The mysterious model has gone viral in three days since its launch. Netizens have found a large number of traces of OpenAI.

The mysterious model has gone viral in three days since its launch. Netizens have discovered a large number of traces of OpenAI.

Some netizens tried to challenge it with MC-Bench to generate Minecraft-style scenes and compared it with 4o-mini. The result was clear at a glance. Some people also systematically tested its programming level and found that Optimus Alpha is the best-performing model in the Ruby language. Some even directly praised that Optimus Alpha must be the state-of-the-art (SOTA). While amazed by its excellent performance, the mysterious identity of Optimus Alpha has also sparked speculation... Million-context window, for real-world tasks. Optimus Alpha supports a million-context window with a maximum output of 32K. And the response speed is very fast, with a median first Token latency of only 0.81 seconds and a median output speed of 24.8 Tokens per second. At the same time, the introduction mentions that Optimus Alpha is mainly for real-world tasks, especially programming. A blogger asked it to write an e-commerce website with a shopping cart function. As a result, Optimus Alpha designed a reasonable UI interface, and the shopping cart function, which many other AIs failed to handle, can work normally. It also works fine when crossing different files. Or when asked to write a Snake game, not only can it work properly, but it also adds clever designs such as color changes of the snake's head and body color gradients, outperforming some other AI programming tools in terms of new ideas. Some even used it to develop an OCR text recognition application that supports handwritten text. In terms of performance, its Elo score is 1338, ranking second on the list, second only to Claude 3.7 Sonnet, ahead of DeepSeek-R1 and Quasar Alpha, which is suspected to be the predecessor of Optimus Alpha. Especially in SQL database query tasks, Optimus Alpha achieved the highest average score. The Aider list shows that the programming ability of Optimus Alpha is close to that of Quasar Alpha, Grok 3 and medium-sized o3-mini, slightly better than GPT-4.5-preview. In addition to programming, Optimus Alpha also performs excellently in creative writing, with its Elo score ranking fourth, behind DeepSeek-V3. Does the mysterious model come from OpenAI? The simplest and most straightforward way to investigate is to directly ask the model to confess. Since the purpose of releasing the model is to collect feedback, Optimus Alpha can be used for free on OpenRouter, making it possible to conduct experiments. When asked about its identity, Optimus Alpha said without hesitation that it is ChatGPT. If asked about the specific version, the answer is "Based on GPT-4, with knowledge cutoff in June 2024." In addition, some people directly associate the name Optimus with Tesla's Optimus robot and think that the mysterious model comes from Musk. But some people think this is a trick by Altman. If you believe it comes from a company under Musk, you will fall into Altman's trap. More convincing evidence comes from Quasar Alpha, which has been taken offline. It first appeared on the 2nd of this month. A netizen on Reddit found that when trying to use Quasar Alpha for illegal operations, the model's way of refusal was very similar to that of OpenAI. The Tokenizer bug mentioned by this netizen refers to the phenomenon that someone earlier found that Quasar Alpha had the same "read but randomly reply" problem as GPT-4o when performing Chinese-English translation tasks. This bug seems to be unique to OpenAI, and it does not occur on Grok, Claude, or DeepSeek. Some people even carried out more complex analysis - AI researcher Sam Paech (he also initiated the creative writing list mentioned above) tried to establish the relationship between models using informatics methods through the differences in the model's answers. As a result, Paech found that Quasar Alpha is extremely similar to OpenAI's models, especially pointing out GPT-4.5-preview. Later, Altman also hinted at the identity of Quasar Alpha in a tweet. Finally, we can return to Optimus Alpha. Tests found that the same bug in ChatGPT and Quasar Alpha also appeared in it. Paech also has new results. Optimus Alpha was added to the latest pedigree chart, and the model closest to it is ChatGPT-4o updated on March 27 this year. From a time perspective, Quasar Alpha was taken off the shelves the day after Optimus Alpha was launched. Therefore, some people think that Optimus Alpha is a replacement for Quasar Alpha. In addition to various signs observed in experiments, community testing of new models in the form of mysterious models has been a traditional practice of OpenAI. Coupled with Altman's hint about Quasar Alpha, the probability that Optimus Alpha comes from OpenAI is overall very high. As for more specific details, combined with the recently leaked "GPT-4.1" of OpenAI, which is regarded as an upgrade of GPT-4o, and the confirmation of Paech's latest pedigree chart... What do you think is the true identity of this mysterious model? Reference links: [1]https://x.com/TheMattBerman/status/1910813233008509191[2]https://www.reddit.com/r/LocalLLaMA/comments/1jrd0a9/chinese_response_bug_in_tokenizer_suggests/[3]https://x.com/sam_paech/status/1910346895110848553 ...

PC version: https://www.cnbeta.com.tw/articles/soft/1492502.htm
Mobile version: https://m.cnbeta.com.tw/view/1492502.htm

via cnBeta.COM Chinese Industry Information Station - Telegram Channel