After the strong performance exhibited by Meta’s chatbots BlenderBot and BlenderBot 2, Meta AI launched BlenderBot 3, a 175B-parameter, publicly available chatbot. Built with all the skills of its previous versions such as long-term memory, personality and empathy, BlenderBot 3 shows a 31% improvement in overall performance related to conversational tasks. 

Meta AI incorporated several new features in the latest version of the chatbot. For example, the model has been trained using a new learning algorithm, Director, that enables the bot not only to provide the most relevant responses to users’ input but also distinguish between right and wrong. New algorithms were used to ensure that the bot is able to distinguish between harmful responses and harmful examples. BlenderBot 3 uses a new safety recovery technique to respond to challenging conversations in a more civil manner. The most remarkable aspect is the option of a live interactive demo that enables BlenderBot 3 to learn from organic interactions with all kinds of people.

Notwithstanding the new and improved features of the latest version, what has triggered a lot of discussions is the ‘publicly available’ aspect of BlenderBot 3. 

How ‘public’ is BlenderBot3?

In its blog, Meta announced it is a publicly available chatbot. In fact, it has provided links to get the code and datasets. “In addition to sharing the model weights and code, we’re publishing new datasets and model cards so that other researchers can leverage BlenderBot 3 in their own work. We believe this open research approach will accelerate progress in conversational AI,” tweeted Meta AI. 

For a moment, any reader might be tricked into thinking that like BlenderBot 2, the latest version is also open-source. Well, like they say, all that glitters is not gold. In the case of BlenderBot 3, one can say all that is publicly available and available for open research is not open source.

The moment one goes to access the model code, one gets to know that only the smaller models, 3b and 30b, can be accessed. For the larger 175b model, one needs to request access. 

While requesting access, one needs to provide details like the organisation one is affiliated to, intended use and others. In fact, the Google Doc page displays a message – “To access the largest model, please fill out this form, and we’ll let you know if your use case is approved. If you have a .edu email address, please use that instead of any other personal email address”. From the message it is tacit that someone with a ‘.edu’ email address would be preferred more than the others. 

As a footnote to the blog, Meta cites that access to the 175b parameter model will be limited. Access will be granted to academic researchers and people affiliated to government organisations, civil society groups, academia and global industry research labs. Thus, what Meta says to be publicly available isn’t easily accessible to many sections of the public.

Further, discussions on Reddit and Twitter bring to light that access to Blender Bot 3 is restricted not just to certain groups of people but also based on region. A Twitter user shares a screenshot of a window that says “BlenderBot is not available in your location”. 

In fact, if any user who isn’t from the US, attempts to deploy the demo version, the window conveys that ‘BlenderBot is US-only at the moment’.

That’s startling. If access is indeed region-based why call it ‘publicly available’? Well, there could be several reasons for this restricted availability.

Some opinions doing the rounds in social media regarding geographical restrictions is probably because of the imminent blackout Facebook could face in Europe if the Irish Data Protection Commission’s draft decision to block Meta from sending data across the Atlantic turns true. Perhaps, Meta isn’t very happy with such decisions being deliberated upon. 

“Zuck’s probably angry with the EU forcing them out and all that controversy about the metaverse,” commented a Reddit user. 

As stated by Meta, BlenderBot 3 is available for public demo. However, a public demo is not without challenges. There is no guarantee that users deploying the demo version are well-intentioned and would not engage in toxic conversations that could be offensive and disrespectful. This  could be a reason for the restricted access. “We believe that long-term safety is an important component of quality chatbots — even if it means sacrificing engagingness in the short term,” notes Meta in the blog. 

Meta seemed to be aware of the criticisms it could face once the BlenderBot 3 releases. Thus, it conveyed its plan to improve the models using feedback from demo users’ interaction and release updated models for the benefit of the larger AI community.