"PRICE WARS are ten a penny in China. The emergence of hundreds of lookalike companies seemingly overnight is pushing down retail prices of everything from electric vehicles to bike-sharing and bubble tea. The latest products to enter the ruinous fray are artificial-intelligence (AI) chatbots. This may seem surprising. Until recently China’s problem was not a surfeit of large language models (LLMs), the sort that makes ChatGPT a humanlike content-creator, but their dearth. At the start of 2023 experts reckoned that the Chinese LLMs that did exist were a decade behind the American cutting edge.
Now the same experts reckon that the transpacific model gap has shrunk to a year or less.
LMSYS Chatbot Arena, which ranks LLMs’ performance, puts five Chinese ones in the world’s top 20.
On some measures, a model called SenseNova 5.0, created by SenseTime, a Shanghainese AI firm, outperforms the LLM which until an upgrade in May powered ChatGPT in logical reasoning and creative writing. In February AskCI, a market-research firm, forecast that Chinese LLMs will generate 22bn yuan ($3bn) in revenue this year, up from 15bn yuan in 2023. By 2028 they could be generating five times as much.
All these eye-catching technological strides and the promise of riches have motivated plenty of AI hopefuls to pile in. Last year Robin Li, chief executive of Baidu, a Chinese search giant with AI ambitions of its own, observed that China was producing one LLM a day. He may have been exaggerating for effect but was not wide of the mark. According to government estimates, the country now boasts more than 100 models with more than a billion parameters, which is to say at least roughly as complex as some versions of the popular Llama models made by Meta, Facebook’s parent company.
Chinese AI developers are now turning these models into products left and right. But in contrast to America, where OpenAI commands a clear technological lead, in China users lack such an industry benchmark to help them differentiate between the sundry products on offer, says Jeffrey Ding, a scholar of Chinese AI at George Washington University. Unable to compete on technology, they are therefore competing on price.
On May 6th High-Flyer, a quant hedge fund that built its own model, slashed prices for the latest version to one-hundredth of one yuan cent per 1,000 tokens (the preferred unit of LLM pricing).
That is approximately 1% of what OpenAI charges for its snazziest model, GPT-4 Turbo. On May 15th ByteDance, which owns TikTok and its Chinese sister video-app, Douyin, began offering its newest model at a similar discount. A week later Alibaba, an e-commerce conglomerate, cut prices for its flagship LLM by 97%. Hours later Baidu announced that its Ernie chatbots will be free of charge for all business users. Tencent, China’s biggest internet firm, is also giving away one of its LLMs for nothing.
One Tencent executive now talks of “the war of a hundred models”. Lee Kai-fu, a Taiwanese tech investor and author of “AI Superpowers”, a bestselling book from 2018, has warned that this conflict is a “lose-lose” one.
Attracting more customers—and therefore hoarding more data on which to train ever cleverer algorithms—counts for little if the resulting lower revenues mean much less money to spend on the pricey computing resources necessary to do the training.
This problem afflicts model-makers everywhere but it is especially acute for Chinese ones. That is because they are barred by American sanctions from gaining access to American-made AI chips, which are the most powerful and efficient in the world.
In the longer run, the likeliest consequence of the price war is consolidation of the Chinese AI industry in the hands of a few deep-pocketed digital giants such as Alibaba, Baidu, ByteDance and Tencent. Promising smaller model-makers, such as Baichuan, Moonshot and Zhipu AI, could be squeezed out of the domestic market—never mind that, as OpenAI and fellow upstarts such as Anthropic and Mistral have shown in the West, newcomers have a record of outinnovating incumbents in machine learning.
Some of the Chinese challengers may try their luck in international markets instead. 01AI, Mr Lee’s latest venture, has signalled that if the price war gets out of control it is prepared to look abroad for growth. But many domestic rivals would struggle to navigate foreign technology markets and so are stuck in their hypercompetitive home market. A hundred LLMs may be blooming in China. Many of them may wither much sooner than their creators expected." [1]
What about AI in Chinese dragon killer Lithuania? Crickets... Lithuanian gut is to thin to produce anything contemporary. We are only buying overpriced and obsolete tanks, Chinese drones, and diesel smoke belching military transporters. Country side ladies are excited to see our diesel smoke, and can't stop waiving at us enthusiastically. Power... So much pleasure for so little work.
1. The LLM summer sale. The Economist; London Vol. 451, Iss. 9401, (Jun 15, 2024): 56, 57.
Komentarų nėra:
Rašyti komentarą