Source URL: https://www.theregister.com/2024/10/02/china_telecom_model_trained_local_tech/
Source: The Register
Title: China trains 100-billion-parameter AI model on home grown infrastructure
Feedly Summary: Research institute seems to have found Huawei to do it – perhaps with Arm cores
China Telcom’s AI Research Institute claims it trained a 100-billion-parameter model using only domestically produced computing power – a feat that suggests Middle Kingdom entities aren’t colossally perturbed by sanctions that stifle exports of Western tech to the country.…
AI Summary and Description: Yes
Summary: China Telecom’s AI Research Institute has successfully trained a 100-billion-parameter model, TeleChat2-115B, utilizing solely domestically produced computing power, demonstrating resilience against Western tech export sanctions. This development highlights the increasing capabilities of Chinese AI infrastructure and its implications for global AI competitiveness.
Detailed Description:
The announcement from China Telecom’s AI Research Institute marks a significant advancement in domestic AI development amidst mounting external pressures:
– **Model Training**: The TeleChat2-115B model, boasting 100 billion parameters, was trained entirely on domestic infrastructure, utilizing 10 trillion tokens from a high-quality corpus in both Chinese and English.
– **Technological Independence**: This achievement underscores China’s ability to circumvent sanctions that restrict the export of advanced Western technology, revealing a shift towards self-reliance in AI capabilities.
– **Infrastructure Details**:
– The training utilized the Ascend Atlas 800T A2 training server, a product by Huawei, indicating a growing synergy between domestic chip production and AI model training.
– Huawei’s Kunpeng processors, created with the Arm architecture, played a central role in this setup, although the model’s relatively low parameter count suggests that it may not have required cutting-edge GPU technology traditionally associated with high-performance AI computations.
– **Comparison with International Models**: TeleChat2-115B’s parameter count is less than leading models such as Llama and OpenAI’s o1, suggesting a competitive but not superior position in terms of sheer scale, although parameter count does not alone determine a model’s efficacy.
– **Strategic Implications**: China Telecom’s position as a major player, with over 500 million subscriptions and significant revenue, indicates a robust resource pool for AI development. Coupled with its advocacy for open-source communities such as OpenStack, this suggests that China is bolstering its infrastructure capabilities while fostering innovation in the AI landscape.
Overall, this development not only illustrates the strides China is making in AI but also indicates potential shifts in the global tech ecosystem as countries adapt to new limitations on technology access. Security and compliance professionals should be aware of these advancements and the geopolitical implications surrounding AI technology ownership and production.