GoogleDeepMind recently announced the launch of PaLM 2, a massive language model (LLM) that supports Bard and 25 other product features. PaLM 2 outperforms the previous version of PaLM in various benchmark tests, while being smaller in scale and lower in operating costs.
Google CEO Sundar Pichai announced this model at Google I/O '23. PaLM 2 now performs well in various tasks, including code generation, inference, and multi-processing language, and offers four different model sizes, including a lightweight version called Gecko, designed for mobile devices. When evaluated in natural language processing (NLP) benchmark tests, PaLM 2 performs better than PaLM and achieves new state-of-the-art levels in many tasks, especially in the BIG-bench benchmark test. In addition to powering Bard, the new model serves as the foundation for many other products, including Med-PaLM 2, an LLM fine-tuned specifically for the medical field, and Sec-PaLM, a model for cybersecurity. According to Google, PaLM 2 marks an important milestone in the development of LLM and natural language processing technology, demonstrating Google's commitment to advancing the field.
PaLM 2 showcases the impact of high-performance models with different scales and speeds, as well as the practical benefits of multi-functional AI models for everyone. However, as we strive to release the most helpful and responsible AI tools, we are also working to create Google's best foundational model to date.
In 2022, InfoQ reported on the initial release of Pathways Language Model (PaLM), a massive language model (LLM) with 540 billion parameters. PaLM achieved state-of-the-art performance on several inference benchmarks and demonstrated its capabilities in two novel inference tasks: logical reasoning and joke interpretation.
For PaLM 2, Google implemented several improvements to enhance model performance. First, they studied the model scaling laws to determine the optimal combination of training computation, model size, and data size. They found that, given a fixed computational budget, data and model size should scale roughly in a 1:1 ratio, whereas previous researchers scaled model size up to three times the data size.
The team improved PaLM 2's multilingual capabilities by adding more languages to the training dataset and updating the model training objectives. The original dataset was English-centric, while the new dataset was extracted from a more diverse range of languages and domains. PaLM 2 not only uses language modeling objectives but also employs "adapted mixtures" training objectives.
Google evaluated PaLM 2 on six broad NLP benchmark tests, including inference, encoding, translation, question answering, classification, and natural language generation. The focus of the evaluation was to compare its performance with the original PaLM. On the BIG-bench, PaLM 2 showed "significant improvements," and even the smallest PaLM 2 model achieved "competitive" performance in classification and question answering compared to larger PaLM models. In inference tasks, PaLM 2 demonstrated similar performance to GPT-4 in a "competitive" manner, surpassing GPT-4 on the GSM8K mathematical reasoning benchmark.
"In discussions on Reddit, some users commented that while the output of PaLM 2 may not be as good as GPT-4, it is clearly better." One user said, "They may want this model to be highly scalable so that it can be implemented for free or at a low cost in their products. At the same time, it is able to provide search results without taking too long. I have been using GPT-4 and I really like it, but its speed is well known. With the new Bard based on PaLM 2, it is even faster than GPT 3.5 Turbo."
Source: https://www.infoq.com/news/2023/06/google-palm2-bard/
Alphabet Q1 paid subscribers reached 350M, fueled by YouTube & Google One. YouTube ad revenue slightly missed at $9.88B amid Premium growth. Total Q1 revenue hit $109.9B, beating forecasts.
Anthropic releases Claude connectors for Adobe, Blender, and Ableton, enabling AI workflow automation and app control. It also sponsors the Blender Development Foundation.
Apple introduces an App Store model splitting 12-month plans into monthly installments. On iOS 26.4+, it lowers upfront costs without reducing total price, adding billing tracking.
Google launches Aadhaar digital IDs in Wallet for India, enabling secure QR/NFC verification. A phased partner rollout targets key sectors, while passport-based credentials expand globally.
YouTube tests "Ask YouTube," an AI tool blending text and video for complex queries. Initially for US Premium users, it supports interactive follow-ups, expanding Google's AI search.
Google grants US DoD classified AI access, joining OpenAI and xAI in defense deals amid Anthropic's dispute. Contractual safeguards against surveillance and lethal weapons remain unenforceable.
Mistral AI launches Workflows, a Temporal-powered orchestration platform for secure, Python-driven enterprise AI automation featuring human-in-the-loop approvals.