AI 2023.pdf
- 1. AI (and ChatGPT) DAVID M. CIESLAK, CPA.CITP, CGMA, GSEC RKL ESOLUTIONS, LLC 1
- 2. David M. Cieslak, CPA.CITP, CGMA, GSEC RKL eSolutions, LLC - EVP, Chief Cloud Officer Frequent speaker for AICPA, CalCPA and numerous accounting industry groups and conferences Named one of Accounting Today’s 100 Most Influential People in Accounting 20 times CPA Practice Advisor – 2011-23 “Top 25 Thought Leader,” Hall of Fame Inductee – 2020 LA Business Journal Top 100 Accountant - 2022 AKA “Inspector Gadget” 3 dcieslak@rklesolutions.com Phone: 717-409-8835 www.rklesolutions.com Twitter: @dcieslak
- 3. (www.rklesolutions.com) Subsidiary of RKL LLP (#59 in the Top 100 CPA firms) ERP sales and consulting since 2001: Sage Intacct, Sage 100, 500, X3 100+ employees in 22 states Sage Intacct success Rookie of the Year (2012) Premier Partner (2015 – 2022) Growth Partner of the Year (2022) 12 full-time Sage Intacct consultants (and growing) 125 implementations in the last 3 years
- 5. 7
- 6. 8
- 10. Robotic Process Automation (RPA) - Definition Non-electromechanical “Robotics” or “bots” / software that is programmed to mimic the keystrokes humans make in completing a process. 12
- 11. RPA - Examples • Employee onboarding • Password resets • Account reconciliation • Compliance reporting • System integration • A/P invoice entry • Claims verification • EFT or ACH processing • Rekeying data (“dirty” interface) – no APIs • Add missing functionality to a system • One person enters data from multiple systems • Payroll tax deposits • Account verification • Customer service • Social media review • Web data collection • Document management • Industry reporting • Chat bots • Review all files for a certain issue • Convert unstructured data to structured data • Manual checking, decisions and calculations 13
- 12. RPA Challenges Process – Optimizing a sub-optimal process flow (vs. replacing with more capable applications and/or process)? Risk – If poorly implemented or controlled, can introduce security and/or internal control risk. Often done at the desktop level, and may escape proper organizational control and governance. Zero-sum – Time & efficiency gains offset by monitoring, security & assurance requirements. Knowledge loss – will users still understand process if automated? 14
- 13. 2019 trends
- 14. AI – Enabling our lives today 16
- 15. Adobe Acrobat Liquid Mode view – automatically resizes documents for mobile devices (page count, table of content) Processing done on Adobe’s Document Cloud servers – for display only, does not change original doc. 17
- 16. Salesforce - Einstein Discovery – discover relevant patterns in your data. Prediction Builder - Predict business outcomes, such as churn or lifetime value. Create custom AI models on any Salesforce field or object with clicks, not code. Next Best Action - Define recommendations, create action strategies, build predictive models, display recommendations, and activate automation. Bots - build, train, and deploy custom bots on digital channels that are connected to your CRM data. Vision - see the entire conversation about your brand on social media and beyond. 18
- 17. Ensure accuracy as transactions are recorded into the system Proactively catch errors • Leverage GL approvals • Score outliers using machine learning • Call attention to what requires a closer look This combination is unusual – Account, Department, Location, Amount Sage Intacct - GL Outlier Detection 19
- 18. MindBridge AI Auditor 100% audit with risk scoring via advanced AI and Machine Learning algorithms Assessed against 28 control points Identified risk areas include ◦ Monetary flows ◦ Manual entry ◦ Unusual coding ◦ Period over period ◦ Industry Out of the box integration with leading ERP systems 20+ standard financial reports & ratios 20
- 20. Definitions AI (Artificial Intelligence) - ability of computer systems or machines to imitate human-like intelligence/behavior; Includes the ability to perceive, reason, learn, and make decisions ◦ Artificial narrow intelligence (ANI)- Focused. Narrow range of abilities ◦ Examples: Smart voice assistants, spam filters, RPA, TV show recommendations….and ChatGPT! ◦ Artificial general intelligence (AGI) - On par with human capabilities. Also called Strong AI or Deep AI. Mimics human behavior. Includes Common Sense, Background Knowledge, Transfer Learning, Abstraction & Causality. Not limited to single domain or tasks. ◦ Artificial superintelligence (ASI) / Singularity - hypothetical point when artificial intelligence surpasses human intelligence and becomes unstoppable. AGI & ASI are theoretical, i.e. not yet possible Many experts are very worried about AGI – Elon Musk warns: “AGI is humanity’s biggest existential threat. Efforts to bring it about, are like “summoning the demon.”
- 21. Definitions (Cont.) *Generative AI - type of artificial intelligence that uses unstructured deep learning models to produce content based on user input. Content includes written materials, images, video, audio and music, and computer code. LLM (Large Language Model) - a massive database that has been trained on a vast quantities of data (in the case of ChatGPT, the entire internet through 2021) to produce human-like responses to dialogue or other inputs. LLMs make use of deep learning models to process, analyze, and make predictions with complex data. Common data sources include: ◦ Literature ◦ On-line content ◦ News and current affairs ◦ Social media
- 22. 24
- 23. Definitions (Cont.) RLHF - advanced approach to training AI systems that combines reinforcement learning with human feedback. It is a way to create a more robust learning process by incorporating the wisdom and experience of human trainers in the model training process Chatbot - (also called AI writer) refers to a type of artificial intelligence-powered program that is capable of generating written content from a user's input prompt.
- 24. 26
- 25. Open AI - ChatGPT Interactive, conversational chatbot Stands for “Chat Generative Pre-Trained Transformer” Overnight sensation – over 100M users in under 2 mos. Conversational – so results can be refined. More detailed questions (prompts) yield more specific responses. Cost: ◦ GPT-3.6 – free to the general public (US only) ◦ GPT-4 – $20/mo. Full access to newest version, even during peak times ◦ Business subscription coming (data won’t be used to train models) Microsoft has pledged more than $10B to OpenAI (49% stake). Native integration with Edge (Power Bing) and Office (Copilot) applications. Other OpenAI solutions: Whisper, DALL*E 2, CLIP, Jukebox https://openai.com 27
- 26. GPT-4 Data through Aug-22 Multi-modal, i.e. works with both text and images. Example: GPT-4 can describe the content of a photo, identify trends in a graph, or even generate captions for images Fun use case: generate recipe based on pic of what’s in the refrigerator. Output up to 25,000 words of text Complex problem solving Reduction of biased / inappropriate responses Aware/responds to emotions expressed in text Sentient or sense of humor? No, but able to mimic more convincingly
- 28. Prompts Prompt engineering is the process of designing and optimizing prompts for AI language models, such as ChatGPT, to generate high quality responses. New books and training courses arriving daily “Prompt Engineers” making $300K? External marketplaces & repositories already available – some free, others for a fee. Eg. ◦ Github ◦ PromptBase ◦ Arvin
- 29. ChatGPT Prompts According to ChatGPT, the anatomy of a prompt is as follows: 1. Act Instruction – specify role & expertise of AI 2. Context – describe situation, setting, or topic 3. Task or Question - define the specific objective or inquiry the AI is expected to address 4. Constraints or Limitations - set any boundaries or conditions the AI should consider while generating a response 5. Additional Guidance - provide further instructions, like tone or formatting, to fine-tune the AI's output
- 30. ChatGPT Prompts A more basic approach: 1. Talk like you would a person 2. Set the stage and provide context 3. Tell the AI to assume and identity or profession 4. Keep AI on track (re-direct and/or challenge responses)
- 31. ChatGPT Prompts You can add more aspects to ensure more accurate and nuanced responses: specify the desired length request a particular tone of voice request examples/analogies incorporate multiple perspectives cite sources/reference materials address potential misconceptions/pitfalls
- 32. ChatGPT Commands Continue Elaborate Summarize List Compare & Contrast Pros and Cons In Simple/Layman's Terms Act As Imagine Clarify Step by Step Brainstorm Rephrase Rank Devil's Advocate Roleplay Translate Retrofit Critique Troubleshoot Analogous
- 34. ChatGPT Plug-ins
- 35. ChatGPT API ChatGPT API is different from ChatGPT plugins. The API brings ChatGPT's tools to other sites, whereas the ChatGPT plugins take other sites and add their functionality into ChatGPT. ChatGPT API uses GPT-4 rather than GPT-3.5, so apps using the ChatGPT API could be more powerful and have greater functionality than the free version of ChatGPT API data usage policy – end user data not used to train underlying LLM
- 36. Edge & Bing search integration Visual input & output Bing search engine in MS Edge integrated with ChatGPT-4 (newest LLM), and results better due to use of “Prometheus Model” (merges search and chat data) Saves chat history Avail for all – no wait list 16% bump in search traffic 38
- 37. Microsoft 365 Copilot ChatGPT integration with Office suite Prompts from/to Copilot filtered through Microsoft Graph Sample uses include: ◦ Excel – VBA macro coding or data visualization ◦ Word – first draft to edit and iterate on ◦ PowerPoint – presentations ◦ Outlook – summarize email threads and respond to emails 39
- 38. 40 ChatGPT for iOS (by OpenAI)
- 40. Bard Google announced internal code red in Jan-23; Bard avail beginning Mar-23 Considered “experimental” Uses Google’s own LaMDA (Language model for dialogue applications) model. Moving to PaLM 2 (Pathways Language Model 2)- better w/ math, logic, security & medical. Trained on over 100 languages. Not multi-modal yet. Less tech-heavy responses Scrapes Google search results daily (more up to date responses) Integration with Google search – avail?? https://bard.google.com 42
- 41. Google Workspace AI-powered features avail. to “Trusted Testers” in Docs and GMail ◦ draft, reply, summarize, and prioritize your Gmail ◦ brainstorm, proofread, write, and rewrite in Docs ◦ bring your creative vision to life with auto- generated images, audio, and video in Slides ◦ go from raw data to insights and analysis via auto completion, formula generation, and contextual categorization in Sheets ◦ generate new backgrounds and capture notes in Meet ◦ enable workflows for getting things done in Chat 43
- 42. Amazon Announced series of generative AI services on April 13th Hosted on AWS: ◦ Bedrock – API accessible foundation models ◦ CodeWhisperer – AI based coding assistant 44
- 43. Stanford Alpaca AI Stanford research team started with Meta's open- source LLaMA 7B language model (the smallest and cheapest of several LLaMA models available). Then asked GPT to take 175 human-written instruction/output pairs, and start generating more in the same style and format, 20 at a time. This was automated through one of OpenAI's helpfully provided APIs. In a short time, team had some 52,000 sample conversations to use in post-training the LLaMA model. Cost less than US$500. Then, used that data to fine-tune the LLaMA model – process took about three hours and cost less than US$100. Note: OpenAI says this violates their terms of service, but fully expect others will do similar
- 44. Many more coming soon! 46
- 45. Auto-GPT Open source “recursive AI” application, using GPT-4 Able to carry out more complex, multi-step procedures than existing LLM-powered applications by creating its own prompts and feeding them back to itself, creating a loop. Breaks larger tasks into smaller sub-tasks. Original instance acts as “project manager.” https://autogpt.net 47
- 46. Impact of Generative AI “More impactful than the printing press”?! Report from analysts at Goldman Sachs. ◦ Two-thirds of Euro/American jobs are set to change due to AI automation, and up to a quarter of all current work will be taken over by AI. ◦ AI could eventually increase annual global GDP by 7%. 80 per cent of the US workforce could have at least 10 per cent of their work tasks affected by GPTs Knowledge workers most impacted Staffing shortage relief? “Fighting against Generative AI is like fighting against calculators!”
- 48. 50
- 49. Impact of Generative AI Possible immediate uses by accountants: ◦ Technical research / documentation ◦ Summarize PDFs ◦ Excel formulas, VBA automation, coding ◦ Transaction ledger entry & error detection ◦ Financial analysis and data rendering (charts & graphs) ◦ Forecasting / predictive analytics ◦ Fraud detection / risk mgmt ◦ Tax compliance Other significant use-cases ◦ Cybersecurity
- 50. Generative AI Concerns Use/Misuse ◦ Can be used for good, or nefarious purposes – cheat on schoolwork, malware/phishing scams, deepfakes, etc. ◦ Job augmentation vs. replacement? ◦ Proliferating web-sites Application ◦ Prompts can be difficult to write ◦ Prone to "hallucinations," fabricated responses that sound plausible but are inaccurate. Includes citing books/articles that actually don’t exist! According to Sam Altman, OpenAI CEO – “It's a mistake to be relying on it for anything important right now.” ◦ Only as good as the LLM, i.e. limited data set (ChatGPT data thru 2021), bias baked-in, etc. ◦ Copyright considerations – original vs. proprietary content? (demand citations!) 52
- 52. Generative AI Concerns Environment ◦ Insecure environments – vulnerable to compromised apps/extensions ◦ Risk of exposing confidential data based on additional information fed into chat bots (searchable by others). ◦ Chat GPT “DAN” mode (“Do Anything Now”) - “jailbroken” versions that allow users to go against ChatGPT’s guidelines. Existential threat to humanity!? 54
- 53. Future of Life Institute (FLI) Non-profit organization that aims to reduce global catastrophic and existential risks facing humanity, particularly existential risk from advanced artificial intelligence (AI) Recently published open letter, signed by thousands of AI researchers and concerned others, including Apple cofounder Steve Wozniak, SpaceX, Tesla and Twitter CEO Elon Musk; Stability AI CEO Emad Mostaque; Sapiens author Yuval Noah Harari; and Yoshua Bengio, founder of AI research institute Mila. Citing “an out-of-control race to develop and deploy ever more powerful digital minds that no one — not even their creators — can understand, predict or reliably control,” the letter called for a 6-month pause in the development of anything more powerful than GPT-4 Additional time would allow ethical, regulatory and safety concerns to be considered and states that “powerful AI systems should be developed only once we are confident that their effects will be positive and their risks will be manageable. 55
- 55. Predictive AI - Definition Predictive AI ◦ Studies historical data, identifies patterns and makes predictions about the future that can better inform business decisions. ◦ Can detect data flow anomalies and extrapolate how they will play out in the future in terms of results or behavior; enhance business decisions by identifying a customer’s purchasing propensity as well as upsell potential; and improve business outcomes. 57
- 56. Predictive AI vs. Generative AI Both generative AI and predictive AI use artificial intelligence algorithms to obtain their results. Key differences include: ◦ Creativity – generative AI is creative and produces things that have never existed before. Predictive AI lacks the element of content creation. ◦ Inferring the future – predictive AI is all about using historical and current data to spot patterns and extrapolate potential futures. Generative AI also spots patterns but combines them into unique new forms. ◦ Different algorithms – generative AI uses complex algorithms and deep learning to generate new content based on the data it is trained on. Predictive AI generally relies on statistical algorithms and machine learning to analyze data and make predictions. Bottom line – Generative – more creative, Predictive – more analytical 58
- 57. What’s Next (already here)? Use cases ◦ Medicine ◦ Personalized medicine ◦ Disease diagnosis ◦ Drive-thru AI ◦ AI lawyers (verify references!) Platform ◦ GPT-5 ◦ AGI(?) ◦ Avail late 23/early 24 ◦ Quantum computing Other ◦ Humanoid robots 59
- 58. OpenAI Superalignment team Led by Ilya Sutskever (Chief Scientist & co-founder) Predict AI with intelligence exceeding that of humans could arrive within the decade. Won’t necessarily be benevolent, to researching ways to control/restrict Team aims to solve the core technical challenges of controlling superintelligent AI over the next four years (currently cant steer or control superintelligent AI, and stop it from going rogue). 60
- 60. Connect on LinkedIn and Twitter David Cieslak, CPA.CITP, GSEC EVP, Chief Cloud Officer RKL eSolutions, LLC @dcieslak linkedin.com/in/davidmcieslak