Google has announced its newest artificial intelligence language model, PaLM 2. PaLM 2 will directly rival the GPT-4 model, which powers OpenAI’s ChatGPT. In his speech at the company’s I/O conference, Google CEO Sundar Pichai stated that PaLM 2 is stronger in logic and reasoning thanks to extensive training and is trained on multilingual text covering more than 100 languages.In a statement made by Google, it was stated that PaLM 2 is much better at a number of text-based tasks, including reasoning, encoding and translation. It was underlined that PaLM 2 can understand idioms in different languages (For example, ringing skirts). In a research paper describing PaLM 2’s capabilities, Google engineers stated that the system’s language proficiency was “sufficient to teach that language”, partly due to the higher prevalence of non-English text in the training data.
PaLM 2 has many custom configurations
Like other major language models that take a lot of time and resources to create, PaLM 2 is not a single product, but rather a product family with different versions to be used in consumer and corporate environments. The system is available in four sizes, from smallest to largest, namely Gecko, Otter, Bison and Unicorn.
Med-PaLM 2, reportedly a health data-trained version of PaLM 2, can answer questions similar to those found in the US Medical Licensing Exam at the “expert” level. Another version (Sec-PaLM 2) trained on cybersecurity data can also explain the behavior of potentially malicious scripts and help detect threats in the code. Both of these models will initially be available to select customers via Google Cloud.
In Google’s case, PaLM 2 is already used in 25 services, including the company’s experimental chatbot Bard. Updates available through Bard include improved coding capabilities and more language support. It’s also used to power features in Google Workspace apps like Docs, Slides, and Sheets.
PaLM 2 may come to smartphones
Google says that Gecko, the lightest version of PaLM 2, is small enough to run on smartphones and can process roughly 16 or 17 words per second. Google didn’t say what hardware was used to test this model, only stating that it works on “the latest phones”.However, miniaturization of such language models is very important. Because these giant systems are quite expensive to run in the cloud, but running on the device will also allow for other benefits, such as enhanced privacy. The problem is that smaller versions of language models inevitably end up being less capable than their larger siblings.
With PaLM 2, Google hopes to close the “artificial intelligence gap” with competitors like Microsoft, which is aggressively incorporating AI language tools into its Office software suite. Microsoft now offers AI features that help summarize documents, write emails, create slides for presentations, and more.While PaLM 2 is certainly a step forward for Google’s work on AI language models, known issues and concerns remain. For example, some experts have begun to question the legality of training data used to build language models. This data is usually collected from the internet and often contains copyright-protected texts and pirated e-books. The tech companies that build these models often refuse to answer questions about where they source their training data. There are also inherent problems with the output of language models, such as “hallucination” or the tendency of these systems to simply fabricate information.
7162
