Just now GPT-4.5 Turbo was suspected to have been leaked by the OpenAI network team!

Search engines such as Bing and DuckDuck Go have already indexed GPT-4.5 Turbo's product pages before GPT-4.5 Turbo was officially released.

When you click on the index link to enter GPT-4.5 Turbo, you will enter a 404 page, but you can see some teaser text on the page, saying that GPT-4.5 Turbo is the fastest, most accurate, and most scalable model OpenAI has ever created.

Netizens speculate that the official release date of GPT-4.5 Turbo may be this Thursday - the anniversary of the release of GPT-4, and OpenAI CEO Sam Altman has also recently subscribed to Lex Fridman's podcast. After the release of GPT-4 last year, Altman Went to Fridman's podcast.

On March 9, Sam Altman hinted on his Twitter that the wait was worth it ~

So let me take a look at what information is "spoiled" by GPT-4.5 Turbo?

Have a larger context window

The size of the context window of the large language model determines how many words the model can process at the same time. In the preview, the window size of GPT-4.5 Turbo is 256,000 Tokens, which is approximately equivalent to 200,000 words, twice the size of GPT-4 Turbo 128K.

According to experts, Google is also testing the large model Gemini 1.5, which also has similar potential.

More important than the amount of input data is the reliability of the data. Current models with large context windows tend to selectively ignore information, making them less valuable for text analysis of large documents. If the GPT-4.5 Turbo can perform better at this match point, then it will be a huge victory, even if some current performance points do not widen the gap with other models.

Have the latest information

The trailer also revealed that the latest version of GPT-4.5 Turbo is as of June 2024, the so-called "knowledge deadline". Netizens speculate that this date seems to indicate that the new Turbo will be released in June.

Do you have video capabilities?

Since December last year, there have been rumors that GPT-4.5 Turbo is coming. The new features mentioned include text and image processing, and may also include video or 3D functions. Netizens have speculated whether it will be compatible with Sora's Is it combined with video capabilities? However, the leaked information did not mention this capability.

Is it a leak or an error?

The OpenAI developer community has already discussed this matter, and many people think this is a big mistake! People who frequently read technology blogs said that they have not seen relevant information on other platforms.

Whether it is an error or a leak, it seems that we need to further observe!

Is the Transformer Debugger a smoke bomb?

Just yesterday, the OpenAI Super Alignment Team officially announced that it would open source the Transformer Debugger (TDB) for internal use.

In short, researchers can use the TDB tool sparse autoencoder, combined with the "automatic interpretability" technology developed by OpenAI - that is, using large models to automatically explain small models, technology.

At this point in time, it’s hard not to suspect that OpenAI is starting to release smoke bombs!

In addition, the University of California, Irvine (UCI) recently issued an announcement:

They are testing their own GPT4.5-based chatbot ZotGPT, which provides cost-free access to GPT-4.5 and supports chat history, file uploads and large context windows. This may also hint that GPT-4.5 Turbo may be coming soon!

The current spoiler is that GPT-4.5 Turbo may be released in June 2024. However, such a delayed release is not OpenAI’s usual method. In the past, OpenAI's usual practice was to launch a new model every time and make it available immediately after release. So netizens speculate that OpenAI may have broken some defenses recently! Competitor Anthropic recently released Claude 3, which surpassed GPT-4 in many tests.

Therefore, if OpenAI wants to continue to be the leader in large models, it must defeat Claude 3 as soon as possible. Even if the current development progress of "Winning Match Point" is only on paper, OpenAI still needs to find a way to return to the throne as soon as possible to prevent the loss of customers to Anthropic.