ChatGPT isn’t available in China—it’s not blocked, but OpenAI, which built the tool, hasn’t made it available there—so Guo went onto Taobao, China’s biggest ecommerce site, where hundreds of thousands of merchants offer everything from iPhone cases to foreign driver’s licenses.
ChatGPT logins have become a hot commodity on Taobao, as have foreign phone numbers—particularly virtual ones that can receive verification codes. A simple search on the platform in early February returned more than 600 stores selling logins, with prices ranging from 1-30 RMB ($0.17-$4.28). Some stores have made thousands of sales. On Tencent’s WeChat, a thriving market for ChatGPT knockoffs has sprung up—mainly via mini programs (sub-applications on the platform) like “ChatGPT Online.” These offer users a handful of free questions before charging for time using a chatbot. Most of these are intermediaries—they ask ChatGPT questions for users and then send the answers back. On Baidu, China’s biggest search engine, “How to use ChatGPT within China” has been consistently trending for weeks.
[…]
Chinese tech platforms have begun to crack down on black market ChatGPT access. By late February, WIRED found that the keywords “ChatGPT” and “OpenAI” have been banned on Taobao. On WeChat, “ChatGPT Online” and similar services have rebranded to neutral-sounding names like “AI Smart Chat.”
The intermediaries depend on APIs (which offer programmers access to the backend of the ChatGPT system) and on bulk-registered accounts. “These intermediaries profit by relaying ChatGPT’s service to users who do not have direct access. Just in this process alone, the parties involved would have violated ChatGPT’s terms and conditions, and other related trademarks and applicable patents,” says Ivan Wang, a New York-based IP attorney.
A creature is formed of clay. A puppet becomes a boy. A monster rises in a lab. A computer takes over a spaceship. And all manner of robots serve or control us. For generations we’ve told ourselves stories, using themes of magic and science, about inanimate things that we bring to life or imbue with power beyond human capacity. Is it any wonder that we can be primed to accept what marketers say about new tools and devices that supposedly reflect the abilities and benefits of artificial intelligence (AI)?
And what exactly is “artificial intelligence” anyway? It’s an ambiguous term with many possible definitions. It often refers to a variety of technological tools and techniques that use computation to perform tasks such as predictions, decisions, or recommendations. But one thing is for sure: it’s a marketing term. Right now it’s a hot one. And at the FTC, one thing we know about hot marketing terms is that some advertisers won’t be able to stop themselves from overusing and abusing them.
In less than two months, ChatGPT has sparked discussions around its potential to disrupt various industries, from publishing to law. But it’s already having a tangible impact on how a number of real estate agents around the country do their jobs – where much of the written work can be formulaic and time consuming – to the extent that some can no longer imagine working without it.
[…]
Asion has also used the tool to draft legally binding addendums and other documents, and sent them to lawyers for approval. “I fine-tune all kinds of drafts with ChatGPT,” he said. “Sometimes I’ll tell it to make it shorter or funnier, and it gives you so many samples to pick and edit from.”
[…]
Trelles said he uses ChatGPT to look up the permitted uses for certain land and zones in Miami-Dade County, and calculate what mortgage payments or return on investment might be for a client, which typically involve formulas and mortgage calculators.
“I can be in a car with a client when they ask me what their mortgage payments might be,” said Trelles. “I can ask ChatGPT what a mortgage payment would be on a $14 million purchase at a 7.2% interest rate amortized over 25 years with two origination points at closing, and in two seconds, it gives me that information. It also explains how it got the answer. It’s amazing.”
There are some limitations, however. The tool has, for example, struggled with some basic math before. Trelles said it’s helpful for approximations on the go, not for exact numbers.
I have a friend who works at a legal compliance/oversight agency and he's worried about this use case when it reaches the actual lawyers. Partly because the chatbots just aren't good enough yet,...
Asion has also used the tool to draft legally binding addendums and other documents, and sent them to lawyers for approval.
I have a friend who works at a legal compliance/oversight agency and he's worried about this use case when it reaches the actual lawyers. Partly because the chatbots just aren't good enough yet, they tend to turn out things that look like legal documents and there's potentially a very expensive difference between something that looks legally watertight to a robot and something that actually is. But that's probably solvable, a more specialised training set of actual legal documents, country specific and so on, would likely go a long way and I'm fairly sure there are already people doing that.
His main concern is that by automating the lower end part of the job of lawyering, junior staff won't get the experience they need to become less junior staff. If all your legal drafting is done by pressing a button, it's much harder to learn how to actually do legal drafting, which means you could end up never learning how to properly evaluate drafts - so eventually, when the people who check the chatbot output retire, who is left? Again, this is solvable but it means some fairly major changes in how legal firms operate and more relevant to my friend, in how the compliance system is structured.
One concern is that firms may need to train people to mid/high level before they start being financially useful, which isn't going to be popular at all. Imagine taking on staff and then spending 3-4 years training them before they make you any money! Especially when you consider that some people don't make it past being junior staff at all. You might train someone for a couple of years before you realise they're not cut out to be a higher level lawyer, or they might change their mind about wanting to be and then you're out a lot of time and money.
I second your friend's concern. I'm seeing exactly that play out in bookkeeping. New bookkeepers in high cost of living areas can't possibly compete with bookkeepers in low cost of living areas,...
I second your friend's concern. I'm seeing exactly that play out in bookkeeping. New bookkeepers in high cost of living areas can't possibly compete with bookkeepers in low cost of living areas, let alone complete automation. How do we get people with ten years of experience in an industry if we don't have ten years of increasingly complex tasks for them to do? More education before people can start working? If people need the equivalent of a masters degree before they're useful in the workplace, that's going to weed out a lot of people who might be great one day but can't afford six to eight years of higher education before they start making money.
So without being too much of a shill - a company I used to work for has a ChatGPT-like product that can also search the internet. It's on https://copy.ai (They also bought https://chat.new). I use...
So without being too much of a shill - a company I used to work for has a ChatGPT-like product that can also search the internet. It's on https://copy.ai (They also bought https://chat.new). I use ChatGPT every day for code but for writing or researching a tool that has more real-time data could be useful.
The official ChatGPT API went into general availability last week, with a 10x price drop compared to previous GPT APIs and a variety of developer-facing changes: [...] [...]
The official ChatGPT API went into general availability last week, with a 10x price drop compared to previous GPT APIs and a variety of developer-facing changes:
The ChatGPT model family we are releasing today, gpt-3.5-turbo, is the same model used in the ChatGPT product. It is priced at $0.002 per 1k tokens, which is 10x cheaper than our existing GPT-3.5 models. It’s also our best model for many non-chat use cases—we’ve seen early testers migrate from text-davinci-003 to gpt-3.5-turbo with only a small amount of adjustment needed to their prompts.
[...]
We are also now offering dedicated instances for users who want deeper control over the specific model version and system performance. By default, requests are run on compute infrastructure shared with other users, who pay per request. Our API runs on Azure, and with dedicated instances, developers will pay by time period for an allocation of compute infrastructure that’s reserved for serving their requests.
[...]
Over the past six months, we’ve been collecting feedback from our API customers to understand how we can better serve them. We’ve made concrete changes, such as:
Data submitted through the API is no longer used for service improvements (including model training) unless the organization opts in
Implementing a default 30-day data retention policy for API users, with options for stricter retention depending on user needs.
Removing our pre-launch review (unlocked by improving our automated monitoring)
A group from Stanford released Alpaca, an instruction-tuned version of Facebook's LLaMA 7B model: There's an online demo, but it is completely overloaded at the moment. First you have to wait 2-5...
A group from Stanford released Alpaca, an instruction-tuned version of Facebook's LLaMA 7B model:
We introduce Alpaca 7B, a model fine-tuned from the LLaMA 7B model on 52K instruction-following demonstrations. Alpaca behaves similarly to OpenAI’s text-davinci-003, while being surprisingly small and easy/cheap to reproduce (<600$).
There's an online demo, but it is completely overloaded at the moment. First you have to wait 2-5 minutes after you agree to the terms, with no visible progress (unless you crack open the developer tools and look at the websocket messages), and then, after you get to input your prompt, you have to wait in a queue again. I gave up after 15 minutes; I think it bugged out after it started the actual processing.
Llama is interesting in so far that it is supposed to give near-GPT3 levels of performance, but appears to be capable of running on consumer-grade hardware when quantized down to 4 bit precision.
I would expect to see near chatgpt3 levels of performance from open source autoregressive LLMs that can run on consumerish hardware within a month. Alpaca is mainly a proof of concept; 5 Stanford...
I would expect to see near chatgpt3 levels of performance from open source autoregressive LLMs that can run on consumerish hardware within a month. Alpaca is mainly a proof of concept; 5 Stanford students and $600 got them a lot of the way there. It's just a matter of time before, for better or for worse, something chatgpt3-like is out there, public and available, just with no guardrails this time.
@stu2b50 linked this here yesterday, but a ChatGPT-like locally executable LLaMa clone already exists. Edit: I just realized I was replying to you with your own link, I must have missed something...
The new one from the Stanford students is improved because it's trained to follow instructions. For plain LLaMA, you'd need to prompt it with some examples of what you want. I don't think there's...
The new one from the Stanford students is improved because it's trained to follow instructions. For plain LLaMA, you'd need to prompt it with some examples of what you want.
I don't think there's much point in downloading it either way, though? ChatGPT is convenient to try out and the standalone programs will be better in a month.
How to Learn and Teach Economics with Large Language Models, Including GPT (Tyler Cowen and Alexander T. Tabarrok) This is for economics but much of it applies more generally. Here's their summary...
So I finally sat down to play with one of these things. I used chat.openai.com. I found the experience fascinating. There is definitely some zany code crafting going on. But at no time did I feel...
So I finally sat down to play with one of these things. I used chat.openai.com. I found the experience fascinating. There is definitely some zany code crafting going on. But at no time did I feel like I was interacting with a human type intelligence.
One interesting dimension is that it can't seem to detect it's own logical consistencies and errors. I was inquiring about how it worked, and it maintained both that none of our conversation would have any impact on subsequent instances, and also that it continuously evolved by learning from user interaction.
There's also this exchange. I had previously pasted some code and asked it for debugging help, which it had none to offer. I asked it whether a span element could be nested, and whether there were any nested span elements in my code. It agreed that the element could be nested, but would not agree that there was a nested span in my code (there was). When I asked it (see the bottom of the quote) how its model would allow it to miss a detail, it paused for about 60 seconds before responding (the only delay in my few hours of experimentation), but I have to give it credit for a very sensible response.
can you provide an example of nested <span> element?
Sure, here's an example of nested <span> elements:
html
Copy code
<div>
<p>This is a <span class="red">red <span class="bold">and bold</span></span> text.</p>
</div>
In this example, there is a <span> element with a class of "red" that contains another <span> element with a class of "bold". The inner <span> element is nested inside the outer <span> element, and both are nested inside a <p> element, which is inside a <div> element.
The text "and bold" will be both red and bold because it inherits the styles of its parent <span> elements.
and you don't detect a similar span structure within the code I provided?
No, there are no nested span elements in the code you provided. Each span is a direct child of the div element.
<div class="center-screen">
<span class="rotateRed">W<span class="rotateBlue">h<span class="rotatePink">a<span class="rotateOrange">t</span></span></span></span>
</div>
There are no nested <span> elements in the code sample you provided. Each <span> element is a direct child of the <div> element.
does the code snippet just provided have a nested span?
Yes, the code snippet provided has a nested <span> element. The innermost <span> element with class "rotateOrange" is nested inside the <span> element with class "rotatePink", which is nested inside the <span> element with class "rotateBlue", which is in turn nested inside the <span> element with class "rotateRed".
How is that different from the code I previously provided?
The code snippet I provided has a nested span structure, where the span element with the class "rotateBlue" is nested within the span element with the class "rotateRed". In the code you previously provided, each span element was not nested within another span element.
I believe you are mistaken, you may want to let the development team know.
I apologize for any confusion. Can you please clarify what you believe I am mistaken about so that I can better address your concerns?
I believe the earlier code I provided also has nested span elements.
You are correct, I apologize for the confusion. The earlier code you provided also contains nested <span> elements. Specifically, each <span> element with a different class is nested within the previous <span> element. For example, the <span class="rotateBlue"> element is nested within the <span class="rotateRed"> element.
Why did you say differently previously?
I apologize for the confusion. You are correct that the earlier code sample you provided did have nested <span> elements. I must have missed that detail earlier. Thank you for bringing it to my attention.
How does your model allow you to miss a detail like that?
As an AI language model, I am capable of providing answers based on patterns and correlations within a vast corpus of text data. However, I do not have the capability to analyze code or syntax directly. My responses are based on statistical patterns in the text data that I have been trained on. Therefore, there may be instances where my responses may not be accurate or may miss specific details, especially in the case of code snippets where syntax is important. It is always important to verify information and double-check any code or technical information with a reliable source.
Overall I'm quite impressed with the product. I intend to utilize it in my work.
Other than the bing product, anyone know of interesting ai powered general purpose chatbots to play with?
There's character AI, which is more just for fun, but it's an LLM of some kind where users fine tune the model to behave like a specific character. Now that Facebook's LLaMa model has "leaked"...
There's character AI, which is more just for fun, but it's an LLM of some kind where users fine tune the model to behave like a specific character.
Now that Facebook's LLaMa model has "leaked" (not that it was that exclusive to begin with), there's a whole bunch of activity trying to get it to run on consumer hardware. If you have an Apple Silicon Mac, there's people running it at impressive speeds on M* hardware: https://github.com/ggerganov/llama.cpp
You can also "make your own" with OpenAI's chat completion API. I've been played around with it for some transfer tasks (mostly grammar correction - there's some translated things I want to read, but the translator's grasp on English was evidently not the best). It's very cheap.
My understanding is that each chat session is normally independent. However, OpenAI keeps chat session logs and can use them to improve the product via fine-tuning. So, it could have some effect...
My understanding is that each chat session is normally independent. However, OpenAI keeps chat session logs and can use them to improve the product via fine-tuning. So, it could have some effect on future chat sessions eventually, but it's unlikely to remember any specifics of what happened.
Anthropic claims that Claude is "much less likely to produce harmful outputs, easier to converse with, and more steerable" than other AI chatbots while maintaining "a high degree of reliability and predictability." The company cites use cases such as search, summarization, collaborative writing, and coding. And, like ChatGPT's API, Claude can change personality, tone, or behavior depending on use preference.
[…]
According to Anthropic, Claude has already been integrated into several products available through partners, such as DuckAssist instant summaries from DuckDuckGo, a portion of Notion AI, and an AI chat app called Poe that was created by Quora.
China’s ChatGPT Black Market Is Thriving (Wired)
[…]
ChatGPT results in someone getting an email about a non-existent book and gives someone Dave's phone number
Don't ask an AI for plant advice
The FTC has surprisingly poetic warning to businesses about false advertising.
From January: Real estate agents say they can’t imagine working without ChatGPT now (CNN)
[…]
[…]
I have a friend who works at a legal compliance/oversight agency and he's worried about this use case when it reaches the actual lawyers. Partly because the chatbots just aren't good enough yet, they tend to turn out things that look like legal documents and there's potentially a very expensive difference between something that looks legally watertight to a robot and something that actually is. But that's probably solvable, a more specialised training set of actual legal documents, country specific and so on, would likely go a long way and I'm fairly sure there are already people doing that.
His main concern is that by automating the lower end part of the job of lawyering, junior staff won't get the experience they need to become less junior staff. If all your legal drafting is done by pressing a button, it's much harder to learn how to actually do legal drafting, which means you could end up never learning how to properly evaluate drafts - so eventually, when the people who check the chatbot output retire, who is left? Again, this is solvable but it means some fairly major changes in how legal firms operate and more relevant to my friend, in how the compliance system is structured.
One concern is that firms may need to train people to mid/high level before they start being financially useful, which isn't going to be popular at all. Imagine taking on staff and then spending 3-4 years training them before they make you any money! Especially when you consider that some people don't make it past being junior staff at all. You might train someone for a couple of years before you realise they're not cut out to be a higher level lawyer, or they might change their mind about wanting to be and then you're out a lot of time and money.
I second your friend's concern. I'm seeing exactly that play out in bookkeeping. New bookkeepers in high cost of living areas can't possibly compete with bookkeepers in low cost of living areas, let alone complete automation. How do we get people with ten years of experience in an industry if we don't have ten years of increasingly complex tasks for them to do? More education before people can start working? If people need the equivalent of a masters degree before they're useful in the workplace, that's going to weed out a lot of people who might be great one day but can't afford six to eight years of higher education before they start making money.
Jack Clark's "Import AI" newsletter is a pretty good summary of what's happening in AI research. Here are some headlines from the latest issue.
So without being too much of a shill - a company I used to work for has a ChatGPT-like product that can also search the internet. It's on https://copy.ai (They also bought https://chat.new). I use ChatGPT every day for code but for writing or researching a tool that has more real-time data could be useful.
The official ChatGPT API went into general availability last week, with a 10x price drop compared to previous GPT APIs and a variety of developer-facing changes:
[...]
[...]
A group from Stanford released Alpaca, an instruction-tuned version of Facebook's LLaMA 7B model:
There's an online demo, but it is completely overloaded at the moment. First you have to wait 2-5 minutes after you agree to the terms, with no visible progress (unless you crack open the developer tools and look at the websocket messages), and then, after you get to input your prompt, you have to wait in a queue again. I gave up after 15 minutes; I think it bugged out after it started the actual processing.
Llama is interesting in so far that it is supposed to give near-GPT3 levels of performance, but appears to be capable of running on consumer-grade hardware when quantized down to 4 bit precision.
I would expect to see near chatgpt3 levels of performance from open source autoregressive LLMs that can run on consumerish hardware within a month. Alpaca is mainly a proof of concept; 5 Stanford students and $600 got them a lot of the way there. It's just a matter of time before, for better or for worse, something chatgpt3-like is out there, public and available, just with no guardrails this time.
Exciting times, in any case, in the LLM space.
@stu2b50 linked this here yesterday, but a ChatGPT-like locally executable LLaMa clone already exists.
Edit: I just realized I was replying to you with your own link, I must have missed something here. lol
The new one from the Stanford students is improved because it's trained to follow instructions. For plain LLaMA, you'd need to prompt it with some examples of what you want.
I don't think there's much point in downloading it either way, though? ChatGPT is convenient to try out and the standalone programs will be better in a month.
ChatGPT's Chess Elo is 1400
https://dkb.blog/p/chatgpts-chess-elo-is-1400
How to Learn and Teach Economics with Large Language Models, Including GPT (Tyler Cowen and Alexander T. Tabarrok)
This is for economics but much of it applies more generally. Here's their summary of their advice:
So I finally sat down to play with one of these things. I used chat.openai.com. I found the experience fascinating. There is definitely some zany code crafting going on. But at no time did I feel like I was interacting with a human type intelligence.
One interesting dimension is that it can't seem to detect it's own logical consistencies and errors. I was inquiring about how it worked, and it maintained both that none of our conversation would have any impact on subsequent instances, and also that it continuously evolved by learning from user interaction.
There's also this exchange. I had previously pasted some code and asked it for debugging help, which it had none to offer. I asked it whether a span element could be nested, and whether there were any nested span elements in my code. It agreed that the element could be nested, but would not agree that there was a nested span in my code (there was). When I asked it (see the bottom of the quote) how its model would allow it to miss a detail, it paused for about 60 seconds before responding (the only delay in my few hours of experimentation), but I have to give it credit for a very sensible response.
Overall I'm quite impressed with the product. I intend to utilize it in my work.
Other than the bing product, anyone know of interesting ai powered general purpose chatbots to play with?
There's character AI, which is more just for fun, but it's an LLM of some kind where users fine tune the model to behave like a specific character.
Now that Facebook's LLaMa model has "leaked" (not that it was that exclusive to begin with), there's a whole bunch of activity trying to get it to run on consumer hardware. If you have an Apple Silicon Mac, there's people running it at impressive speeds on M* hardware: https://github.com/ggerganov/llama.cpp
You can also "make your own" with OpenAI's chat completion API. I've been played around with it for some transfer tasks (mostly grammar correction - there's some translated things I want to read, but the translator's grasp on English was evidently not the best). It's very cheap.
My understanding is that each chat session is normally independent. However, OpenAI keeps chat session logs and can use them to improve the product via fine-tuning. So, it could have some effect on future chat sessions eventually, but it's unlikely to remember any specifics of what happened.
The stupidity of AI (archived)
Anthropic introduces Claude, a “more steerable” AI competitor to ChatGPT
[…]