Probably the hottest word in tech these days is ChatGPT, a conversational bot powered by GPT, in generative AI space. It looks like they are really at some usable level, so let’s look what is out there.
Create content
Generative AI can be used for a range of activities such as creating software code, facilitating drug development and targeted marketing, but also misused for scams, fraud, political disinformation, forged identities and more. By 2025, Gartner expects generative AI to account for 10% of all data produced, up from less than 1% today.
The use cases currently under discussion include new architectures of search engines; explaining complex algorithms; creating personalized therapy bots, helping build apps from scratch; explaining scientific concepts; writing recipes; and college essays, suggests code and assists developers in autocompleting their programming tasks, among others, to increase productivity in human life/work.
Text-to-image programs such as Midjourney, DALL-E and Stable Diffusion have the potential to change how art, animation, gaming, movies and architecture, among others, are being rendered.
Landscape
Landscape map by 9 VCs: Sequoia Capital, Base10 VC, FoundationCapital, Scalevp, Antler, Leonis Capital, NFX, A16z, Dealroom
Market vendors
Risks
Job loss: large parts of the creative workforce, including commercial artists working in entertainment, video games, advertising, and publishing, could lose their jobs/2 because of generative AI models (e.g. An A.I.-Generated Picture Won an Art Prize). Even AI can pass Amazon’s software engineer interview, Google Engineer and Wharton MBA (ChatGPT answer), and other.
False and misleading content: Meta’s Galactica — a model trained on 48 million science articles with claims to summarize academic papers, solve math problems, and write scientific code — was taken down after less than three days of being online as the scientific community found it was producing incorrect results after misconstruing scientific facts and knowledge. Even worse, it can be misused to generate fake news and disinformation across platforms and ecosystems.
Toxic and bias: OpenAI acknowledges that their models can still generate toxic and biased outputs.
For pieces of work produced by LLMs, the lack of transparency around which sources of data the output is drawing on means that the answers provided by ChatGPT are impossible to properly cite and therefore impossible for users to validate or trust its output. The ramifications of such a lack of transparency and trustworthiness are particularly troubling in the era of fake news and misinformation, where LLMs could be fine-tuned to spread misinformation and threaten political stability.
Difficulty in tracing the origin of a perfectly crafted ChatGPT essay naturally leads to conversations on plagiarism 1/,2, but detectors 1, 2, 3, 4 are also being created/improved.
Sensitive information can be leaked if user sends information to Generative AI tools like ChatGPT.
Since LLMs come pre-trained and are subsequently fine tuned to specific tasks, they create a number of issues and security risks (e.g. insecure code). Notably, LLMs lack the ability to provide uncertainty estimates, which makes it difficult for us to decide when to trust the model’s output.
Summary from ChatGPT
- Bias and discrimination: Generative models are only as good as the data they are trained on, and if the training data contains biases or discriminatory information, the model may generate outputs that perpetuate these biases.
- Misuse and abuse: Generative models can be used to create fake or misleading information, such as deepfakes, fake news, and fake reviews. This can harm individuals and organizations and undermine public trust in information and technology.
- Privacy concerns: Generative models may be trained on sensitive personal data, such as images, text, or speech, increasing the risk of privacy violations.
- Unintended consequences: Generative models may generate outputs that have unexpected and unintended consequences, such as creating fake identities, generating fake financial data, or spreading false information.
- Cost of ChatGPT is 10x of Google search
- Challenge of publisher monetization of content
chatbot
ChatGPT is the hottest in this space. However, Google plans to answer with Bard.
In my last post, ChatGPT seems to really have some understanding on human languages and can accomplish basic tasks. Here is where it really shines
- Packaging, repackaging (reformatting), integrating information
- Prompting AI art models
- Good for idea exploration. Bad for reliable information retrieval
codegen
Explain code with Codex (Python, Javascript, call Word API)
Imagegen
With the understanding of language, it does make it easy for non-professionals generate images even without much skill.
stable diffusion
Dalle 2
MidJourney
Comparisons
Dall-E 2 vs. Stable Diffusion:
- If you need a higher resolution image, Stable Diffusion is the way to go. It can generate up to 1024x1024 images while Dall-E 2 is stuck at 512x512.
- When it comes to the quality, Dall-E 2 looks to be more capable than Stable Diffusion. However, the latter is more permissive with the text prompt because it allows generations of famous people like celebrities and politicians (more realistic).
- When it comes to pricing, Stable Diffusion (DeamStudio) is almost ten times cheaper than Dall-E 2. Stable Diffusion is open source, while Dall-E 2 is not.
Dall-E 2 vs. Stable Diffusion vs Midjourney
Midjourney is more artistic
GAN
Most generative AI models are based on diffusion/transformer (another explanation) architecture. But GAN models were there long time ago to generate stuff, create synthetic data to help train model without revealing individuals’ data. This helps current deep learning paradigm which needs enormous amount of data.
Business model
Appendix
Algorithm Invention
Data Augmentation
Neural Network Design
Data Synthesis
Text Generation
Image Generation
Music Generation
Artificial Creativity
AI Art: Midjourney
Video broadcasting: Nvidia Broadcast
Video editing: Descript
Voice: Adobe Podcast AI, Resemble AI
Music generation: Soundraw AI
Learning Code: Codedamn
Note generation: Notion AI
AI tools 1, 2, 3, 4, 5, 6, 7, 8, 9, AI-powered presentation software
ChatGPT
Underlying ChatGPT’s excellent text generation capability, it needs to have some general understanding of trained data. If it can have general understanding of general type of text, table data, we can use it to analyze those data.
Analyze customer feedback and sentiment, providing businesses with valuable insights into customer needs and preferences; analyze large amounts of financial data, identify patterns, and make predictions, which could help businesses make more informed financial decisions; analyze data from various sources such as weather forecasts, traffic conditions, and product demand, to optimize logistics, production, and inventory management; automate tasks such as resume screening, candidate interviewing, and employee onboarding, which can save businesses time and money; analyze data from scientific papers, patents, and other sources to identify trends and potential areas for innovation; analyze customer data and create personalized recommendations, advertisements, and offers for businesses.
Microsoft integrates Generative AI into Bing, and other productivity tools.
Transcript, search result with citation, e.g. product price history, available coupons for shoppers; use Edge to summarize key takeaway of company quarterly report and translate code from Python to Rust. Backend
Microsoft 365 AI-copilot based 3 pillars: M365 apps, Microsoft Graph, LLM, Powerpoint (generates slides, rephrase, customize images, speaker notes), Onenote (to-do list), Excel (summarize trends, refine visuals), Outlook (prioritize/summarize emails, generate reply), Teams (summarize across different sources, e.g. chat, emails, documents), Loop (bring data from different types of documents)/more info
Security Copilot
Interact with chatbot when creating application with low-code platform.
1 million users in 5 days.
100m in 2 months
Wife is always right with ChatGPT
Data analysis and data science
General programming and query language
- Create Learning Plan
- Break the Python Learning Steps into Lessons
- Learning Python Basics
- Coding Exercises and Feedback
Query Neo4J
ChatGPT competitors
IQ study of GPT3, better than the average human
- Generative pretraining (GPT-3)
- Supervised fine-tuning (trained on dataset that has prompts and answers, so the model generates responses in a way that is based on the patterns and structures learned from the labeled training data)
- Reinforcement learning from human feedback (Rewards Model, Proximal Policy Optimization/PPO)
Appstore: ChatGPT as interface
Industry
1. Streamlined drug discovery and development: Generative AI algorithms can help speed up the process of drug discovery and development by identifying potential drug candidates and testing their effectiveness in silico (i.e. using computer simulations) before moving on to the clinical trials on animals and humans.
2. Personalized medicine: Generative AI algorithms can potentially help create personalized treatment plans for patients by taking into account their medical history, symptoms, and other factors. However, this is a hypothetical benefit of generative AI and we have not yet seen a real-life case study of this yet.
Healthcare Startups worth watching, use cases 1
Logistics and transportation: Generative AI can accurately convert satellite images into map views, enabling the exploration of previously unknown locations.
Travel industry: Generative AI can help with face identification and verification systems at airports. By creating a full-face picture of a passenger from photos taken from different angles, the technology can make it easier to identify and verify the identity of travelers.
Marketing: synthetically generate outbound marketing messages, enhancing upselling and cross-selling strategies.
Create synthetic data
- boost our collective familiarity with AI and how to use it
- assist educators in preparing and reviewing sessions, e.g. tests questions, cases analysis to be discussed in class
- save educators time by automatically grading students’ assignments, providing feedback to essays
- used for training purposes, e.g. students can use ChatGPT to emulate conversations and develop their language skills and abilities through conversational interactions
EdTech 1 (what to do with ChatGPT, ban? avoid? student tool? instructor tool?), 2, 3, 4, 5, 6, 7
Financial services
Financial advisors can summarize earnings calls and create transcripts of important meetings using large language models. And credit-card companies can use LLMs for anomaly detection and fraud analysis to protect consumers, also contact center, claims processing, trading prediction, risk factor modeling use cases.
Investment research, Customer service
The business model of financial services companies uniquely requires near perfection to avoid high cost outcomes. Generative AI gives the 90% answer and financial services demands 100% accuracy (underwriting, regulatory checks (AML, KYC, BSA/AML, OFAC screening), and accounting / financial reporting.)
Applications
- Distribution: How the financial services product is sold to customers, e.g. Customized marketing, Document processing (Information verification, extraction, integration from different sources)
- Manufacturing: How the financial services product is created — once created, a financial services product is a unique instance of a pre-determined product, as agreed to with the relevant regulating authority, for a particular customer, e.g. Fraud identification, Memo writing, Risk identification and portfolio construction, Pricing and fee optimization, Product selection
- Servicing: How the financial services product is delivered to the customer, e.g. Automated relationship management, Customer service
Other industries 1
Other fun scenarios: text-based adventure game 1, 2
