Join daily and weekly newsletters to obtain the latest updates and exclusive content to cover the leading artificial intelligence in the industry. Learn more
A tripartite partnership between the artificial intelligence phone support company VoiceThe improvement platform for reasoning MaitaiAnd chips maker Groc It has achieved a penetration that addresses one of the most stable artificial intelligence problems in the conversation: the embarrassing delay that immediately indicates the callers speaking to a machine.
Vocal cooperation has enabled more than 70 % response times to increase accuracy from 81.5 % to 99.2 % through four typical repetitions, bypassing GPT-4O’s 94.7 % measurement By 4.5 degrees Celsius. Improvements from the possibility of the new GROQ stem on the alteration immediately between multiple artificial intelligence models without an additional transition time, coordinated through the MAITAI improvement platform.
Achievement is what the industry experts call “Wadi Gharib“From AI Voice-Micro Signals that make automatic conversations feel clearly inhuman. For communication centers and customer service operations, transformational effects can be: one of Phonely customers is to replace 350 human agents this month alone.
Why still phone calls from artificial intelligence seem robotic: four seconds problem
Traditional large language models such as Openai’s GPT-4O I have long fought with what appears to be a simple challenge: responding quickly enough to maintain the flow of natural conversation. While a few seconds of delay barely recorded in the text -based reactions, the stopping itself appears to be unacceptable during live phone conversations.
“One of the things that most people do not realize is that the main LLM service providers, such as Openai and Claude, and others who have a very high degree of cumin contrast,” said Will Bodies, founder and executive head in Phoenay, in an exclusive interview with Venturebeat. “4 seconds looks eternal if you are talking to AI’s voice on the phone-this delay is what makes most of the artificial intelligence voice today does not feel the non-human form.”
The problem occurs almost once all requests every ten requests, which means that standard conversations include at least one or two of the temporary suspension that immediately reveal the artificial nature of the interaction. For companies that are considering artificial intelligence agents, these delays have created a large barrier in front of adoption.
“This type of cumin is unacceptable to support the phone in real time,” Bodes explained. “Regardless of the cumin, the accuracy of the conversation and the human -like responses are something that Llm Legacy has not met in the sound world.”
How did three startups solve the biggest challenge in artificial intelligence?
The solution appeared from the development of Groq for what the company calls “Zero Lora Hotswapping-The ability to switch immediately between multiple artificial intelligence model variables without any performance penalty. Lora, or low-ranking adaptation, allow developers to create lightweight modifications specific to current models instead of completely new training from scratch.
“Groq Mix from the architectural engineering controlled by programs with microcredit, high -speed memory on chip, architecture, and inevitable implementation, means that it is possible to access the multiple hot Loras,” Chelsey Kantor, the chief marketing employee in GROQ, explained in an interview with Venturebeat. “Loras is stored and managed in SRAM alongside the original typical weights.”
This progress in the MAIAI infrastructure has enabled the creation of what was described by the founder Christian Dalingo as a “coincidence of Box” system that constantly improves the performance of the model. “Maitai works as a high agent between customers and their models service providers,” said Dalsanto. “This allows us to define the best model for each request and improve it in a dynamic way, apply evaluation, improvements and flexibility strategies automatically such as retreat.”
The system works by collecting performance data from each reaction, identifying weaknesses, and improving models frequently without the customer’s intervention. “Since Maitai sits in the middle of the inference flow, we collect strong signals that determine the place of weakness models,” Dalsanto explained. “These” soft spots “are assembled and increasingly adjusted to treat specific weaknesses without causing slopes.”
From 81 % to 99 % accuracy: the numbers behind the hacker of the month of artificial intelligence
The results show significant improvements through multiple performance dimensions. It is time to the first distinctive symbol – how quickly artificial intelligence begins – by 73.4 % from 661 mm to 176 milliliters in the ninety percent. The completion times in general decreased 74.6 % from 1,446 milliseconds to 339 milliseconds.
Perhaps most importantly, accuracy improvements followed a clear upward path through four typical repetitions, starting from 81.5 % and up to 99.2 % – a level that exceeds human performance in many customer service scenarios.
“We have seen about 70 %+ people who are calling for our artificial intelligence who cannot distinguish between the difference between a person,” Bodes told Venturebeat. “Cumin, or was, the dead gift that was of artificial intelligence. With a dedicated model that is speaking like a person like a person, and low -end devices, it does not prevent us from completely transit of the socking valley of humans.”
Performance gains are translated directly into business results. “One of our largest customers has seen a 32 % increase in qualified threads compared to a previous version using the previous modern models,” Bodewes note.
350 human agents have been replaced in one month: Communication centers go to artificial intelligence
Improvements with communication centers face increasing pressure to reduce costs while maintaining service quality. Traditional human agents require training, scheduling of coordination, and large general costs that artificial intelligence agents can eliminate.
“The communication centers are really witnessing huge benefits of audio use to replace human agents,” said Bodies. “One of the communication centers with which we work is to replace 350 human agents completely this month. From the perspective of the call center, this is a change for games, because they do not have to manage the schedules of the human support agent, training agents, and match supply and demand.”
Technology shows a special strength in specific use cases. “It really excels in some areas, including the leading performance in the industry in schedule the main dates and qualification specifically, exceeding what can be able to do so,” Bodewees explained. The company has made a partnership with the major companies that deal with customer and cars.
Edge of devices: Why do you make GROQ chips of artificial intelligence
Groq specialized inferences, called Language units (LPUS), providing the basis for devices that make the multi -style models approach. Unlike graphics processors for general purposes that are usually used to infer artificial intelligence, LPU improves the serial nature of language processing.
“The LPU structure has been improved accurately in data movement and an account at an accurate level with high speed and prediction, allowing effective management of multiple small Delta weights (Loras) on a common basic model with no additional access time,” Kantor said.
The infrastructure -based infrastructure also deals with the expansion capacity that limits the spread of artificial intelligence historically. “The beauty of the use of a cloud -based solution like Groqcloud is that Groq deals with a dynamic synchronization and measurement of our customers for any model of artificial intelligence that we offer, including the arrested Lora models.”
For institutions, economic advantages seem great. “The simplicity and efficiency of the design of our system, low energy consumption, and high performance of our devices allow Groq to provide customers at the lowest cost for each symbol without sacrificing performance with the expansion of its scope,” said Kantor.
Spreading artificial intelligence on the same day: How companies exceed months of integration
One of the most persuasive aspects of partnership is the speed of implementation. Unlike traditional artificial intelligence spread that can require months of integration work, Maitai’s approach allows transformations on the same day for companies that already use models for general purposes.
“For companies that are already produced using models for general purposes, we usually transfer them to Maitai on the same day, with zero disruption,” said Dalsanto. “We start collecting immediate data, and within days to a week, we can provide a faster and more reliable form of original preparation.”
This rapid publishing capacity addresses the Foundation’s joint concerns about artificial intelligence projects: the long -term implementation schedules that delay the return on investment. The agent’s class approach means that companies can maintain current API integration while accessing performance improved continuously.
The Foundation of the Foundation AI: Specialized models replace one size-everything
Cooperation indicates a broader shift in the institution’s AI structure, and to stay away from homogeneous models and general purposes towards specialized systems for the task. “We are monitoring the increasing demand from the teams that break their applications to a very smaller and specialized work burden, each of which benefits from individual transformers.”
This trend reflects the challenges of spreading artificial intelligence. Instead of expecting one models to excel across all tasks, companies are increasingly aware of the value of the solutions built for this purpose, which can be improved continuously based on performance data in the real world.
“HotSwapping Multi-Lora allows companies to spread faster and more accurate models for their applications, remove traditional cost and complexity,” explained Dalsanto. “This mainly turns into how the AI is built and published.”
The technical establishment also provides more advanced applications with technology maturity. Groq’s infrastructure can support dozens of specialized models in one way, which may allow institutions to create very specialized Amnesty International experiences across different customer sectors or cases of use.
“HotSwapping Multi-Lora allows low inference and high mass specifically designed for specific tasks,” said Dalsanto. “Our road map gives the priority of more investment in infrastructure, tools and improvement to create a specific precise reasoning of the application as the new standard.”
For the broader artificial intelligence market, the partnership shows that one day technical restrictions can be processed through specialized infrastructure and exact system design. Since more institutions publish artificial intelligence agents, the competitive advantages showed by sound may create new basic forecasts for performance and response to customer interactions.
Success is also achieved by the emerging model of Amnesty International’s infrastructure companies that work together to solve the complex publishing challenges. This cooperative approach may accelerate innovation through the AI sector, as it collects specialized capabilities to provide solutions that exceed what one provider can achieve independently. If this partnership is any indicator, the era of artificial telephone conversations may end faster than anyone expected.
https://venturebeat.com/wp-content/uploads/2025/06/nuneybits_Vector_art_of_diverse_call_center_0a39d368-40f6-464f-9487-a614964bec67.webp?w=1024?w=1200&strip=all
Source link