ChatGPT's memory function has been completely upgraded, making Altman excited to the point of insomnia last night. However, an even greater surprise awaits - GPT-4.1, o3, and o4 models may be released next week.
Last night, Altman's preview post received over 2 million views, completely igniting widespread speculation across the internet.
What excited him to the point of sleeplessness was ChatGPT's enhanced memory capability!
Starting today, ChatGPT's memory function can reference all past chat records and provide more personalized responses.
Moreover, it will provide more ease in writing, obtaining advice, learning, and other aspects based on personal preferences and interests.
Additionally, it can quote previous chat records to offer more tailored and valuable responses.
New conversations will naturally unfold based on its existing understanding of you, making interactions smoother and truly customized for you.
Once the memory space is full, no new memories can be created. However, you can also forget existing memories to free up space.
As before, ChatGPT's memory function can be turned off at any time in the settings by choosing to stop "referencing past chat records" or completely disabling the memory function.
Currently, all Plus and Pro users can use the latest memory function, while Team, Enterprise, and Edu users will receive this feature in the coming weeks.
ChatGPT Will Remember Everything You've Told It
What does an AI that remembers everything about your past mean?
OpenAI researcher and poker pioneer Noam Brown wrote that memory signifies not just the birth of a product, but a paradigm shift from situational interaction to continuous interaction.
OpenAI co-founder Greg Brockman stated that ChatGPT now has long-term memory of your past conversations, which is a true step towards a personal intelligent agent!
A KOL in AI said that ChatGPT's cross-session memory is a massive leap and could be a significant turning point.
Because for a truly useful AI, memory is a crucial missing piece. From now on, ChatGPT will transform from a one-time tool to a genuine assistant, and may ultimately become a "second self".
Finally, ChatGPT asked a heart-piercing question: Do you want me to turn this into a Twitter-length rant or continue with sarcasm?
AI KOL Bao Yu also had ChatGPT create a self-portrait based on historical records.
However, after actual testing, some people expressed great disappointment. They originally thought that ChatGPT with memory could search for related questions in past conversations, but it turned out that it absolutely cannot!
On the contrary, it produces severe hallucinations. For example, when discussing a paper called Skillweaver, it started talking nonsense. This problem even exists in professional accounts.
Some also complained that after most conversations reach a certain length, AI remembers all incorrect information and becomes ineffective.
Although nothing big was discovered this time, OpenAI's new model is really about to go online.
GPT-4.1 to Launch Next Week?
Verge exclusively revealed that OpenAI is preparing to release a series of brand new models.
Among them is a model expected to be named GPT-4.1, reportedly an improved version of the GPT-4o multimodal model, which may debut as early as next week.
At the same time, they will also release smaller GPT-4.1 mini and nano versions.
Additionally, the highly anticipated full-version o3 and mysterious o4 series models (o4-mini, o4-mini-high) will also be unveiled simultaneously.
Last night, AI engineer Tibor Blaho discovered references to these models in the ChatGPT webpage code.
The day before, internal code had already leaked that OpenAI is preparing to launch o4 and o3 models
On the other hand, Altman had previewed the release of o3 and o4-mini models in a post last weekend.
The new models are confirmed, and now we just wait for the release.
References:
https://x.com/sama/status/1910380644972265603
https://x.com/OpenAI/status/1910378768172212636
https://x.com/dotey/status/1910436681330897405
https://www.theverge.com/news/646458/openai-gpt-4-1-ai-model
This article is from the WeChat public account "New Intelligence", author: HYZ, published by 36Kr with authorization.
