Microsoft are holding a lot of great cards in the AI game, especially ChatGPT-3, but Google also have a great hand, in fact they have a bird in the hand:
LamDA (language Model for Dialogue Applications) is Google's main LLM, branded as Bard. This has been in the wings for some time but will be launched soon.
Hassabis has promised some sort of release in 2023.
Their goal is to reduce wrong and invented information by linking it to Google Search and Scholar for citations. Stating the source(s) is important to overcome the nervousness people have about the validity of the content. It may also frighten those who see these tools as blowing up citation-based assessment and hey may have a solution to this problem.
Google’s goal is a more laid back goal to build on Google Tools such as Search and Scholar to deliver factually correct text. They also have Lamda, remember the chatbot that a Google developer thought was sentient? RanBran , is AI deeply embedded inside tier search engine and they have acquired well.
To achieve this they want to use a form of reinforcement learning (RL) based on people's actual feedback, to improve performance. This is clever as it can determine whether output is adequate, needs more or even a citation.They also stress test the system by getting humans to break the system.
They recognise that real dialogue with trustworthy answers, requires additional rules that constrain that dialogue. The quickest way to determine those rules is from real human judgements.
Of course, OpenAI will release with ChatGPT-4, a product that is several orders of magnitude bigger than Chat GPT-3. Google are old hands, have the processing power to deliver and may want the ChatGTP to take the heat, then release afterwards. The stakes are high. The winner adds billions in value.
No comments:
Post a Comment