary for the creation of the work are undertaken" Does copyright on input training data give rights on output? • Maybe, if it replicates and competes with the original creator Diligently document your output • Make clear what was generated by humans vs AI Legal implications are complex and uncertain 19/07/2023 30
Based on existing data • Identifies patterns, predicts words Pre-trained • Trained on existing, large dataset • Tuned for specific tasks Transformer • Neural net architecture • Good at context, parallelisation 19/07/2023 33
without custom built algorithms No need to program the model Machine learns from data • Supervised learning – human gives examples/labels input • Unsupervised learning – finds patterns in data • Reinforcement learning – "earns rewards" for achieving goals Targets specific tasks – predict / classify 19/07/2023 34
human intelligence • Perform complex tasks • Learning, reasoning, self-correction • Handles unstructured data GPT • Predicts the next word in a sentence • Based on provided words and training data • Appends to text • Repeat until stopping condition encountered 19/07/2023 35 https://org62.my.salesforce.com/sfc/p/#000000000062/a/3y000000iOrb/li90p54YaGginl_Ajb1a7z5Tg81rJLkaWQQB4Rdy1Bg
encryption Dynamic Grounding • Additional information specific to use case Data Masking • Replace PII in request with similar, inauthentic data. Reverse in response. Toxicity Detection • Filtering harmful/inappropriate content Audit Trail • Log actions 19/07/2023 40
trust • Your company is still responsible and accountable • Understand how Einstein GPT Trust Layer works in depth General • Verify accuracy of output • Make sure everyone involved is aware they are using AI • All actions taken by humans – AI assisting • Not 'one and done' - models learn, retest regularly 19/07/2023 41
how to get the most out of generative AI • 48% worry they will lose their job if they don't learn • 53% of sales reps don't know how to get the most out of generative AI • 39% worry they will lose their job if they don't learn Currently using Generative AI 19/07/2023 45 https://www.salesforce.com/uk/news/stories/sales-service-research-generative-ai SALES 35% MARKETING 51% SERVICE 24%
/ AI • How specific tool works • What tasks it is safe to use for • Risks associated with use • How to evaluate output • Where to go with concerns • Never use real data to learn 19/07/2023 46
to talk to customers • Understand your products • Handle customers that AI can't • Recognise accurate output Keep training juniors, or there will be no seniors 19/07/2023 47
• Creates early drafts, not final versions • Can't do the creative work, just the mundane, repetitive tasks • Staff can work at a higher level • Humans need to understand, correct and approve content Not less staff, the same staff doing more Just because AI can replace humans, doesn't mean it will 19/07/2023 50
Quest : bit.ly/3NSN7Wt Building Ethical and Inclusive Products : bit.ly/46WqIQI Data Ethics and Trusted Personalization : bit.ly/3pWXfp6 Webinar: Unlock ROI with Generative AI and GPT : bit.ly/43F3sUM DeepLearning.ai Short Courses : https://bit.ly/46UB3N5 19/07/2023 53
keirbowden
nishiuma
aki_moon
yanaemon
rishigami
ysknsid25
ma2shita
yukukotani
hanacchi
abnoumaru
oracle4engineer
openwall
jcasabona
jennybc
caitiem20
carmenintech
addyosmani
tanoku
aleyda
brad_frost
ddemaree
pedronauck
chrisshort
chriscoyier
