Sunny

← All weeks

What I’ve been reading (W11, 2023)

How ChatGPT was born and OpenAI revenue numbers · The story of Rust, the programming language (cc: @ernest)

How ChatGPT was born and OpenAI revenue numbers

  • Most generative A.I. systems have an LLM at their core. They are created by taking very large neural networks—an A.I. based very loosely on connections in the human brain—and applying them to vast amounts of human-created text. The chat interface allows users to converse with the A.I. in plain English (and many other languages) and was initially conceived by OpenAI to improve its “large language models,” or LLMs.
  • OpenAI found getting GPT-3 to produce exactly what a user wanted could be tricky. One team suggested using reinforcement learning—an A.I. system that learns from trial and error to maximise a reward—to perfect the model. The team thought that a chatbot might be a great candidate for this method since constant feedback in human dialogue would make it easy for the A.I. software to know when it had done a good job and where it needed to be improved.
  • OpenAI switched gears and tried to build expert chatbots that could help professionals in specific domains because it wasn’t clear to people what they were supposed to talk to the chatbot about. But that effort also ran into problems—partly because OpenAI lacked the right data to train expert bots.
  • After a while, they decided to pull ChatGPT off the bench, put it in the wild for the public to use, and let beta testers play with ChatGPT.
  • It is one of the products that they managed to commercialise. The company was projected to bring in just under $30 million in revenue in 2022 with a burn of $544.5 million. They forecast $200 million in revenue for 2023 and expect revenues to top $1 billion in 2024.

The story of Rust, the programming language (cc: @ernest)

  • Graydon Hoare, a computer programmer working for Mozilla, designed a new computer language called Rust after finding the elevator in his Vancouver apartment building was out of order due to a software crash.
  • Hoare decided to do something about it. He opened his laptop and began designing a new computer language that he hoped would make writing small, fast code without memory bugs possible. He named it Rust, after a group of remarkably hardy fungi that are, he says, “over-engineered for survival.”
  • The software used in some of these old systems (incl. elevators) is written using C++ or C, which is usually fast but not memory safe. Microsoft estimates that 70% of the vulnerabilities in its code are due to memory errors from code written in these languages. Rust is designed to make writing small, fast code without memory bugs possible. Mozilla officially sponsored Rust in 2009, and over the next 10 years, Mozilla employed over a dozen engineers to work on Rust full-time.
  • Rust splits the difference between older languages like C or C++ and newer garbage-collected languages like Java, JavaScript, and Python. Rust doesn’t require programmers to figure out where in memory they are putting data manually, but it imposes many strict rules on how data can be used or copied inside a program.
  • Rust has become one of the hottest new languages on the planet—maybe the hottest. 2.8 million coders are writing in Rust, and companies from Microsoft to Amazon regard it as key to their future. The chat platform Discord used Rust to speed up its system, Dropbox uses it to sync files to your computer, and Cloudflare uses it to process more than 20% of all internet traffic.
  • A study of Rust-based code found it runs so efficiently that it uses half as much electricity as a similar program written in Java, a language commonly used at AWS. Rust could help create a data centre that runs twice the workloads or does the same work in one that’s half the size.
  • Rust could make our software landscape more reliable, less crash-prone, and less insecure. However, all old C and C++ code already kicking around won’t vanish and will remain used for many decades.

Facebook’s new restructuring plans and notes from Zuck TLDR; Facebook is expected to reduce our team size by around 10,000 people and close around 5,000 additional open roles. Zuck shared some sensible thoughts about why they are doing it.

Making the team leaner and executing the highest priorities

Facebook is asking many managers to become individual contributors (ICs) to reduce layers of hierarchy which adds latency and risk aversion in information flow and decision-making. They are going to make ICs report into almost every level — not just the bottom — so information flow between people doing the work and management will be faster. A leaner org will execute its highest priorities faster. People will be more productive, and their work will be more fun and fulfilling. We will become an even greater magnet for the most talented people. Focusing on strengths and the right outcomes We are a technology company, and our ultimate output is what we build for people. Everything else we do is in service of that. We’re focusing on returning to a more optimal ratio of engineers to other roles. It’s important for all groups to get leaner and more efficient to enable our technology groups to get as lean and efficient as possible. We will make sure we continue to meet all our critical and legal obligations as we find ways to operate more efficiently. On using AI to improve productivity We’re focused on the long term. That means investing in tools that will make us most effective over many years, not just this year — whether that’s building AI tools to help engineers write better code faster, enabling us to automate workloads over time, or identifying obsolete processes that we can phase out. Profitability enables innovation. Operating our business more efficiently will give us the resources and confidence to achieve our long term vision by delivering sustainable financial results that make us an attractive company to work at and invest in. Other thoughts a new economic reality Higher interest rates lead to the economy running leaner, more geopolitical instability leads to more volatility, and increased regulation leads to slower growth and increased costs of innovation.

Voice assistants lost the race to AI

  • Virtual assistants had more than a decade to become indispensable. But they were hampered by clunky design and miscalculations, leaving room for chatbots to rise.
  • John Burkey, an ex-Apple engineer, said there was no path for Siri to become a creative assistant like ChatGPT. Apple Siri has a cumbersome design, making it time-consuming to add new features. Siri’s database contains a gigantic list of words, including the names of musical artists and locations like restaurants, in nearly two dozen languages and to his made it “one big snowball”. If someone wanted to add a word to Siri’s database, he added, “it goes in one big pile.” So seemingly simple updates, like adding some new phrases to the data set, would require rebuilding the entire database and could take up to six weeks. Adding more complex features like new search tools could take nearly a year. He said t
  • Amazon’s Echo, a smart speaker powered by Alexa, in 2014, the company hoped the product would help it increase sales in its online store by enabling consumers to talk to Alexa to place orders, said a former Amazon leader involved with Alexa. But while people had fun playing with Alexa’s ability to answer weather prompts and set alarms, few asked Alexa to order items.
  • The company also underinvested in creating an ecosystem for people to easily expand Alexa’s abilities as Apple had done with its App Store, which helped stoke interest in the iPhone, the person said. While Amazon offered a “skills” store to make Alexa control third-party accessories like light switches, it was difficult for people to find and set up skills for the speakers — unlike the friction-free experience of downloading mobile apps from app stores.
  • Alexa and Google Assistant relied on technology similar to Siri’s, but the companies struggled to generate meaningful revenue with the assistants, former managers at Amazon and Google said.

Others

Substack edition

View the original post

How ChatGPT was born and OpenAI revenue numbers

  • Most generative A.I. systems have an LLM at their core. They are created by taking very large neural networks—an A.I. based very loosely on connections in the human brain—and applying them to vast amounts of human-created text. The chat interface allows users to converse with the A.I. in plain English (and many other languages) and was initially conceived by OpenAI to improve its “large language models,” or LLMs.

  • OpenAI found getting GPT-3 to produce exactly what a user wanted could be tricky. One team suggested using reinforcement learning—an A.I. system that learns from trial and error to maximise a reward—to perfect the model. The team thought that a chatbot might be a great candidate for this method since constant feedback in human dialogue would make it easy for the A.I. software to know when it had done a good job and where it needed to be improved.

  • OpenAI switched gears and tried to build expert chatbots that could help professionals in specific domains because it wasn’t clear to people what they were supposed to talk to the chatbot about. But that effort also ran into problems—partly because OpenAI lacked the right data to train expert bots.

  • After a while, they decided to pull ChatGPT off the bench, put it in the wild for the public to use, and let beta testers play with ChatGPT.

  • It is one of the products that they managed to commercialise. The company was projected to bring in just under $30 million in revenue in 2022 with a burn of $544.5 million. They forecast $200 million in revenue for 2023 and expect revenues to top $1 billion in 2024.

The story of Rust, the programming language

  • Graydon Hoare, a computer programmer working for Mozilla, designed a new computer language called Rust after finding the elevator in his Vancouver apartment building was out of order due to a software crash.

  • Hoare decided to do something about it. He opened his laptop and began designing a new computer language that he hoped would make writing small, fast code without memory bugs possible. He named it Rust, after a group of remarkably hardy fungi that are, he says, “over-engineered for survival.”

  • The software used in some of these old systems (incl. elevators) is written using C++ or C, which is usually fast but not memory safe. Microsoft estimates that 70% of the vulnerabilities in its code are due to memory errors from code written in these languages. Rust is designed to make writing small, fast code without memory bugs possible. Mozilla officially sponsored Rust in 2009, and over the next 10 years, Mozilla employed over a dozen engineers to work on Rust full-time.

  • Rust splits the difference between older languages like C or C++ and newer garbage-collected languages like Java, JavaScript, and Python. Rust doesn’t require programmers to figure out where in memory they are putting data manually, but it imposes many strict rules on how data can be used or copied inside a program.

  • Rust has become one of the hottest new languages on the planet—maybe the hottest. 2.8 million coders are writing in Rust, and companies from Microsoft to Amazon regard it as key to their future. The chat platform Discord used Rust to speed up its system, Dropbox uses it to sync files to your computer, and Cloudflare uses it to process more than 20% of all internet traffic.

  • A study of Rust-based code found it runs so efficiently that it uses half as much electricity as a similar program written in Java, a language commonly used at AWS. Rust could help create a data centre that runs twice the workloads or does the same work in one that’s half the size.

  • Rust could make our software landscape more reliable, less crash-prone, and less insecure. However, all old C and C++ code already kicking around won’t vanish and will remain used for many decades.

Facebook’s new restructuring plans and notes from Zuck

TLDR; Facebook is expected to reduce our team size by around 10,000 people and close around 5,000 additional open roles. Zuck shared some sensible thoughts about why they are doing it.

Making the team leaner and executing the highest priorities

Facebook is asking many managers to become individual contributors (ICs) to reduce layers of hierarchy which adds latency and risk aversion in information flow and decision-making.

They are going to make ICs report into almost every level — not just the bottom — so information flow between people doing the work and management will be faster.

A leaner org will execute its highest priorities faster. People will be more productive, and their work will be more fun and fulfilling. We will become an even greater magnet for the most talented people.

Focusing on strengths and the right outcomes

We are a technology company, and our ultimate output is what we build for people. Everything else we do is in service of that.

We’re focusing on returning to a more optimal ratio of engineers to other roles. It’s important for all groups to get leaner and more efficient to enable our technology groups to get as lean and efficient as possible. We will make sure we continue to meet all our critical and legal obligations as we find ways to operate more efficiently.

On using AI to improve productivity

We’re focused on the long term. That means investing in tools that will make us most effective over many years, not just this year — whether that’s building AI tools to help engineers write better code faster, enabling us to automate workloads over time, or identifying obsolete processes that we can phase out.

Profitability enables innovation. Operating our business more efficiently will give us the resources and confidence to achieve our long term vision by delivering sustainable financial results that make us an attractive company to work at and invest in.

Other thoughts a new economic reality

Higher interest rates lead to the economy running leaner, more geopolitical instability leads to more volatility, and increased regulation leads to slower growth and increased costs of innovation.

Voice assistants lost the race to AI

  • Virtual assistants had more than a decade to become indispensable. But they were hampered by clunky design and miscalculations, leaving room for chatbots to rise.

  • John Burkey, an ex-Apple engineer, said there was no path for Siri to become a creative assistant like ChatGPT. Apple Siri has a cumbersome design, making it time-consuming to add new features. Siri’s database contains a gigantic list of words, including the names of musical artists and locations like restaurants, in nearly two dozen languages and to his made it “one big snowball”. If someone wanted to add a word to Siri’s database, he added, “it goes in one big pile.” So seemingly simple updates, like adding some new phrases to the data set, would require rebuilding the entire database and could take up to six weeks. Adding more complex features like new search tools could take nearly a year. He said t

  • Amazon’s Echo, a smart speaker powered by Alexa, in 2014, the company hoped the product would help it increase sales in its online store by enabling consumers to talk to Alexa to place orders, said a former Amazon leader involved with Alexa. But while people had fun playing with Alexa’s ability to answer weather prompts and set alarms, few asked Alexa to order items.

  • The company also underinvested in creating an ecosystem for people to easily expand Alexa’s abilities as Apple had done with its App Store, which helped stoke interest in the iPhone, the person said. While Amazon offered a “skills” store to make Alexa control third-party accessories like light switches, it was difficult for people to find and set up skills for the speakers — unlike the friction-free experience of downloading mobile apps from app stores.

  • Alexa and Google Assistant relied on technology similar to Siri’s, but the companies struggled to generate meaningful revenue with the assistants, former managers at Amazon and Google said.

Originally on Substack