Archive for February 24, 2023

Friday, February 24, 2023

What Is ChatGPT Doing and Why Does It Work?

Stephen Wolfram (via Hacker News):

And the remarkable thing is that when ChatGPT does something like write an essay what it’s essentially doing is just asking over and over again “given the text so far, what should the next word be?”—and each time adding a word.


The fact that there’s randomness here means that if we use the same prompt multiple times, we’re likely to get different essays each time. And, in keeping with the idea of voodoo, there’s a particular so-called “temperature” parameter that determines how often lower-ranked words will be used, and for essay generation, it turns out that a “temperature” of 0.8 seems best.


In the first neural nets we discussed above, every neuron at any given layer was basically connected (at least with some weight) to every neuron on the layer before. But this kind of fully connected network is (presumably) overkill if one’s working with data that has particular, known structure. And thus, for example, in the early stages of dealing with images, it’s typical to use so-called convolutional neural nets (“convnets”) in which neurons are effectively laid out on a grid analogous to the pixels in the image—and connected only to neurons nearby on the grid.

The idea of transformers is to do something at least somewhat similar for sequences of tokens that make up a piece of text. But instead of just defining a fixed region in the sequence over which there can be connections, transformers instead introduce the notion of “attention”—and the idea of “paying attention” more to some parts of the sequence than others.


First, it takes the sequence of tokens that corresponds to the text so far, and finds an embedding (i.e. an array of numbers) that represents these. Then it operates on this embedding—in a “standard neural net way”, with values “rippling through” successive layers in a network—to produce a new embedding (i.e. a new array of numbers). It then takes the last part of this array and generates from it an array of about 50,000 values that turn into probabilities for different possible next tokens.


Update (2023-03-08): See also: ChatGPT Explained: A Normie's Guide To How It Works (via Hacker News).

Quora’s Poe

Sarah Perez (via John Gruber):

Q&A platform Quora has opened up public access to its new AI chatbot app, Poe, which lets users ask questions and get answers from a range of AI chatbots, including those from ChatGPT maker, OpenAI, and other companies like Anthropic. Beyond allowing users to experiment with new AI technologies, Poe’s content will ultimately help to evolve Quora itself, the company says.

Quora first announced Poe’s mobile app in December, but at the time, it required an invite to try it out. With the public launch on Friday, anyone can now use Poe’s app. For now, it’s available only to iOS users, but Quora says the service will arrive on other platforms in a few months.


Google’s Bard

Sundar Pichai (via John Gruber, Hacker News):

We’ve been working on an experimental conversational AI service, powered by LaMDA, that we’re calling Bard. And today, we’re taking another step forward by opening it up to trusted testers ahead of making it more widely available to the public in the coming weeks.

Bard seeks to combine the breadth of the world’s knowledge with the power, intelligence and creativity of our large language models. It draws on information from the web to provide fresh, high-quality responses. Bard can be an outlet for creativity, and a launchpad for curiosity, helping you to explain new discoveries from NASA’s James Webb Space Telescope to a 9-year-old, or learn more about the best strikers in football right now, and then get drills to build your skills.


Now, our newest AI technologies — like LaMDA, PaLM, Imagen and MusicLM — are building on this, creating entirely new ways to engage with information, from language and images to video and audio. We’re working to bring these latest AI advancements into our products, starting with Search.

Isabel Angelo (via Hacker News):

Unfortunately a simple google search would tell us that JWST actually did not “take the very first picture of a planet outside of our own solar system” and this is literally in the ad for Bard so I wouldn’t trust it yet

Jennifer Elias (via Hacker News):

Staffers took to the popular internal forum Memegen to express their thoughts on the Bard announcement, referring to it as “rushed,” “botched” and “un-Googley,” according to messages and memes viewed by CNBC.

On Monday, Google got ahead of a Microsoft event the following day and had Pichai publicly divulge some details of the company’s chatbot technology.

Paul Graham:

What happens if you take too long to launch: your product is defined by its relationship to whatever launched first.

That casual appositive phrase is worth more to Microsoft than any news story.

Nick Heer:

The original point of search engines was to be directed to websites of interest. But that has not been the case for years. People are not interested in visiting websites about a topic; they, by and large, just want answers to their questions. Google has been strip-mining the web for years, leveraging its unique position as the world’s most popular website and its de facto directory to replace what made it great with what allows it to retain its dominance. Artificial intelligence — or some simulation of it — really does make things better for searchers, and I bet it could reduce some tired search optimization tactics. But it comes at the cost of making us all into uncompensated producers for the benefit of trillion-dollar companies like Google and Microsoft.


Update (2023-03-21): Google (via Hacker News):

Join the waitlist and try it for yourself.

ChatGPT in Bing and Edge

James Vincent (Hacker News):

In demos today the company showed what it’s calling “the new Bing” working in various configurations. One of these shows traditional search results side-by-side with AI annotations (above), while another mode lets users talk directly to the Bing chatbot, asking it questions in a chat interface like ChatGPT (below).


Unlike ChatGPT, the new Bing can also retrieve news about recent events. In The Verge’s demos, the search engine was even able to answer questions about its own launch, citing stories published by news sites in the last hour.


In addition to the new Bing, Microsoft is launching two new AI-enhanced features for its Edge browser: “chat” and “compose.” These will be embedded within Edge’s sidebar.

“Chat” allow users to summarize the webpage or document they’re looking at and ask questions about its contents, while “compose” acts as a writing assistant; helping to generate text, from emails to social media posts, based on a few starting prompts.

Nick Heer:

Microsoft announced today’s event unveiling these developments midday yesterday, hours after Google announced its efforts in the space, as it has done before. I am not sure whether to read this as panic or excitement, though Meta’s caution is notable.


The big question right now is, I think, where Amazon and Apple are at internally. Are they racing to compete with Alexa and Siri? Are they maybe waiting it out to see if this is a real, exciting development, or yet more baseless hype like so many technology land rushes before it?

Noor Al-Sibai:

The lawyer also revealed, per Insider, that Amazon is developing “similar technology” to ChatGPT — a revelation that appeared to pique the interest of employees who said that using the AI to assist their code-writing had resulted in a tenfold productivity boost.

Katyanna Quach (via Hacker News, Nick Heer):

Microsoft’s new AI-powered Bing search engine generated false information on products, places, and could not accurately summarize financial documents, according to the company’s promo video used to launch the product last week.


In reality, both Microsoft’s Bing and Google’s Bard are just as bad as each other. Both companies launched shoddy AI chatbots that generated text containing false information, but Microsoft’s mistakes were not immediately caught. Now, some of its errors have been spotted by Dmitri Brereton, a search engine researcher.


Update (2023-03-28): Akash Sriram and Chavi Mehta (via Hacker News):

The integration of OpenAI’s technology into Microsoft-owned Bing has driven people to the little-used search engine and helped it compete better with market leader Google in page visits growth, according to data from analytics firm Similarweb.

Page visits on Bing have risen 15.8% since Microsoft Corp unveiled its artificial intelligence-powered version on Feb. 7, compared with a near 1% decline for the Alphabet Inc-owned search engine, data till March 20 showed.

ChatGPT Is Ingesting Corporate Secrets

Noor Al-Sibai (Hacker News, Bruce Schneier):

After catching snippets of text generated by OpenAI’s powerful ChatGPT tool that looked a lot like company secrets, Amazon is now trying to head its employees off from leaking anything else to the algorithm.

According to internal Slack messages that were leaked to Insider, an Amazon lawyer told workers that they had “already seen instances” of text generated by ChatGPT that “closely” resembled internal company data.

This issue seems to have come to a head recently because Amazon staffers and other tech workers throughout the industry have begun using ChatGPT as a “coding assistant” of sorts to help them write or improve strings of code, the report notes.

Just like you once had to pay Github to keep your repositories private, perhaps ChatGPT will let you pay not to have your inputs become part of its training data.


Update (2023-02-27): Damien Petrilli:

There is this form you can use to opt out of data training.

I did it and never got any reply. So not sure you can trust them.

It’s a Google Doc.

ChatGPT for Apple Platforms Development

Steve Troughton-Smith:

At this point, I really would appreciate a tool that just parses an Xcode SceneKit .scn scene and spits out a built-in-code representation of it. Xcode’s SceneKit editor has been a liability for years, and it exhausts all my energy to fix its compatibility issues every time I update my projects 😪 PaintCode, but for SceneKit scenes

Just had my mind blown: So I NSLogged a description of my SceneKit hierarchy for the luls and gave it to ChatGPT and told it to turn it into Swift code. Gibberish, right? Well it had NO PROBLEM figuring out what I meant, and it just did what I wanted 🤯

omg ChatGPT can parse Interface Builder XML and output code too? I just dropped my Launch Storyboard into it 😱

So I crafted a storyboard in Interface Builder (pic 1). I then took that XML, pasted it into ChatGPT, and asked it to rewrite it in SwiftUI. It gave me SwiftUI code that, with the tiniest bit of massaging, I could drop into Playgrounds to get the following app (pic 2). It did take some liberties and didn’t get it 100% correct, but it’s pretty darn close

Quentin Zervaas:

This particular screen uses SwiftUI, but off the top of my head, I wasn’t sure of the easiest to create this typing effect. Instead of using Google or Stack Overflow, I thought this might be a good chance to try ChatGPT[…]


I didn’t love how it generated the timers - seemed a bit wasteful to create a repeating timer, only to cancel it every time, so I made a few changes.


On balance, using ChatGPT was a very useful way to mock up some working code. While some tweaks were needed for my specific application, it was ultimately a huge time saver.

• • •

Benjamin Mayo:

Better error messages for SwiftUI result builders coming soon

Michael Steeber:

honestly i just paste my code block and the error into ChatGPT now and have it explain the error. The “solution” usually doesn’t work, but it at least makes the problem clear to me


my success rate is a lot higher than stackoverflow because it understands the exact context of your specific use case, not just the error in an abstract sense of the language

• • •


I also am looking for “UNIBOX” alternative […] I asked if it’s possible to use extensions to make apple mail function better.

p.s. I asked ChatGPT: Maybe someone with more skill can just develop this plugin which turns apple mail into unibox

Scott Morrison:

Don’t believe everything (or anything) ChatGPT tells you.

I have been developing Plugins for Mail App for 20 years.

The ChatGPT response is 100% fiction.