ChatGPT-4 Also ‘Hallucinates’ Facts a Bit, but Already a Hit
After igniting a global frenzy with its AI chatbot ChatGPT four months ago, a product that racked up 100 million subscribers in record time, OpenAI’s follow-up version GPT-4 is already here.
The ChatGPT 3.5 model has created such excitement that even those who cannot access it officially, like in China and some countries in Africa, want it and are using it.
Harnessing the hype, OpenAI was already hard at work improving their model and have now launched a newer edition, ChatGPT-4. Although at the moment it is only available to ChatGPT Plus subscribers per the OpenAI website, reviews are pouring in to reveal how the new model exceeds the older version, ChatGPT 3.5.
According to OpenAI, ChatGPT-4 is a much enhanced chatbot as engineers “found and fixed some bugs” and improved some “theoretical foundations.”
“As a result, our GPT-4 training run was (for us at least!) unprecedentedly stable, becoming our first large model whose training performance we were able to accurately predict ahead of time,” explained the company.
“As we continue to focus on reliable scaling, we aim to hone our methodology to help us predict and prepare for future capabilities increasingly far in advance – something we view as critical for safety.”
Users give latest chatbot thumbs up
One user named Linus shared a Twitter thread of users discussing how favorably ChatGPT-4 compares to its predecessor. Now, with the ability to take image and text as input, the latest model lets users express themselves better.
Rowan Cheung posted a video showing how ChatGPT can turn a website sketch into a functional website, which probably takes less time that it would a web developer. He captioned the video “This is insane.”
Another user, Ammaar, posted how he created a snake game using the technology and deploy it to Replit in 20 minutes. But that’s not all. ChatGPT-4 is also being used for match-making purposes.
“It takes profile data & preferences, determines if the match is worth pursuing and automates the followup. With computer vision for the physical, you can filter anything and find your ideal partner,” said Jake Kozloski.
Conor also shared a post in which he uploaded an Ethereum contract into GPT-4 to see if it would highlight some vulnerabilities. Amazingly, ChatGPT-4 managed to come up with some issues with the provided contract. The contract, which is from 2018 for the Proof of Weak Hands (PoWH) coin, was hacked a few hours after the bug had been identified.
The same vulnerabilities that were exposed by the hackers are the same ones highlighted by ChatGPT in Conor’s post. This could be a game changer in assessing contract vulnerabilities before launching a coin.
The latest ChatGPT is already making its way into some third-party products, including Microsoft’s Bing.
“We are happy to confirm that the new Bing is running on ChatGPT-4, which we’ve customized for search,” Microsoft confirmed on Tuesday.
“If you’ve used Bing preview at any time in the last few weeks, you’ve already experienced an early version of this powerful model.”
Microsoft invested billions of dollars into OpenAI and integrated the conversational AI tool into its Bing search engine early last month, helping it cross 100 million daily users for the first time.
There are still flaws with ChatGPT -4
OpenAI claims the new ChatGPT-4 can pass a simulated bar test with a score in the top 10% of test takers, whilst the 3.5 model test score was in the bottom 10%.
On its website, OpenAI claims the differences between ChatGPT -4 and ChatGPT 3.5 are subtle in a casual conversation, however the newer model is more creative and reliable once a certain complexity threshold is passed.
ChatGPT-4 has the ability to use both text and images as inputs, though the result provided is still text-based. The addition of extra inputs is undoubtedly a huge plus for users.
Although the new model comes with several upgrades, it still faces the same challenges as previous models. OpenAI highlighted how it “hallucinates” facts but still makes reasoning errors, issuing a fair warning to users to be extra cautious when using outputs from language models by always applying human review, especially on important contexts.
“It is still flawed, still limited, and it still seems more impressive on first use than it does after you spend more time with it,” said OpenAI CEO Sam Altman in a series of tweets announcing the update on Tuesday.
Despite the notable improvements, the company still advises users to be cautious.
“It is more creative than previous models, it hallucinates significantly less, and it is less biased,” wrote Altman.
This article is originally from MetaNews.