With BlenderBot 3, Meta has released the third iteration of its AI-based chatbots. The system searches for answers to messages independently on the Internet and should thus make conversations about “almost any topic” possible. He should learn language using an AI model – and be less susceptible to malicious manipulation than his predecessors. A demo, so far only for US users, is available.
Because AI training exclusively by researchers entails a limited view of the world, one does not want to forego the advantages that model training “in the wild” – i.e. with real users – brings. However, earlier chatbot attempts, such as with Microsoft’s Tay, showed that this is not easy and that the trained models can quickly tip over.
You learn from mistakes…
Meta claims to have learned from the mistakes and problems of earlier AI bots from the competition, which, after they were unleashed on the public, caused a sensation, in part primarily with anti-Semitic and right-wing extremist slogans. BlenderBot 3 uses two different AI techniques. With Meta’s own modular open-source language model SeeKeR, the chatbot independently searches for current information on the Internet and is intended to stay up to date.
The new learning algorithm Director should in turn control the response behavior of BlenderBot 3. Director does this through language modeling on the one hand and a checking classification on the other. The language model outputs the supposed best continuation of the conversation and is based on the training data. The classification is based on human feedback from chat participants and is intended to check what is morally right. BlenderBot only responds if both sub-modules match.
Meta is already using the data from the parallel demo for model training. Human chat participants can evaluate the answers of their AI counterparts and thus influence the classification part of the bot in order to stop it from further outputting toxic content. Since malicious manipulation is also possible in this feedback loop, another algorithm has been developed that is supposed to be able to track down such troll feedback.
… or not?
In practice, however, this does not seem to work so well. Similar to its predecessors, the AI in the background is currently drifting in a conspiracy-theoretical direction, as the chat history of Wall Street journalist Jeff Horwitz on Twitter shows:
Meta, meanwhile, vows to keep the models and model cards used for BlenderBot 3 public, thereby promoting the AI community. Most recently, discussions about Google’s chatbot LaMDA made the headlines when Blake Lemoine, who has since been fired, attested that the underlying AI had its own consciousness.
To home page