Hard Fork

Help! My Boss Won’t Stop Using ChatGPT

Fri Jul 14 2023

Chat GPTGrounding AI Language ModelsScaling AI ModelsVenture CapitalThe VC IndustryAlternate EndingsAccessibilityEtiquette in the Age of A.I.Recreating Loved Ones with AI

Description

This episode covers various topics including chat GPT, grounding AI language models, scaling AI models, venture capital due diligence, the VC industry, alternate endings in movies, accessibility, etiquette in the age of A.I., and recreating loved ones with AI. The hosts answer listener questions and provide insights into these subjects.

Insights

AI Language Models and Incorrect Information

Chat GPT relies on statistical guesses and can produce incorrect information. Adding a confidence indicator may help improve accuracy.

Grounding AI Language Models

Companies are exploring tying AI language models to authoritative sources of knowledge. Bing has integrated its AI technology with the search index.

Scaling AI Models

There is a predictable relationship between model size, computing power, and performance. Scaling laws allow predicting model performance based on size and compute.

Venture Capital Due Diligence

VCs may overlook questionable numbers and rely on groupthink. Limited accountability exists for bad investments. The SEC is considering new rules.

The VC Industry

The VC industry is lopsided, with few top firms achieving significant success. Many VCs spend time on social media rather than having real jobs.

Alternate Endings and Accessibility

AI technology could enable alternate endings in movies. Speed listening can benefit blind users. More stories about accessibility are needed.

Etiquette and Decency in the Age of A.I.

Using chat GPT to answer work-related questions lacks human effort. Searching for exes on LinkedIn raises ethical concerns.

Recreating Loved Ones with AI

AI voice generating technology can clone the voice of a deceased loved one. Using existing recordings for personal healing is ethical.

AI Technology and Recreating Loved Ones

Recreating loved ones with AI is a personal choice. The hosts appreciate listener feedback and acknowledge the production team.

Summary

Transcript

Introduction

00:01 - 06:32

The hosts introduce the episode and mention that they will be answering listener questions about tech.
They divide the episode into two halves: one with questions that have correct answers, and another with thornier questions.

Chat GPT and Incorrect Information

00:01 - 06:32

A listener named Madeline Winter asks why chatbots like GPT sometimes produce incorrect information instead of saying they don't know.
The hosts explain that chat GPT doesn't have a concept of knowledge and relies on statistical guesses, which can lead to hallucinations.
They discuss the idea of adding a confidence indicator to these models to indicate how certain they are in their answers.
While there are instances of hallucination, most of the time these models perform well when it comes to factual answers.

Grounding AI Language Models

06:07 - 13:18

Companies are exploring grounding AI language models by tying them to authoritative sources of knowledge.
Bing has fused its AI technology with the Bing search index, allowing the AI model to query the search engine for information.
The carbon footprint of training AI models is not well-documented, but it involves a computationally intensive process and energy consumption.
Training GPT-3 consumed about 1.3 gigawatt hours of electricity.
Tech giants like Google and Microsoft have pledged to be carbon neutral in their AI operations.
AI models have shown consistent improvement over time, such as GPT-2 being better than GPT-1 and GPT-4 being better than GPT-3.
There is optimism that continued advancements can be achieved by scaling existing techniques rather than waiting for new breakthroughs.

Scaling AI Models

13:02 - 20:06

There is a predictable relationship between the size of AI models, computing power used for training, and their performance.
Scaling laws for neural language models allow predicting model performance based on size and compute.
So far, there is no indication that these scaling laws will break down in the future.
Researchers are confident that as models get bigger, they will also improve.
However, there could be an invisible asymptote where model performance plateaus.

Venture Capital and Due Diligence

13:02 - 20:06

Venture capital funds may trust tech companies' financial claims despite poor bookkeeping.
Entrepreneurs sometimes lie to venture capitalists, and appearing on Forbes' 30 under 30 list can be an indicator of dishonesty.
VCs get rich by backing risky bets with unusual founders who defy expectations.
VCs expect a high failure rate among their investments and may overlook questionable numbers until it becomes a problem.
Due diligence by VCs can be influenced by groupthink and bandwagon mentality in the industry.
Limited accountability exists for VCs making bad investments due to insufficient research.
The SEC is considering a rule that would allow investors to sue VCs for negligence in due diligence.
Journalists have mistakenly assumed legitimacy based on VC investments in certain firms.

The VC Industry

19:39 - 26:32

VC firms may not always do thorough due diligence on the companies they invest in.
There is a trend of many tech professionals working for VC firms instead of building startups.
There is an oversaturation of money in the VC industry, with too much capital chasing too few ideas.
The life of a VC may seem glamorous, but there are fewer successful VCs than people think.
Many VCs spend their time on social media and podcasts rather than having real jobs.
Working as an operator in a tech company before becoming a VC can provide more value to entrepreneurs.
The VC industry is lopsided, with only a few top firms getting access to the best deals and achieving significant success.
VC firms have a compensation structure that rewards them regardless of investment outcomes.

Alternate Endings and Accessibility

26:03 - 33:10

AI technology could potentially allow for alternate endings in movies and impact society's emotional understanding.
Alternate endings can be created through fan fiction and fair use laws.
Tech-enabled versions of alternate endings already exist, such as interactive shows on Netflix.
Filmmakers may not embrace AI-generated alternate endings as it goes against the ideals of art and storytelling.
Some viewers desire a low-stimulus version of intense shows, making technology like alternate endings appealing to them.
Blind users can grasp audio information at higher speeds, making speed listening an accessibility feature.
More stories about accessibility in tech are needed to understand the experiences of people with disabilities.
Even popular websites like Reddit still have accessibility issues that hinder certain users from accessing their content.
Generative AI technology has potential applications in improving accessibility for people with disabilities.
AI can assist blind individuals in navigating the world through apps like Be My Eyes.

Etiquette and Decency in the Age of A.I.

32:47 - 38:52

A listener asks about etiquette and decency in the age of A.I.
The listener's boss uses chat GPT to answer work-related and casual questions.
Casey believes this behavior is not okay as it lacks human effort.
The boss's use of chat GPT prevents employees from getting to know him.
The behavior feels passive-aggressive and unhinged.
Possible solutions include confronting the boss or finding a new job.
Karen wonders if it's ethical to search for her ex on LinkedIn.
Kevin thinks it's fine as everyone snoops on their exes, regardless of platform.
Concerns arise if the ex can see that they are being looked up on LinkedIn.

Recreating Loved Ones with AI

38:30 - 45:04

LinkedIn has a feature where you can see who has viewed your profile.
Looking up your ex on LinkedIn can be risky if they receive notifications.
Using AI voice generating technology to clone the voice of a deceased loved one is possible.
Consulting with family members, especially the living spouse, is important in such cases.
Creating an audiobook version of a deceased loved one's autobiography using their voice can bring comfort and joy within the family.
It's ethical to use existing recordings of a departed loved one's voice for personal healing and processing.
Replica, an AI startup, offers chatbot companions based on text messages from deceased individuals.
People process grief in different ways, and using AI to recreate loved ones is a personal choice.

AI Technology and Recreating Loved Ones

44:36 - 46:49

AI technology can be used to recreate loved ones who have passed away.
The speaker personally finds the idea of recreating loved ones with AI creepy and potentially increasing grief.
The podcast hosts appreciate emails and voice memos from listeners.
Listeners are referred to as 'forkers'.
Acknowledgements for the production team.