AI21 CEO says transformers not right for AI agents due to error perpetuation

As extra enterprise organizations look to the so-called agentic future, one barrier could also be how AI fashions are constructed. For enterprise AI developer A121, the reply is obvious, the {industry} must look to different mannequin architectures to allow extra environment friendly AI brokers. 

Ari Goshen, AI21 CEO, mentioned in an interview with VentureBeat that Transformers, the preferred mannequin structure, has limitations that might make a multi-agent ecosystem troublesome.

“One development I’m seeing is the rise of architectures that aren’t Transformers, and these various architectures can be extra environment friendly,” Goshen mentioned. “Transformers operate by creating so many tokens that may get very costly.” 

AI21, which focuses on growing enterprise AI options, has made the case earlier than that Transformers needs to be an choice for mannequin structure however not the default. It’s growing basis fashions utilizing its JAMBA structure, brief for Joint Consideration and Mamba structure. It’s primarily based on the Mamba structure developed by researchers from Princeton College and Carnegie Mellon College, which may supply sooner inference instances and longer context. 

Goshen mentioned various architectures, like Mamba and Jamba, can typically make agentic buildings extra environment friendly and, most significantly, reasonably priced. For him, Mamba-based fashions have higher reminiscence efficiency, which might make brokers, significantly brokers that hook up with different fashions, work higher. 

He attributes the rationale why AI brokers are solely now gaining recognition — and why most brokers haven’t but gone into product — to the reliance on LLMs constructed with transforms. 

“The primary cause brokers aren’t in manufacturing mode but is reliability or the dearth of reliability,” Goshen mentioned. “Once you break down a transformer mannequin, you recognize it’s very stochastic, so any errors will perpetuate.”

Enterprise brokers are rising in recognition

AI brokers emerged as one of many largest traits in enterprise AI this 12 months. A number of firms launched AI brokers and platforms to make it straightforward to construct brokers. 

ServiceNow introduced updates to its Now Help AI platform, together with a library of AI brokers for patrons. Salesforce has its secure of brokers referred to as Agentforce whereas Slack has begun permitting customers to combine brokers from Salesforce, Cohere, Workday, Asana, Adobe and extra. 

Goshen believes that this development will develop into much more fashionable with the right combination of fashions and mannequin architectures. 

“Some use circumstances that we see now, like query and solutions from a chatbot, are mainly glorified search,” he mentioned. “I feel actual intelligence is in connecting and retrieving completely different data from sources.”

Goshen added that AI21 is within the means of growing choices round AI brokers.

Different architectures vying for consideration

Goshen strongly helps various architectures like Mamba and AI21’s Jamba, primarily as a result of he believes transformer fashions are too costly and unwieldy to run. 

As a substitute of an consideration mechanism that kinds the spine of transformer fashions, Mamba can prioritize completely different knowledge and assign weights to inputs, optimize reminiscence utilization, and use a GPU’s processing energy. 

Mamba is rising in recognition. Different open-source and open-weight AI builders have begun releasing Mamba-based fashions up to now few months. Mistral launched Codestral Mamba 7B in July, and in August, Falcon got here out with its personal Mamba-based mannequin, Falcon Mamba 7B.  

Nonetheless, the transformer structure has develop into the default, if not normal, selection when growing basis fashions. OpenAI’s GPT is, after all, a transformer mannequin—it’s actually in its identify—however so are most different fashionable fashions. 

Goshen mentioned that, finally, enterprises need whichever strategy is extra dependable. However organizations should even be cautious of flashy demos promising to unravel a lot of their issues. 

“We’re on the section the place charismatic demos are straightforward to do, however we’re nearer to that than to the product section,” Goshen mentioned. “It’s okay to make use of enterprise AI for analysis, but it surely’s not but on the level the place enterprises can use it to tell selections.”