TF Securities: Claude 3 releases a report emphasizing the ongoing focus on AI investment opportunities.

Zhitong App learned that TF Securities has released a research report stating that AI has made continuous progress recently. Overseas, multimodal technologies continue to catalyze with the successive launch of Sora, StableDiffusion3, and Google's world model Genie, showcasing a flourishing development of AI models. The continuous evolution of multimodal models and the iterative upgrades of AI models are expected to drive a competition in computing power under the forefront of AI models. The demand for computing power is expected to remain high, emphasizing the benefits of AI hard technology core targets, such as optical modules and server switch upgrades accelerating further. Looking forward to the practical application of AI, it is crucial to continue focusing on core beneficiary varieties in the AI sector.

Investment recommendations: ① Optical modules & devices: Zhongji Innolight (300308.SZ) and others; ② Server switches and PCB: WUS (002463.SZ) and others; ③ Smart computing: Runze Technology (300442.SZ) and others.

Preface:

In the early hours of March 5th, Anthropic, mainly invested by Google and Amazon, launched three new AI models, namely Claude3Haiku, Claude3Sonnet, and Claude3Opus, in ascending order of intelligence level and cost.

▍Key points from TF Securities:

Significant performance improvement, model performance surpassing GPT-4

① Each Claude3 model has its own characteristics: Opus leads in performance among the three, as shown in the test results released by Anthropic, outperforming GPT-4 in 8 tests such as MMLU, GPQA, and GSM8K, demonstrating human-like understanding and fluency; Sonnet strikes a balance between processing speed and computational efficiency, while maintaining cost-effectiveness, making it advantageous in enterprise-level applications requiring large-scale deployment; Haiku is the fastest and smallest model of the three, capable of reading a highly dense paper (about 10,000 tokens) on the arXiv website in just 3 seconds.

② In comparison, Claude3Opus model's performance completely surpasses GPT-4 and Gemini1.0Ultra. Claude3Sonnet outperforms GPT-4 in some benchmarks like GSM8K. Claude3Haiku can compete with Gemini1.0Pro. This progress in large model iterations continuously breaks through in performance, raising expectations for the dynamics of GPT-5 and its performance improvement.

③ The instruction compliance capability of Claude3 models has significantly improved, reducing the likelihood of refusing to answer prompts close to system guardrails, showing a more detailed understanding of requests, being able to identify real violations of content policies, and significantly reducing the frequency of refusing harmless prompts.

• Claude3Opus surpasses GPT-4Turbo in reducing model "illusions," which means providing incorrect information output. Compared to Claude2.1, Claude3Opus has doubled its performance in giving correct answers and reducing incorrect answers.

• The Claude3 model family offers a context window of 200K and can handle inputs of over 1 million tokens, which helps in better understanding and memorizing long-text information.

Advanced multimodal processing capabilities, upgraded consumer usability; soaring professional domain performance, leading in business applicability

Claude3 has sophisticated visual capabilities comparable to other leading models in the industry. It can handle visual formats including photos, charts, graphics, and technical diagrams, enhancing visual abilities similar to GPT-4V and Google's Gemini. Currently, the multimodal functionality is open to enterprise customers, with some encoding up to 50% of their knowledge base in various formats such as PDFs, flowcharts, or presentation slides.

At the same time, Claude3 has released performance benchmarks in specific fields such as finance, medicine, and philosophy, showing significant improvements compared to earlier versions like Claude2.1.

Risk Warning:

Risks include AI progress falling below expectations, algorithm construction falling short, and increased competition.