谷歌、Anthropic和OpenAI齐陷“僵局”，AI模型开发瓶颈如何破？

Source: Zhitong Finance "Since 1950, the S&P 500 index has risen more than 10% 21 times as of the end of May. In about 90% of these cases, the S&P 500 index rose for the rest of the year. There were only two instances of declines for the rest of the year, in 1987 (-13%) and 1986 (-0.1%)." With the rebound of the stock market, the old adage "Sell in May and Go Away" seems to have been a bad advice once again. Last month, the S&P 500 index rose 4.8%, the best May performance since 2009. The NASDAQ 100 index rose nearly 6.2%, and the NASDAQ Composite Index rose 6.9%. Goldman Sachs FICC & Equities Trading Division said: "History doesn't really support this saying. Don't sell, leave the market (go on vacation), and enjoy the good times." The rising trend is still to be continued? If history is any guide, it may indicate that the rise of the stock market is not over yet. Looking ahead to the rest of 2024, Scott Rubner, Managing Director of the Goldman Sachs Global Markets Division and tactical expert, pointed out the following historical background for investors. Rubner stated that the S&P 500 index has risen 10.7% year-to-date, and since 1950, the S&P 500 index has risen more than 10% 21 times as of the end of May. In about 90% of these cases, the S&P 500 index rose for the rest of the year. There were only two instances of declines for the rest of the year, in 1987 (-13%) and 1986 (-0.1%). "Since 1950, the median return of the last 7 months of each year (June 1 to December 31) is 5.4%. In the aforementioned 21 cases, the average performance of the last 7 months increased to 8.1%." Rubner added. Rubner also pointed out that the NASDAQ index has risen for 16 consecutive Julys, with an average return of about 4.64%.

According to insiders, like OpenAI, the performance of siasun robot&automation's large language model supported by Anthropic has also reached a bottleneck period of stagnation. Three informants revealed that the upcoming Gemini has not met internal expectations. In addition, Anthropic has postponed the next version of its Claude model, 3.5 Opus. Companies state that there is still a lack of artificial data for training models.

OpenAI, currently at a milestone key moment, completed the first round of training for a large new ai model in September, hoping the model will significantly surpass the previous versions of the technology behind ChatGPT, moving closer to the goal of powerful ai that surpasses human capabilities. However, according to two informed sources, this internally named ‘Orion’ model did not meet the company's expected performance.

Informed sources stated that, for example, by the end of summer, Orion performed poorly when attempting to answer programming questions that had not been trained. Overall, so far, the enhancement of Orion over OpenAI's existing models is not as significant as the upgrade from GPT-3.5 to GPT-4. GPT-3.5 was initially the driving system of OpenAI's flagship chat robot.

OpenAI is not the only company that has encountered obstacles recently. After launching increasingly complex ai products at a rapid pace for years, the three leading ai giants are now finding that the returns from the high costs spent to develop new models are diminishing.

Anthropic released a new version named Claude 3.5 Haiku last month, along with an upgraded version of Claude 3.5 Sonnet. In September of this year, OpenAI launched a series of new ai models named o1 and o1-mini, capable of reasoning over complex tasks and solving problems that are more difficult than previous scientific, programming, and mathematical models.

However, according to three informed sources, the upcoming Gemini software version from Google, under Alphabet, did not meet internal expectations. Meanwhile, the long-awaited release schedule for Anthropic's Claude model 3.5 Opus has also been delayed, without a specific release date provided.

These AI giants are currently facing some challenges. Finding new and untapped high-quality human training data sources to build more advanced ai systems is becoming increasingly difficult. Two insiders stated that Orion's programming performance is not ideal, partly due to the lack of sufficient programming data for training. Meanwhile, even a modest improvement may not justify the enormous costs associated with manufacturing and operating new models, as well as meeting the expectations of marketing the product as a significant upgrade.

However, these models still have considerable potential for improvement. According to one of the insiders, OpenAI has been putting Orion through a months-long process known as "post-training." This is a routine procedure the company follows before publicly releasing new ai software, which includes incorporating human feedback to enhance its responses and improving the model's tone of interaction with users. However, an insider indicated that Orion still has not reached the level that OpenAI wishes to be ready for user release, and the company is unlikely to launch the system before early next year. The Information had previously reported some details about the challenges OpenAI faces in developing new models, including programming tasks.

Additionally, OpenAI CEO Sam Altman has previously stated that the overall level of computing power is also an issue that the latest wave of ai models needs to address. Altman recently said on Reddit: "All of these models are becoming quite complex, and we cannot launch so many things simultaneously." "We also face many constraints and tough decisions about how to allocate our computing resources to many great ideas."

Editor / jayden