SenseNova

What SenseNova is

SenseNova (日日新) is a set of foundation models from the Chinese artificial intelligence company SenseTime, presented publicly at the company's Tech Day event on 10 April 2023^[1][3]. The launch introduced not a single model but a group of large-model products: the SenseChat conversational chatbot, a text-to-image generator and an avatar video generator^[1]. A foundation model, in this sense, is a large model trained on broad data that can later be adapted to many different tasks, rather than a system built for one narrow job.

Where it came from

Chinese Wikipedia records the April 2023 launch of the SenseNova large model system at a SenseTime technology day^[4]. The launch placed SenseTime beside other Chinese companies that had already put their own chatbots in front of the general public: Baidu's ERNIE Bot and Alibaba's Tongyi Qianwen had opened to public testing^[1].

What was presented

SenseTime said SenseChat has more than 100 billion parameters, while the image generator has 1 billion and a visual model 32 billion^[1]. TechNode reported that the over-100-billion figure for SenseChat came from co-founder Wang Xiaogang^[1]. Parameters are the numbers inside a neural network that are adjusted during training; a larger count generally gives a model more capacity, at the cost of more computing power to train and run. On stage, CEO Xu Li demonstrated SenseChat live, with storytelling, code writing and medical advice^[1]. The three demonstrations covered free-form text, program code and health-related questions.

The three products

SenseChat handles text conversation; the text-to-image generator turns written descriptions into pictures; the avatar video generator produces video of a virtual presenter. They take different inputs and produce different outputs, and the models behind them differ in size: more than 100 billion parameters for SenseChat, 1 billion for the image generator and 32 billion for the visual model^[1]. Parameter counts are not directly comparable across tasks, because models built for different jobs use different structures and training data.

Who could use it at launch

At launch the products were not open for public testing; access was reserved for government and corporate clients^[1]. That set SenseNova apart from Baidu's ERNIE Bot and Alibaba's Tongyi Qianwen, which had already opened to public testing^[1]. The first users were therefore organisations putting the models into their own systems rather than individuals signing up for a consumer app.

Market reaction

The day after the event, SenseTime shares rose as high as HK$3.70, up 11.1%, though that price was still below the HK$3.85 IPO price; the stock closed at HK$3.30^[2].

Later use

By February 2024 SenseTime said that more than 3,000 enterprise users had adopted SenseNova since its launch^[3]. The figure describes business adoption rather than consumer sign-ups, in line with the arrangement at launch, when access was limited to government and corporate clients^[1].