Pacific Gas and Electric Company

1. ”Generative” means “producing:” (“creative,” “originative” or “procreative”) in short, this new type of AI creates new digital realities, such as pictures, music, or video. Until this point, we have enjoyed AI’s ability to develop machine intelligence that predicts an outcome or automates tasks. For instance, we have used AI to anticipate events such as weather developments (through climate pattern simulations), or to use virtual, voice- controlled assistants (such as Alexa or Siri) to simplify everyday tasks by setting reminders, playing music, and answering questions. These advanced AI

implementations are ancillary to a particular fact or reality, such as anticipating a weather status that is already in process (like a forming storm or tornado) or retrieving existing music on one’s Amazon playlist or information on the web to answer a question.

Conversely, in Generative AI algorithms produce new content: they create novel images, paint pictures, write songs or poems, and compose and play music they’ve created from scratch. Some algorithms go as far as getting input in one media and producing output in another media: DALL-E, for instance, gets text as content input to create images depicting that content.

2. Generative AI creates new digital objects by extracting and applying patterns from existing data samples: thus, creating new data after learning attributes and rules

inferred from real data inputs. Let’s dig a bit deeper into this. Up to this point, the most widely used AI models take an existing data sample to anticipate a reality in development through performing tasks like categorizing an event as fraud or grouping similar items into recommended/similar product lists. Analytics teams in many companies have successfully applied these techniques to predict defective products or assets (by using historical data on what a failure looks like), or identify factors related to root cause of unwanted events (such as assembly line downtime), among others. These techniques are called “discriminative AI” because they, precisely, “discriminate” or distinguish patterns belonging to specific entities or facts (such as what makes a product fail), differentiating those from all other ones in a dataset (i.e. non-failing products). In contrast, Generative AI models use those learnt patterns from real data to create new (sometimes “unreal”) data. Because generative models learn those underlying attributes of data (like harmonic patterns in a song) to apply them for creating new data (i.e. a new song), we call this form of machine learning “self-supervised”—i.e. learning independently from humans and as part of the generative process (such as learning how harmonic melodies are made up by identifying harmony structures).

  ​Generative AI creates new digital objects by extracting and applying patterns from existing data samples: thus, creating new data after learning attributes and rules inferred from real data inputs  

3. The generative AI model solution group is under development and steadily growing, with plenty of clever acronyms already making up the jargon family, the most popular now being ChatGPT. Let’s review briefly some of the trendiest ones (at the moment of writing):

● ChatGPT (by OpenAI): dedicated to all types of text tasks: summarization, translation, explaining complicated content in plain English, writing poetry or songs, etc. Google has released a competing product called BARD. Currently, the set of generative AI models working on language tasks are called “Large Language Models” or LLM.

● DALL-E (OpenAI product as well; pronounced “dal-ee”; very clever indeed): generates images from text inputs. The user types the image they want (say, “draw a group of diverse people sharing information in a meeting”) and after a few seconds DALL-E creates an image representing that.

● AIVA: music composer that creates original music using either pre-defined styles (such as “modern cinematic”, pop, rock, “fantasy,” or jazz) or a personalized style extracted from any file uploaded by the user.

● LUMEN: video content creator that converts blog posts and podcasts to video. It also adds call-outs and captions. SYNTHESIA is a competitor.

● JASPER AI: comprehensive suit of generative AI models to create all types of digital on-brand content.

● ARTBREEDER: collaborative website where users can generate and modify images of faces, landscapes, and paintings, among other categories.

● GITHUB COPILOT: creates code automatically, composing entire functions in real time, debugging, etc.

4. Best applications are for creative outcomes: because of the free (randomized, actually) way in which generative AI produces new content, it offers a perfect

sandbox for anything that involves creativity and freedom from existing realities. For instance, your company’s product design creatives can use DALL-E to input text describing new features of an existing product and get a picture of how the product would look like. Marketing and communication teams can use ChatGPT or Bard to create product taglines or jingle songs with catchy rhymes. And so on…

5. When applied to real products/processes and facts, close monitoring is needed: precisely because of the creative (in fact: random, remember?) nature of generative models, they can be greatly detached from factual/reality outcomes. We said before that generative AI can create unrealistic data; this happens when the new data created follows patters from real data, but without representing any actual object or event in reality. For instance, ChatGPT reportedly offers inaccurate facts, biased answers, and incorrect information. This happens because these models create content from patterns in a generalist way, without close human validation based on specific cases/content. In fact, if your teams decide to experiment with generative AI technologies, such as Large Language Models, it is imperative to have some form of human-in-the-loop (ideally subject matter experts) continuously checking model outputs for veracity. For instance: if workplace safety teams for your company use ChatGPT to generate written guidelines on a particular topic (i.e. safety when handling hazardous chemicals) automatically, a team of specialists on safety content should review the answers for accuracy. Most generative AI models are not mature enough to match reality accurately, which presents optimization opportunities for analytics teams in companies. Beyond optimization, the use of open-source generative AI models poses serious threats to cybersecurity as well as information security and confidentiality, since sensitive information uploaded in those platforms are used by third parties to develop models further and could also be accessed by anyone accessing the platform. Finally, and specifically referring to Large Language Models, hacking is a real risk since generative AI models can write code that interacts with code given by a user without human control.

Potential applications: Generative AI offers multiple possibilities of adding value to your companys strategic goals. Many potential applications to different industries are offered in multiple web forums. For our purposes, we will divide the main ones according to maturity level, as a feasibility proxy for that industry:

● Ready to use: creative industries such as Advertising, Music Production, Film and Video Production, Creative Writing, Game Design, Digital Design.

● Use with close human supervision: to draft design mocks before prototyping or as a form of simulation: Architecture, Automotive (Design and Autonomous Driving Systems), Fashion Design, Pharmaceutical Research (to investigate new drug compounds and their interactions), Finance (Algorithmic Trading, Forecasting and Risk Management Simulations).