Source: Brokerage China
Author: Shi Qian

OpenAI may release GPT-4.5 to counter Musk's Grok3. Today, OpenAI's CEO and co-founder Altman suddenly stated, "For high-demand testers, the experience of trying GPT-4.5 brings a sense of AGI that is much more profound than I expected!" This indicates that GPT-4.5 has entered the testing phase and is very close to its official release.

Altman's move seems aimed at countering Musk. At noon on February 18, Beijing time, Musk's xAI launched the third generation of large language model Grok3. Musk stated that Grok3 performed "better than any product already released" in testing. Additionally, a post by Musk on Social Media ignited interest in the Technology and Gaming Industry: xAI will establish an AI gaming studio. This also sparked enthusiasm in the A-share market for gaming stocks.

OpenAI suddenly announced.

At 1 AM today, OpenAI's CEO and co-founder Altman stated, "For high-demand testers, the experience of trying GPT-4.5 brings a sense of AGI that is much more profound than I expected!" This indicates that GPT-4.5 has entered the testing phase and is very close to its official release.

Recently, Altman also unveiled the development roadmap for the highly anticipated GPT-5 model. GPT-5 is the successor to the globally sensational GPT-4 AI language model in 2023. Altman mentioned in response to questions on the X platform that GPT-5 will be released in a few months, hinting it may launch in late 2025.

Altman provided a detailed explanation on the X platform, stating that the company plans to release GPT-4.5 (internally codenamed Orion) "within weeks," which will be OpenAI's last non-simulated reasoning model. Simulated reasoning models (like o3) use special techniques to iteratively process user-posed questions but are slower than traditional large language models (like GPT-4o) and are not suitable for all tasks.

Subsequently, GPT-5 will become a system that integrates all current features of OpenAI's AI models, including traditional AI models, SR models, and specialized models for web search and research. "We will release GPT-5 in ChatGPT and the API, as a system that integrates multiple technologies including o3," he wrote, "We will no longer release the o3 model separately."

Altman stated that ChatGPT free users will gain "unlimited GPT-5 conversation access under standard intelligence settings," while ChatGPT Plus users can run GPT-5 at a higher intelligence level, and Pro users can operate GPT-5 at the highest intelligence level. He also mentioned that these integrated AI models will be able to utilize several advanced ChatGPT features from OpenAI, including advanced voice modes, Canvas, search, and deep research. Simplifying the complex product line, Altman previously commented on the arbitrary nature of future GPT-5 labels, suggesting that several AI models released by the company last year could have been named GPT-5.

Musk's big move.

Interestingly, Musk also released what he called the "most powerful model on the planet" Grok3 today.

According to reports, in terms of performance in mathematical reasoning, scientific logical reasoning, and code writing abilities, Grok-3 has achieved better results than DeepSeek-v3, GPT-4o, and Gemini-2pro in multiple benchmark tests. Grok3 has reportedly been running internally at xAI for two weeks.

Musk founded the artificial intelligence company xAI as early as 2023. However, the development of Grok can be traced back to 2015, when Altman and 11 others co-founded OpenAI. However, Musk and the OpenAI team had ideological differences and gradually drifted apart, ultimately leaving. Later, OpenAI launched the AI chatbot program ChatGPT at the end of 2022, sparking a global craze.

The term Grok originates from American author Robert Heinlein's 1961 novel "Stranger in a Strange Land", and it is a word from the Martian language he created, meaning to understand through empathy or intuitively. The term was later included in the Oxford English Dictionary and gradually made its way into computer programming culture.

So, how exactly does this Grok model perform? According to industry professionals, first of all, from the comparison of model performance, in the new high school competition benchmark tests, larger models outperform smaller models, demonstrating stronger generalization capabilities; however, the situation was the opposite in last year's testing, reflecting the true generalization situation of the model. Over the past 17 months, Grok has evolved from being nearly unable to solve high school problems to possessing strong capabilities.

Second, from the perspective of capability transfer, Grok not only trains reasoning skills in mathematical and coding comparison problems, but can also solve various tasks such as games. It has learned to detect and correct its own errors, gaining generalization abilities from math and coding to solve other problems.

Third, in terms of Hardware, the team has built a GPU cluster and tested the stability of training settings through cable insertion and removal tests to enhance reliability. It is predicted that machines will win awards in fields such as the Fields Medal and Nobel Prize within the next two to three years.

Fourth, in the direction of practical application demonstration, Grok provided a Python code solution for the physical problem of spacecraft transfer between Earth and Mars. Although its correctness has not been confirmed, the simulation effect is close to the actual situation. Additionally, Grok generated Python code similar to the game Tetris.

Should computing power and electrical utilities both be on fire?

Currently, it seems that the competition in AI is intensifying, and both computing power and electrical utilities may be key.

Cai Xin Securities stated that although DeepSeek-R1 has significantly reduced training computing power expenses and model inference costs, computing power remains essential for model performance support. The explosive growth of AI application scenarios will push overall computing power demand to a new level.

Xinda Securities also believes that WeChat's integration with DeepSeek-R1 to gray test the AI search function, using public account articles and web search strategies, is expected to promote upgrades in AI applications and improve efficiency across various industries, paving the way for the realization of complex AI scenarios such as humanoid robots and autonomous driving. Currently, the number of DeepSeek users is rapidly increasing, and the demand for computing power is surging, suggesting a focus on the computing power-related industrial chain.

The market may still be neglecting electrical utilities to some extent. Yesterday, GD Power Development saw a large increase with a volume over 300%, and stocks like China National Nuclear Power also strengthened noticeably. However, overall, this Sector has not yet clearly started. According to a data report from January, PJM Interconnection LLC, which manages electrical systems from Washington D.C. to Illinois across 13 states, recently projected that by 2035, summer peak electrical demand will increase by nearly 58 gigawatts (about 38% growth) from the historic high set last year, reaching about 210 gigawatts. According to Statistics, this expected increase in electrical demand will be more than double the highest demand ever recorded in the New England grid. The main reason is the vigorous development of AI.

In addition, the demand for electrical equipment will also increase.

Editor/Rocky