Stephen Wolfram on AI and Chat GPT

In this episode of This Week in Startups, Jason interviews Stephen Wolfram, founder and CEO of Wolfram Research. They discuss the history of neural nets, the development of chat GPT, how it works, and its potential impact on jobs in the future.

History of Neural Nets

• Stephen has been tracking neural nets since 1980.

• Deep learning neural nets started doing interesting things around 2012.

• Large language models like chat GPT are able to do human-like tasks by using statistics from text on the web.

How Chat GPT Works

• The model uses a neural net to predict what word comes next based on the statistics of text on the web.

• The model extrapolates from these statistics in a way that is similar to how humans would do it.

• The model needs a large amount of data to work effectively.

Potential Impact on Jobs

• Chat GPT and other AI technologies have the potential to shape jobs in the future.

• Some jobs may be replaced by AI, while others may be augmented by it.

• It is important for society to prepare for these changes and ensure that people are able to adapt.

Introduction to Neural Nets

In this section, the speaker introduces neural nets and explains how they work.

What is a Neural Net?

• Neurons have incoming connections called dendrites that receive signals.

• When there are enough incoming signals, the neuron fires an electrical pulse that gets sent out to its outgoing wires or neural wires.

• Incoming signals on each wire have a weight, which is multiplied by the signal and added up. The sum goes through a thresholding function to determine whether the neuron fires and sends data on to the next neurons down the line.

History of Neural Nets

• Invented in 1943, people started thinking about formal representation and mathematical ways to represent neural wiring.

• In the 1950s and 1960s, people explored what happens when you have five neurons, ten neurons, thirty connections between neurons but didn't do anything exciting.

• It wasn't obvious how big the number of neurons would have to be or how much data you would need to train with for human-like behavior.

• Between 1940 and now, many clever ideas were tried but didn't work out. The structure of neural nets with a few extra pieces is really close to what people imagined back in the 1940s.

Three Components Needed for Neural Nets

• A large corpus of text available on the web that can be used for training data.

• Enough compute power and storage capacity to process it fast enough.

• A language model built by somebody.

How Neural Nets Work

This section provides more details on how neural nets work.

Importance of Weights

• Each incoming signal has a weight that determines its importance.

• The weights are multiplied by the signal and added up to determine whether the neuron fires.

Language Model

• Neural nets can produce things that match human-like behavior by training on large text corpora available on the web.

• The language model is built using a few clever ideas, but it's not significantly different from what was imagined in the 1940s.

Three Components Needed for Neural Nets

• A large corpus of text available on the web that can be used for training data.

• Enough compute power and storage capacity to process it fast enough.

• A language model built by somebody.

Cast AI: Reduce Your Cloud Spend by 60%

In this section, the speaker introduces Cast AI and explains how it can help reduce cloud spending.

Introduction to Cast AI

• Cast AI is a tool that helps discover what could be reduced in your Cloud mail.

• It eliminates the stuff that you pay for but don't use.

• It searches for less expensive hosting options within your cloud provider.

• On average, Cast AI customers save over 60% on their Cloud spend.

Benefits of Using Cast AI

• By using Cast AI, you could be spending 60% less on your cloud bill.

• This will add more runway to your startup as every dollar counts in this funding environment.

• Before signing any multi-year Cloud contracts or making any drastic personal decisions, check out what cast AI can do for you.

• You get a personal free Cloud cost audit and consultation.

How Chatbots Work

In this section, the speaker explains how chatbots work and the different components involved in building them.

Components of Chatbots

• Words are turned into numbers using embeddings.

• These numbers determine the intensities of firing of neurons in the first layer.

• The data from these neurons goes through a sequence of layers (about 400 layers for GPT).

• The output is a collection of numbers that gives probabilities for possible words that might follow.

Making Chatbots Effective

• Picking not just the most probable word but sometimes other probability words leads to more lively results.

• A critical piece of what's worked in something like chat GPT is this idea of Transformers.

• Transformers help feed words into neural nets by knowing they are in a sequence.

• The neuron that learns it knows given that we're going to add the next word, it says well the word three back.

Neural Net Training

In this section, the speaker explains how neural net training works and how weights are determined.

Determining Weights

• The neural net has 175 billion weights.

• Weights are picked so that the neural net conforms to the statistics of the web.

• The training process involves guessing what the next word is and tweaking all those weights using a mechanism called back propagation.

Training Process

• The training process involves gradually adapting it to get closer and closer to the right answer.

• You find some high-quality piece of text, mask out the words at the end of the text, and train it so that when you take off the mask, it will really be the ones that were there.

• It's trained on a trillion words, which is roughly comparable to the number of weights needed.

Importance of Sock 2 Compliance

In this section, the speaker talks about why sock 2 compliance is important for startups.

Benefits of Sock 2 Compliance

• Tech is getting hit super hard as such you cannot afford to lose sales for silly stuff like not having your sock 2 right now.

• Vanta makes it incredibly easy to get and renew your sock too on average van to customers our stock 2 compliant in just two to four weeks compare that to three to five months without vanta huh.

• One major customer that could be the difference between your startup thriving or going away so get it done right now that is going to give you a thousand dollars off because you'll listen to this podcast.

Neural Nets and Brains

In this section, the discussion is about how close neural nets are to emulating what happens in a human brain. The speaker talks about the similarities and differences between brains and computers.

Similarities and Differences Between Brains and Computers

• The speaker believes that what's happening in neural nets is fairly close to what happens in brains.

• One difference between brains and computers is that every neuron in a brain both computes and stores things, while computer memory typically just sits there storing data.

• Another difference is that there's some kind of feedback loop in our brains, which may be similar to the one that chat GPT has.

• At a computational architecture level, the speaker thinks that brains are surprisingly close to computers.

Computational Irreducibility

In this section, the discussion is about computational irreducibility. The speaker explains what it means and how it relates to deep computation.

What Is Computational Irreducibility?

• Computational irreducibility refers to the feature of deep computation where you have certain rules that you apply over and over again until you see what comes out.

• There's a question of whether you can jump ahead and see what the result will be more quickly than just following all those rules.

Introduction to Computation and Emergent Behavior

In this section, Stephen Wolfram discusses the concept of irreducible computations and how they differ from shallower computations that humans use most of the time. He also talks about emergent behavior and how it relates to the complexity of nature.

Irreducible Computations

• Certain computations cannot be shortcut and require computational work.

• GPT can only do computations up to a certain point before running out of layers in its neural net.

• Nature contains many irreducible computations that are not made for humans, making them difficult to understand without simulation or analysis.

Emergent Behavior

• Emergent behavior refers to when a system's actual behavior is much more complicated than the rules put in place.

• The complexity of nature is due in part to irreducible computations leading to emergent behavior.

• LMs are currently being trained using individual chat sessions, but there is potential for threaded chats to combine information in the future.

Ethics of LM Training

In this section, Stephen Wolfram discusses ethical considerations surrounding LM training and data privacy.

Data Privacy

• There are technological and policy issues surrounding combining information from multiple chat sessions into an LM's training data.

• Individual chat sessions with LMs are being stored for long-term training purposes but not immediately used for real-time learning.

Ethical Considerations

• No bullet points available.

Privacy and Language in Chatbots

In this section, the speaker discusses the issue of privacy in chat sessions and how it relates to medical records. They also explore the idea that language may not be as complicated as previously thought.

Privacy in Chat Sessions

• The question of how private chat sessions are arises.

• The value of aggregate data versus individual privacy is discussed.

• Technical challenges associated with maintaining privacy are mentioned.

Complexity of Language

• The speaker suggests that language may not be as complex as previously thought.

• Rules of language such as syntactic grammar are mentioned.

• LLMS have discovered many more regularities in language than previously classified.

• There are many meaningful structures to language beyond parts of speech.

• Aristotle's discovery of logic is used as an example.

• LLMS have found a collection of other puzzle pieces that make up semantic regularities in language.

The Importance of Language

In this section, Stephen Wolfram discusses the importance of language as a means to represent and understand the world.

Language as a Means to Represent the World

• Humans use language to represent and make decisions about the world.

• Computational language can be used to precisely represent things in the world in a formal way that is both readable and writable by humans and computers.

• Wolfram's long-term project is to create a computational language that can represent things in the world in a precise, formal, and computational way.

Natural Language Understanding

• Natural language understanding involves going from small fragments of human language to computational language.

• Once something is represented in computational language, it can be computed for anything you want from it.

Implications of Computational Language

In this section, Wolfram discusses some implications of computational language.

Connecting LLM Layer with Computational Bedrock

• The Wolfen plugin for Chat GPT connects an LLM layer with what we might think of as computational bedrock.

• Clumio provides tips on how to protect your data, take control of your cloud costs, and not let backups and compliance requirements distract you.

Clumio: Saving Time and Money

In this section, the speaker talks about the three major business models in the world and how Clumio fits into two of them. He also provides a call to action for viewers to visit clumio.com.

Clumio's Value Proposition

• The speaker believes that businesses should save people time, save them money, or entertain them.

• Clumio saves time and money for its users.

• Viewers are encouraged to visit clumio.com to start a free backup or sign up for a demo.

Chat GPT and Wolfram Alpha

In this section, the speaker discusses how Chat GPT can summarize natural language prompts and how Wolfram Alpha can provide precise computational answers. He also mentions some limitations of Tech GPT.

Chat GPT Summarization

• Chat GPT can summarize natural language prompts.

• Wolfram Alpha provides precise computational answers.

• Sometimes Chat GPT needs help with summarizing prompts correctly.

Limitations of Tech GPT

• Tech GPT doesn't always get numbers or equations right.

• Sometimes it gets prompts roughly right but not exactly right.

Using Wolfram Alpha with Natural Language Prompts

In this section, the speaker talks about how Wolfram Alpha is built for people who ask direct questions in natural language. He also discusses how Chat GPT can boil down complicated prompts into precise computations.

Using Wolfram Alpha

• Wolfram Alpha is built for people who ask direct questions in natural language.

• Chat GPT can boil down complicated prompts into precise computations.

Example of Using Wolfram Alpha

• The speaker gives an example of asking for the distance between Los Angeles and London.

Chat GPT Generating Graphics

In this section, the speaker talks about how Chat GPT can generate graphics and real-time feeds of data.

Generating Graphics

• Chat GPT can call the Wolfram plugin to generate graphics.

• Real-time feeds of data can be used to create histograms or charts.

Conclusion

In this section, the speaker concludes by discussing how Chat GPT takes complicated natural language prompts and boils them down into precise computations. He also mentions a new workflow that they have been using in the last two weeks.

Summary

• Chat GPT takes complicated natural language prompts and boils them down into precise computations.

• A new workflow has been discovered in the last two weeks.

Understanding Workflows and Use Cases

In this section, the speaker emphasizes the importance of understanding workflows and use cases in order to effectively utilize AI technology. The speaker also discusses the challenges of prompt engineering.

Importance of Understanding Workflows and Use Cases

• It is important to understand workflows and use cases when working with AI technology.

• The speaker gives an example of how a question can be better answered by asking if an answer is correct rather than generating an answer in the first place.

• Prompt engineering is challenging because it involves understanding how to structure prompts for optimal results.

Challenges of Prompt Engineering

• The theoretical description of how neural nets work is still far from being able to guide prompt engineering decisions.

• Prompt engineering is like animal wrangling - trying different approaches until you find what works best.

Impact on Society and Humans

In this section, the speaker discusses the impact that AI technology has on society and humans. He compares it to other technological advancements throughout history and raises concerns about job displacement.

Comparison to Other Technological Advancements

• The pace at which AI technology is advancing feels qualitatively different from previous technological advancements such as automation of software or online education.

• There are concerns about job displacement in fields such as copywriting, journalism, research, and design due to increased efficiency through AI technology.

Job Displacement Concerns

• While there have been dramatic changes in jobs due to technological advancements in the past, it takes time for these changes to fully work their way through society.

• There are concerns about the availability of jobs in fields that are becoming more automated, but humans have historically found new work to do.

Linguistic User Interfaces

In this section, the speaker discusses how essays can be used as a convenient transport medium for information. He also introduces the concept of linguistic user interfaces and how they are like an interface between two sides that don't quite know what the other side is looking for.

Essays as a Transport Medium

• Essays can be used as a convenient transport medium for information.

• They provide background foundational facts and make it easy to extract information.

• They are useful when both sides aren't aligned or don't know what the other side is looking for.

Linguistic User Interfaces

• Linguistic user interfaces are like an interface between two sides that don't quite know what the other side is looking for.

• They are similar to graphical user interfaces but focus on language instead of graphics.

• They allow humans to communicate with computers in a way that represents things at a human level rather than at the level that happens to be convenient for the computer.

Automation of Knowledge Worker Type Professions

In this section, the speaker talks about knowledge worker type professions and how people have assumed they cannot be automated. He also discusses programming languages and how they serve as a bridge between how humans think about things and what can be done computationally.

Automation of Knowledge Worker Type Professions

• People have assumed that knowledge worker type professions cannot be automated, but this is not true.

• Human judgment plays a significant role in these professions, making them difficult to automate.

• However, there has been progress in automating some aspects of these professions.

Programming Languages

• Programming languages serve as a bridge between how humans think about things and what can be done computationally.

• They allow humans to communicate with computers in a way that represents things at a human level.

• The goal is to represent things at a human level rather than at the level that happens to be convenient for the computer.

The Future of Automation

In this section, Stephen Wolfram discusses the future of automation and how it will impact society. He talks about how automation has historically led to more job opportunities and increased diversity in the workforce.

Impact of Automation on Job Opportunities

• Wolfram believes that some things will happen reasonably quickly due to societal attitudes and momentum.

• As automation increases, more niches become available for people to fill, leading to increased diversity in the workforce.

• Historical examples such as telephone switchboard operators show that automated switching enabled the telecommunications industry, generating a range of jobs.

Human Goals vs AI Goals

• Humans have intrinsic goals while AI does not. The goals humans have come from history and biology.

• When something gets automated, it enables other opportunities. Defining what direction or objective is important for humans because they need to pilot or jockey the technology.

• Humans need to define use cases for technologies like LMS because an LMS could just go spinning random words out without any clear objective.

Limitless Possibilities

• There is no limit to what can be invented theoretically due to computational irreducibility.

• However, there may be a point where humans say they are done inventing everything they care about.

The Future of Work

In this section, the speakers discuss the possibility of automation taking over knowledge work and what that means for humans in terms of their role in the workforce.

Post-Knowledge Work Era

• With the rise of automation, knowledge work may become automated, leading to fewer human workers involved in it.

• This shift could lead to a new era where creativity and human judgment are more important than rote knowledge work.

• Specialized siloed knowledge may decrease in value as automation can handle deep dives into specific areas.

Value of Human Creativity

• The value of globally thinking about problems will increase as it is not something that can be easily automated.

• More creative and arbitrary tasks such as creating products or services that solve pressing problems will become more valuable.

• The post-knowledge work era may be an era where human judgment and creativity are the driving force.

Optimism for the Future

• There are more pieces to the pie chart now than ever before, with different skills and interests becoming available for people to pursue.

• As technology advances, there will be more opportunities for people with different interests and skills.

The Rise of Podcasting

In this section, the speakers discuss the rise of podcasting and how it has become a popular medium for people to share their passions with others.

Podcasting as a Profession

• Podcasting has become a profession for many people who are passionate about certain topics.

• There are now hundreds of thousands of people making a living from podcasts, with millions of listeners tuning in.

• The speakers discuss the different categories that have emerged within podcasting, such as AI Wranglers and AI Psychologists.

Unintended Consequences

In this section, the speakers discuss unintended consequences that can arise from technological advancements and research.

Unintended Consequences in Research

• The pace of technological advancements means that unintended consequences are always possible.

• The speakers use virology research as an example of how unexpected outcomes can occur.

• They briefly touch on the possibility that COVID-19 may have been created by humans or even AI.

Predicting Protein Shapes with Language Models

• Large language models have been used to predict protein shapes based on their sequences.

• This is important because proteins' shapes determine their biological functions.

The Possibility of Using Generative AI to Create Proteins

In this section, the speakers discuss the possibility of using generative AI to create proteins and the potential issues that may arise.

Generative AI for Protein Creation

• Generative AI can be used to create proteins based on a given prompt.

• There are many computational and ethical questions surrounding this technology.

• It will be possible for people to use generative AI to create proteins that work differently from existing ones by pulling in data from other genome databases.

• This technology has both positive and negative implications, depending on how it is used.

Ethics of AIS

• The speakers discuss the need for an ethical framework for AIS (Artificial Intelligence Systems).

• Defining what we want AIS to do is crucial before implementing them.

• One suggestion is making AIS behave like humans aspire to be, but whose aspirations should we follow?

• A worldwide legal code for AIS may not be a good idea as it could make the system brittle.

Open Source vs For-Profit

• The speakers discuss open-source versus for-profit models in relation to dangerous code.

• OpenAI started as an open-source non-profit but later became a for-profit company with restricted access to its code due to safety concerns.

• The question arises whether or not such code should be open-sourced or kept private.

The Importance of Business Models for Innovation

In this section, the speaker discusses the significance of having a viable business model to support innovation. He emphasizes the importance of having direct business models where customers pay for the product or service rather than indirect models with advertising.

Viable Business Models and Innovation

• Having a viable business model is crucial for innovation.

• Direct business models where customers pay for the product or service are preferred over indirect models with advertising.

• Pulling everyone down so that nobody has the war chest or motivation to be a leader is not good for innovation.

• Organizations need to have some independence in deciding what they will do up to a point, as it is implausible that creative innovation will happen if the whole world votes on what should be done next.

Independent Innovation and Chat GBT

In this section, the speaker talks about independent innovation and how it led to creating chat GBT. He highlights that it took a relatively modest-sized group of people to achieve this and gives credit to those who were motivated to innovate.

Independent Innovation and Chat GBT

• Independent innovation is important, as it can lead to significant achievements like creating chat GBT.

• It took a relatively modest-sized group of people (a couple hundred people) to achieve this.

• Those who were motivated to innovate should get all the credit in achieving such an accomplishment.

True Artificial Intelligence

In this section, the speaker shares his thoughts on true artificial intelligence (AGI). He mentions that he has been paying attention to the development of computers for the last 50 years and that he has personally built some of the X's that people have said would lead to true artificial intelligence.

True Artificial Intelligence

• The speaker believes that true human intelligence will be achieved when there is a copy of a human, but even then, people may say it doesn't have certain attributes.

• The speaker does not provide a specific timeline or year for achieving AGI.

The Future of Automation and Robotics

In this section, the speaker discusses the advancements in automation and robotics, including the level of humanity that can be achieved in automated systems. They also touch on the Turing test and its relevance to artificial intelligence.

Advancements in Automation

• The speaker compares automating things to printing, where there was a standardized font for letters A, B, and C. Some people argue that automation is more efficient while others believe it loses the human touch.

• People may continue to prefer handmade items with little errors over machine-made perfect items.

• The speaker mentions Boston Dynamics' robot that could run and do flips as an example of how robotics has surprised people with its advancements.

Universal Computation for Robotics

• Unlike computers, robots have not yet achieved universal computation where a fixed piece of hardware can perform different tasks. This is because dealing with the physical world is more challenging than dealing with information.

• Biology at a molecular scale has solved this problem by using proteins that curl up into different shapes to perform various functions. However, we have not yet solved this problem on a large scale.

Future of Robotics

• Automating 3D objects and animation is already very close to being achieved through language models.

• Machine learning for grasping objects like cell phones has been difficult but will likely be cracked eventually.

• Solving the universal robotics problem will revolutionize robotics as we know it.

The Future of Robotics and Computational Language

In this section, Dr. Wolfram discusses the future of robotics and how computational language can be used to manipulate the physical world.

Manipulating the Physical World with Software

• The main thing that will happen is manipulating the physical world will become a problem of software.

• This means that robots will be able to perform tasks based on software rather than being limited by their physical capabilities.

Chat Interface for Robots

• A chat interface or language model would allow users to communicate with robots in natural language.

• The robot could then use its physical capabilities to carry out tasks such as carrying bags upstairs.

Importance of Computational Language

• Generating a piece of computational language from natural language is important for ensuring that robots carry out tasks correctly.

• Without this intermediate layer, it can be challenging to ensure that robots do what they are supposed to do.

Humanoid-Like Robots

• There is pressure for humanoid-like robots because our built environment was designed for humans.

• However, there are environments where something quite different would be appropriate.

Conclusion and Final Thoughts

In this section, Dr. Wolfram concludes the interview and shares his final thoughts on the topic.

Fast-Paced Advancements in Robotics

• It is crazy how fast advancements in robotics are moving.

• People can check out Wolf from Alpha and start playing with it to see what they can build.

Addressing Dr. Wolfram

• For business purposes, Dr. Wolfram prefers to be addressed as "Mister."

Final Remarks

• Dr. Wolfram thanks everyone for taking the time to listen and encourages them to share whatever they build on Twitter.