Julien Simon, Chief Evangelist, Hugging Face
Use cases:
Productivity - automate manuel process. Document processing. pdf. Translate, summarize, extract insight. code generative
Chatbots - internal knowledge base, customer support agents, customer facing chatbots,
Content generation - e-mail answers, marketing copy, personalized content, image generation for marketing
HF at a glance - open source datasets (60k) and models (~300k), -> Transformers, Diffusers, Accelerate, autotrain -> Spaces
In house models - Bloom, starcoder, idefics, huggingchat, safecoder (fine tuning starcoding)
State of AI Report 2023 - https://www.stateof.ai/ - HF get a shout out.
Why HF?:
Accessibilty - anyone can use models
Transparency - full visibilty of model and training
Privacy - can self host
IP protection
Freedom of choice
IT flexibility
Grading Founding Model Providers Compliance with the Draft EU AI Act - HF came first
Work with LLM pyramid:
Using as it
Prompting - domain specific - multi-shot won't work
Fine-tuning - most reasonable - can cost $5
Initial training - domain adaption,
Working with datasets:
Image/text
Industry data - FinBERT, StarCode, Code LLaMa
Company data - product documentation / internal reports - bloombergGPT
Use case data - user data / application data - safecoder
Models are a cost, datasets are an investment
Models evolve quickly. Data stays good for years
Feb/Mar 2023 - LLaMA
May 2023 - Faclcon-40b
July 2023 - LLaMa 2 - can beat OpenAI for fine tuning - https://ai.meta.com/llama/
September 2023 - Falcon 180B, Mistral 7B (Apache), Zephyr 7B
Dont try to keep up with the model - Work with the best smallest model today - Deploy get ROI. Switch when ready.
Find the best model for each use case
AWS is the best place to train and deploy HF
Laura Edell, Next Level Generative AI Tools
Benefits
Quick product development
Improved customer experience
Legacy code advancement
Risks
Lack of transparency
Limited accuracy
Bias
IP and Copyright
Cybersecruity and fraud
Sustainability - Power consumption of training
https://www.microsoft.com/en-us/research/project/farmbeats-iot-agriculture/ - has wildfire built in
https://github.com/microsoft/farmvibes-ai
Panel
Lisa Cohen (Google, Bard)
Morgan Ingram
Steffin Harris
Julien Simon
Alex Wettreich
Work backwards from leadership goals to what tools exist to help with things
Clean, curate data, Reduce hateful content, reduce bias
AI can help with bias reduction and toxicity
Alex Wettreich, Writer, Sales
Think transformational, not incremental
Get out of the sandbox
Thinking cross-functionaly
Use cases that matter to leadership - example Victoria's secret - non-branded search key to growth stratergy, new SEO, Rewrote 6,000 product description via API in minutes.
Think through all the requirements - broad array of use cases, integration with existing apps, fast time to value. - Workflow (high, low), requirements (high, low).
Indentify all the business impact - Cost savings - productivity gain and headcount avoidance, reduced cost to serve customers. Incremental revenue - higher conversion rates/win rates, faster time to market, net new programs that become viable, impact of unrelated initiatives that become possible due to unlocked capacity.
Lay out your plan to mitigate risks - Will you be able to cite sources of claims? How will build AI guardrails. How will it get adopted
Ryan McClelland, NASA, Evolved Structures: Generative Design and Digital Manufacturing at NASA
Gen AI for materials for space flight
"AI is one of the most important things humanity is working on. it is more profound thatn, I dunno, electricty or file" Sundar Parchi (Google)
AI based design e.g. chair in the shape of an avacado
Create a prompt - encode (1 hour) e.g. image and description -> evolve design -> fabricate the part
EXCITE - satellite to view exoplanets - part tip/tilt assembly. Human process iterate (may have taken two weeks)
Pitfalls and barriers to adoptions - skills gap, cultural intertia, fear of being replaced by AI, Garbage in/Garbage out - human validation
Venu Vasudevan, P&G, GenAI: Thinking Big, Acting Small
How do use AI with untidy data
Data is messy e.g. how long do you brush your teeth? people say 2 minutes, small actual data 47 sections. AI should work on the small sample. It's can make a difference on how you design toothpaste.
Hedonic nature of design
Motion GPT - capture data re. showing with privacy.
Lisa Cohen, Google Bard
Improve time-to-value to search, navigate and extract insights from understanding large amounts of complex data
App integration - e.g. maps, flights, hotels, youtube. Summarize a video
Translate code from one language to another, debug, secruity threats, generate documentation and tutorials, explain, translate
Citations and user feedback
Test and tune - safety thresholds,
Capacity planning - cost management and forecast
Charlotte Munro, Booking.com - AI Trip Planner
ML for search, personalisation, content, trip planning
Shift from being a place where you make your booking to where you search for your vacation
Task force team did the week in 10 weeks.
Use ChatGPT3.5 + latest booking.com documents
Women in AI panel
Stefanie Khan - UPS
Lisa Cohen - Google
Wendy Zhang - CSL Behring
Laura Edell - Microsoft
Richard Halkett, SambaNova Systems, Pioneering in the era of pervasive AI
Customer Ops
Sales & marketing
Product R&D
Finance
Risk & Legal
HR & IT
Supply Chain & Ops
Composition of experts (CoE) - Base LLM an model chooses a model to ask to
Ramesh Menon, Defense Intelligence Agency, Explainable and Responsible AI Principles aligned to the US Constitution
Exploring the Legal Considerations of Generative AI - Ownership vs. Innovation
Amir Ghavi - Fried Frank
Scott Sholder - Cowan DeBaets Abrahams & Sheppard LLP
Franklin Graves - HCA Healthcare
~Can only copyright human works. Not really AI works
Copyright offices are behind the real world nuancies
But Iphone Portrait mode is AI yet copy writable
Training
GitHub Co-pilot is getting sued
Anita Western, LexisNexis - The Disruptive Potential of Generative AI: A legal Industry Revolution
Ask/Search, draft, summarize, extract
Consider real world impact
Prevent unfair bias
Conor Grennan, NYU Stern School of Business - Unlocking Generative AI: Retraining Our Brains to Maximise ChatGPT
tool - ChatGPT
behavior - Us
Mohamad Houri, AirPlus International - Starting you first Gen AI project in a corperate setting
Priya Shivakumar, Lightning AI - Gen AI and Lightning Accelerating AI Innovation while ensuring Scalability and Security
NeurIPS LLM efficience challenge - 1 LMM + 1 GPU + 1 Day = 1 uses lit-gpt
Use fabric for LLM optimization
Pytorch lightning to pytorch is like react js to html
Pytorch lightning framework
powers stable diffusion, open fold, nemo (nvidia)
Noele Roque, Manulife - Crafting a strategic blueprint for operating model transformation
What is out Gen AI ambition?
Now, optimization - quick wins & low hanging fruit. Employee productivity & copoilots, knowledge management and agent assist chatbots
Next, Differentiation - Augmentations over current AI solutions
Later, new business models - new revenue streams, new customers
What is our starting point?
What is our path forward?
Critically of a cross-functional co-creation approach
Increased importance of partnering iwth technology provides
A shared understanding of suitable use cases and value drivers
A consistent approach to use case definition and priotitization
A clear methodology to cost modeling
Agility in decision-making
Anticipate and address questions about ethics, explainability, human agency and other responsible AI topics
Stefanie Khan, UPS - Innovating your strategy from the ground up
Invest In & Upskill High-potential Employees
Evolve Maturity & Adoption of Technology
Unlock the value of data and anlytics
empower teams to think critically
Enable a citizen data science program at UPS
Baby steps, start small. Bring everyone along for the ride, Colloboration is key
Vivek Jetley, EXL - Making Generative AI Real for the Enterprise
Data - the hidden part of the Gen AI iceberg
Internal - code gen, conversation BI
Middle-office (expose to someone aiding customer) - knowledge base, customer
Customer facing - marketing, chat bot
Enterprise data management and migration, AI workbench, secruity and complience, plug and play gen AI accelerators, human-in-the-loop
Smart agent which uses gen AI during calls to help customers e.g. sympathetic
Daniel Hulme, WPP - Rethinking and mastering AI: impact on business and humanity
goal directed adaptive behavior
Task automation - macros,
Content generation - image, video, text, music
Human representation - deepfakes, voice, personas
Insight extraction - ML, DS, analytics
Decision making - optimization
Human augmentation
"Automoney, Mastery, Puporse" - talent managing, Daniel Pink, Drive
singularity - can't see beyond it
Political - we no longer know what is true. e.g. voice cloning. safe words can help
Environmental - ecological collapse
Social - life entension
Technological - create superintelligence
Legal - surveillance
Economic - job lesses