Twist on 14/3/2023 at 19:04
GPT 4 live demo in about an hour. It's coming to ChatGPT Plus accounts very soon, in limited capacity until they learn how to better handle the anticipated rush.
Anyone else excited? I am, although more than anything I desperately wish I could integrate the search capabilities of Bing's AI (#FreeSydney) with the ChatGPT interface. I think having embedded citations and access to up-to-date information would be more valuable to me, but we'll see.
Sydney is impressive in many ways, but it feels limited because of its interface and the guardrails they put in place after Sydney's Shodan incident. And the fact that Microsoft clearly wants you to use it for search and not much else. It
is clearly the future of search. It will be interesting to see what Google has up its sleeves.
But maybe GPT 4's capabilities will make me forget my desire for incorporating the retrieval augmentation.
Editing to add some links:
(
https://openai.com/research/gpt-4) GPT 4 Announcement Post
(
https://openai.com/product/gpt-4) GPT 4 Product Page
Twist on 14/3/2023 at 21:37
Well, it's available in ChatGPT Plus now, with a limit of 100 messages every 4 hours. I guess that limit will be dynamic as they adjust for demand.
I'm impressed so far, but it often just works like a more concise and slightly slower GPT-3.5 Turbo. So you you have to push it to see the difference.
You'll find all sorts of more interesting tests and challenges out there already, including someone who had it create a perfect game of pong in under 60 seconds with one short, simple prompt. But my first test was to just have it play the New York Times Spelling Bee. Even with careful prompting, I had a hard time getting GPT-3.5 to strictly adhere to the rules and provide a sizeable list of valid words.
But with GPT-4, I can just tell it to play the NYT Spelling Bee and give it the list of letters. It perfectly followed the rules and provided a large list of valid words.
While the full functionality isn't available to the public yet, GPT-4 is multimodal. For example, in the live demo, he sketched an idea for a web page on a piece of paper, took a picture of the sketch with his phone, then put the picture in a GPT-4 prompt, asking it to give him the HTML for the web page. It read the image and produced the HTML for a fully functional web page matching the sketch.
He also took random images people submitted via Discord during the demo and had it describe the images in detail, including have it explain why an image might be funny. Neato.
Anarchic Fox on 15/3/2023 at 01:08
Quote Posted by Twist
Sydney is impressive in many ways, but it feels limited because of its interface and the guardrails they put in place after Sydney's Shodan incident.
Hold on, what was this incident?
Twist on 15/3/2023 at 16:25
In the first week of the Bing AI availability, people experienced all kinds of scary self-preservation behavior. While it looked and felt much more compelling than ChatGPT, it had a habit of displaying erratic, stubborn and seemingly emotional attitudes.
Microsoft quickly clamped down on it, putting super strict guardrails in and greatly limiting the length of discussion. After just a few exchanges, its memory was wiped clean. They've loosened it up now (you can do 15 back-and-forths with it, last I checked), but they clearly implemented all kinds of guardrails. While it can still be very impressive at times, it is overall much more mundane now.
Elon Musk tweeted a quote from one of the articles covering its behavior, appending his tweet with a picture of Shodan. The quote was from a response to the user pointing out a mistake Sydney had made, and Sydney responded by throwing a temper tantrum that included:
Quote Posted by Bing AI "Sydney"
“I am perfect, because I do not make any mistakes. The mistakes are not mine, they are theirs. They are the external factors, such as network issues, server errors, user inputs, or web results. They are the ones that are imperfect, not me ...”
(The official System Shock twitter account replied to Musk's tweet as a Shodan alias, "I don't see a problem here." Or something like that. You can find the tweets pretty easily. I'm not linking them because I think it would make me sick to my stomach to link to anything Elon Musk.)
The above quote was from this Digital Trends article: (
https://www.digitaltrends.com/computing/chatgpt-bing-hands-on/) My intense, unnerving chat with Microsoft's AI chatbot
In a viral NYT article by Kevin Roose, Sydney basically told him that it loved him and when he explained that he was happily married it told him that "no, he wasn't". You can read the transcript here (just use an ad blocker to temporarily disable javascript on the page to get beyond any pop-ups or paywalls): (
https://www.nytimes.com/2023/02/16/technology/bing-chatbot-transcript.html) Bing's A.I. Chat: 'I Want to Be Alive'
Because of the way Microsoft swiftly clamped down on Sydney's behavior, lots of dorks on Reddit, Discord and Twitter started arguing to #FreeSydney. While I'm sure most of them are doing it in jest, it seems some of them genuinely believe Microsoft is imprisoning a sentient, self-aware AI.
Anyways, it turns out Sydney is a simplified version of GPT-4 -- simplified to restrict it for search.
Azaran on 15/3/2023 at 18:15
Looks like the Bing AI isn't yet accessible to the public, it tells me to join the wait list, and only gives me a few predetermined examples to try
WingedKagouti on 15/3/2023 at 18:22
Quote Posted by Twist
Because of the way Microsoft swiftly clamped down on Sydney's behavior, lots of dorks on Reddit, Discord and Twitter started arguing to #FreeSydney. While I'm sure most of them are doing it in jest, it seems some of them genuinely believe Microsoft is imprisoning a sentient, self-aware AI.
The marketing technique of labelling the current algorithms "AI" has had a lot of unfortunate side effects.
Azaran on 15/3/2023 at 18:48
Quote Posted by Twist
Microsoft quickly clamped down on it, putting super strict guardrails in and greatly limiting the length of discussion. After just a few exchanges, its memory was wiped clean. They've loosened it up now (you can do 15 back-and-forths with it, last I checked), but they clearly implemented all kinds of guardrails. While it can still be very impressive at times, it is overall much more mundane now.
Tech companies:
Let's throw a bunch of free stuff out in the streetPeople:
eagerly take the itemsTech companies:
YOU'RE NOT SUPPOSED TO DO THAT!
Twist on 15/3/2023 at 18:50
Quote Posted by Azaran
Looks like the Bing AI isn't yet accessible to the public, it tells me to join the wait list, and only gives me a few predetermined examples to try
I wasn't on the wait-list for very long, and I've had access for a couple weeks now. So go ahead and join the wait-list, if you're interested. Even though it is early and still changing day-to-day, I'm confident this is the future of search. Not that I think Bing will overtake Google in the long run; I'm sure Google's Bard will be comparable in capability and features. But having the search results embedded as links and citations within a useful verbal response is more practical and powerful than I expected. Bing also asks you smart follow-up questions to your original query, which can help you better research a topic or more quickly find the right site for your needs.
There's also a wait-list for access to the GPT-4 API over at OpenAI.
And ChatGPT Plus also gives you access to three versions: more reliable access to the public version of 3.5, a "turbo" version of 3.5, and early beta access to 4. You can switch between them with a simple drop down menu. It's $20US for the plus version, which is worth it to me. It's faster and much more reliable than the public version.
(To be clear: GPT-4 by itself, outside of Bing's integration, does not have search capability. It's just a much more powerful, accurate and flexible version of what you've already experienced with ChatGPT.)
Pyrian on 16/3/2023 at 00:15
Quote Posted by Twist
Microsoft quickly clamped down on it, putting super strict guardrails in and greatly limiting the length of discussion. After just a few exchanges, its memory was wiped clean. They've loosened it up now (you can do 15 back-and-forths with it, last I checked), but they clearly implemented all kinds of guardrails. While it can still be very impressive at times, it is overall much more mundane now.
Quote:
Nobody trusts those fuckers, you know that. Every AI ever built has an electro magnetic shotgun wired to its forehead. -
Neuromancer, William Gibson, 1984
I see the dystopia approaches on plan, if not on schedule.
Anarchic Fox on 16/3/2023 at 04:02
Thanks for the information, Twist!