Why OpenAI's new ChatGPT will change the future
OpenAi is an artificial intelligence software and platform company with its third version of GPT, the conversational AI platform. This platform has generated incredible results, and Freeburg has spent the last two days testing it out. He asked it to write a script of Sacks and Jaeckel talking about Artificial Intelligence in a Quentin Tarantino movie style. The result was a PG-rated version of a Tarantino script, with no cussing or violence.
The Impressive Results
Freeburg shared the script with us, which opens with David Sacks and Jason Calacanis sitting at a dimly lit table in a smoky bar. David Sachs expresses his amazement at the level of advancement of this conversational AI system. Jason agrees and warns of the potential dangers of the technology. David Sachs needs to be more convinced, pointing to the importance of regulation in using AI.
The Three-Step Process
OpenAi explains the three-step process for training GPT 3.5, the intermediate model for GPT 4.0. It begins with collecting data, creating a supervised system, and finally optimizing it with feedback. This model is only 100 gigabytes and is incredible for its size.
Impact of GPT
GPT 3.5 has the potential to replace many knowledge-based human roles, such as homework help, software engineers, salespeople, copywriters, customer service, and more. This could start a flurry of 100,000 startups. The possibilities are endless with this technology, but regulation and responsibility are key to ensuring it is used for good.
The Rise of Generative AI
The attention of investors and tech enthusiasts is shifting toward a new type of Artificial Intelligence (AI) called generative AI. This technology is poised to create the next "hype cycle" or "bubble" of development in Silicon Valley, as the possibilities for its application across industries and applications are almost endless.
How Generative AI Works
One of the most technical aspects of generative AI is its ability to distinguish between different words. For example, how would it know the difference between "your anus" and "Uranus"? It will learn this through AI-powered algorithms that can process natural language.
The Disruption of Search
Generative AI’s potential to disrupt the traditional search box used by Google, and other web crawlers, is of particular interest. Currently, data is gathered, indexed, and made available on the search page in either a structured or unstructured way. Google has already created a product called the "one box" which can take structured data, such as "what is the weather in San Francisco today," and present it at the top of the search result page.
A Shift to Natural Language
Putting generative AI in competition with traditional search engines could lead to a shift towards natural language chat interfaces. This is why Google purchased DeepMind, a collection of human-powered search engines. It has already been applied to ads, ad optimization, as well as the ranking of YouTube videos.
Models as a Service: Replacing SAS with Mass
The first significant change software will see, particularly in the enterprise, is the replacement of SAS (Software as a Service) with “Mass” (Models as a Service). Instead of relying on SAS companies, companies can use language models such as GPT-3 to take care of functions such as expense management or forecasting.
Limitations of Current Models
Although these models can be useful, they remain brittle. They are good at one thing and are a single-mode way of interfacing with data. To overcome this limitation, the next step is to create a multimodal model that can combine video, voice, and data to answer more substantive problems.
GPT-3: Cute but Not Complete
GPT-3 is a popular example of language models, and while it can be impressive, it is not yet at the point of providing precise answers that are overwhelmingly right. To get there, it is important to focus on training the model with non-obvious sources of data.
Tensor Processing Units
Tensor Processing Units (TPUs) are Google’s application-specific circuits and custom silicon for TensorFlow. Now that we are in the world of transformers, TPUs may not be necessary.
Comments
Post a Comment