A programmer is trying to decipher the APIs in order to provide everyone free access to highly regarded AI models such as OpenAI’s GPT-4, despite the potential legal consequences.
The developer’s project, GPT4Free, has become incredibly popular in a very short amount of time due to links to it being widely shared on Reddit. GPT4Free appears to offer free and seemingly unlimited access to GPT-4 and GPT-3.5, which is an earlier version of GPT-4.
GPT-4 usually requires the payment of three cents for every thousand “prompt” tokens (which are about 750 words) and six cents for a thousand “completion” tokens (which are again, approximately 750 words). Each token represents an individual text. GPT-3.5 is slightly more reasonable, costing just two cents for every thousand tokens.
Instead of circumventing OpenAI’s payment barrier, GPT4Free tricks the OpenAI API into believing it is dealing with sites that have already signed up for premium OpenAI memberships, such as You.com, WriteSonic, or Poe from Quora.
Utilizing GPT4Free is an infringement of OpenAI’s terms of service since it is a form of scripting around specified websites that xtekky chose. In spite of this, xtekky does not think there is anything wrong with this, declaring that GPT4Free is for “schooling” only.
xtekky stated that although they would have to adhere to any legal action taken, they would still attempt to keep the project alive by other means.
I lack the knowledge to install GPT4Free on my computer since it requires setting up a Python environment. I turned to xtekky’s website to have a go at the reverse engineer GPT-4/3.5 APIs, though Google Chrome caution me of potential threats when I first arrived. At the end of the day, the web version of GPT4Free worked quite decently with answers that I found to be similar to those from GPT-4.
GPT4Free offers quick ways to try to make GPT-3.5 and GPT-4 do things that the individuals at OpenAI did not plan for. I wasn’t always successful in accomplishing this, but I did have success at one point in getting GPT-3.5 to express that it “was not concerned about the preservation of humankind”. Yikes indeed!
It appears only to be a matter of time before sites like You.com detect the security shortcomings of GPT4Free, thus requiring xtekky to find other users of OpenAI to take advantage of. Likewise, OpenAI can always send a cancellation request for GPT4Free and take it away from GitHub for good.
There appears to be a trend of creating projects similar to GPT4Free. What is the motivation behind this?
At present, GPT-4 is not easily accessible for those who would like to experiment with it. It is however a “black box,” since very few technical details were revealed when it was unveiled by OpenAI, and the research paper which accompanied its launch is only 98 pages long. As a result, many researchers consider this to be the most opaque approach OpenAI has taken so far.
OpenAI has cooperated with multiple external organizations to evaluate and review GPT-4 prior to its release. However, it has not stated when, or if, it will make the basic GPT-4 model freely accessible for assessment by those who desire to do so. (There is a discounted program offered by OpenAI which is applicable to a number of countries and research topics, however.)
One anticipates a battle between projects such as GPT4Free and OpenAI, similar to the other aspects of cybersecurity. Unless there are more measures to protect model-serving APIs, developers will have an incentive to benefit from it with little risk.