Do large language models dream of artificial intelligence agents?
During sleep, The human brain is arranged through different memories, and while not important, combines important issues. What if AI can do the same?
Bilt, a company that offers local shopping and restaurant transactions to the lessee has recently set up several million factors with hope to do so.
Bilt uses technology to start a starting name called Letta, which allows delegates to learn from previous conversations and share their memories. Using a process called “Calculating Bedroom”, the delegates decide what information to store in their long -term memory vault and what is needed to call faster.
“We can make a single update for [memory] “This is useful in any scenario you want to control the text of the representatives,” says Andrew Fitz, an artificial intelligence engineer at Bilt.
If information is included in the context window, large language models can usually only “remember” things. If you want to remember a chatbot to remember your latest conversation, it should be attached to the chat.
Most artificial intelligence systems can only control a limited number of information in the context window before their ability to use data and make them hallucinated or confused. In contrast, the human brain is able to destroy useful information and remember it later.
“Your brain is constantly improving and adding more information like sponges,” says Charles Packer, CEO of Letta. “With language models, exactly the opposite. You run these language models long enough in a loop and the field is poisoned; they disappear and you just want to reset.”
Packer and his founder Sarah Wooders had previously developed MEMGPT, an open source project that decided to help the LLMs decided to store in short and long -term memory. With the Letta, this dual has expanded its approach to teach the delegates in the background.
Bilt’s collaboration with Letta is part of the wider pressure to provide useful information to save and remember useful information to AI, which can make your chat more intelligent and less prone to error. According to the experts I talked to, memory is not developed in modern artificial intelligence, which weakens the intelligence and reliability of artificial intelligence tools.
Harrison Chase, founder and CEO of Langchain, another company that has created a way to improve memory in artificial intelligence representatives, says he considers memory as an essential part of the field engineering – where the user or engineer decides what information to feed in the field. Langchain offers companies a variety of memory for delegates, from long -term facts to users to memories of recent experiences. “Memory, I argue, it is some kind of background.” “A large part of a work of an artificial intelligence engineer is essentially acquired [information]. “
Consumer AI tools are also less forgotten. In February, OpenAI announced that ChatGPT stores relevant information to provide more personal experience to users – though the company did not disclose how it works.
Letta and Langchain make the call process more transparent for AI system engineers.
“I think not only is it important for opening models but also for open memory systems,” says Clem Delangue, CEO of AI Hounding Face and an investor in Letta.
Interestingly, the CEO of Letta Packer points out that it may also be important for artificial intelligence models to forget. “If a user says,” The project we were working on, wipe it off your memory “, the agent should be able to go back and rewrite each memory.”
The concept of artificial memories and dreams makes me think of Do Androids see the dream of an electric sheep? By Philip K. Dick, a curved mind novel that inspires the stylish dystopian movie Runner of the bladeHuman beings of the big models of language are still not as impressive as the rebellious repetitions of the story, but their memories appear to be equally fragile.
This is a copy of Will Knight AI Laboratory NewsletterRead the previous newsletters Here
