Scaled Cognition Proposes a More ‘Reliable’ Approach to AI
Scaled Cognition, an AI lab focused on reliability, said Thursday it has raised $100 million in a Series A funding led by Khosla Ventures.
Scaled Cognition Proposes a More ‘Reliable’ Approach to AI
Scaled Cognition, an AI lab focused on reliability, said Thursday it has raised $100 million in a Series A funding led by Khosla Ventures.
Scaled Cognition, an AI lab focused on reliability, said Thursday it has raised $100 million in a Series A funding led by Khosla Ventures.
Genesys, a provider of cloud-based AI customer experience technology, also invested in the round, which valued the company at $750 million.
Scaled Cognition will use the funds to expand its research team and accelerate enterprise deployments, the company said. The Mountain View, Calif.-based startup was founded by Chief Executive Dan Roth and Chief Technology Officer Dan Klein, a natural language processing researcher and professor of AI at the University of California, Berkeley.
They sold their prior startup, Semantic Machines, to Microsoft in 2018.
“These frontier models that are out there are amazing—they’re intelligent in so many different ways—but they’re sort of like schizophrenic geniuses,” said Roth. “They can create incredible answers, and then you can ask them the same question a second time and get a completely different answer that … might not even be correct.
“We really believe that for these systems to really be useful, you have to be able to trust them. And in order for you to trust them, they have to be provably reliable,” he said.
A single error can have disastrous consequences, Roth said. An automated healthcare agent that processes a routine prescription refill can’t afford to hallucinate so much as a single digit in a prescription number, lest the patient receive an incorrect medication with potentially harmful effects.
Adding more structure
Roth and Klein set out to design an alternative AI architecture that delivers reliably correct results. The result was APT, or Agentic Pretrained Transformer, as their flagship model is called.
Beyond the model, Scaled Cognition has also built a platform for enterprise AI deployment that includes agentic tooling, live agent monitoring and simulation and evaluation frameworks, the company said.
It is targeting customer experience as a first market, but Genesys is already using APT within its Genesys Cloud platform for agentic virtual agent capabilities.
Roth said large language models rely on token-by-token text prediction, optimizing for linguistic plausibility. The model is detached from external reality and lacks an inherent understanding of whether the output is correct. Scaled Cognition addresses AI hallucination with a model that predicts structured objects, such as programs and system queries, in addition to token streams.
The concept is especially difficult to apply to general models, Roth said, but lends itself to narrow-domain enterprise applications.
“We take advantage of the fact that the places that we apply our models, which are in these enterprise settings, have more limited domains. And when you narrow the scope of what you’re asking the model to predict, you can take advantage of these techniques. If you were to try to apply this to everything, these techniques don’t really work,” Roth said.
A router for AI models
Scaled Cognition’s architecture directs different portions of a query to the most appropriate system, depending upon variables such as the need for extreme reliability, according to venture capital investor Vinod Khosla, founding partner of Khosla Ventures.
“It is a separate model for those parts of the system that need real reliability and cannot be subject to hallucination,” he said.
The need for such high levels of reliability can stress the capabilities of AI architectures and humans alike, according to computer scientist and Databricks co-founder Ion Stoica.
Stoica, like Klein, is a professor of computer science at the University of California, Berkeley. While he was briefed on the company’s work, he said he has no commercial, investment or other involvement with the firm.
“While a human can easily verify a few suggested lines of code, verifying hundreds of thousands of AI-generated lines is practically impossible,” Stoica said. “This makes programmatic reliability an absolute necessity for enterprise systems.”
Scaled Cognition, an AI lab focused on reliability, said Thursday it has raised $100 million in a Series A funding led by Khosla Ventures.
Genesys, a provider of cloud-based AI customer experience technology, also invested in the round, which valued the company at $750 million.
Scaled Cognition will use the funds to expand its research team and accelerate enterprise deployments, the company said. The Mountain View, Calif.-based startup was founded by Chief Executive Dan Roth and Chief Technology Officer Dan Klein, a natural language processing researcher and professor of AI at the University of California, Berkeley.
They sold their prior startup, Semantic Machines, to Microsoft in 2018.
“These frontier models that are out there are amazing—they’re intelligent in so many different ways—but they’re sort of like schizophrenic geniuses,” said Roth. “They can create incredible answers, and then you can ask them the same question a second time and get a completely different answer that … might not even be correct.
“We really believe that for these systems to really be useful, you have to be able to trust them. And in order for you to trust them, they have to be provably reliable,” he said.
A single error can have disastrous consequences, Roth said. An automated healthcare agent that processes a routine prescription refill can’t afford to hallucinate so much as a single digit in a prescription number, lest the patient receive an incorrect medication with potentially harmful effects.
Adding more structure
Roth and Klein set out to design an alternative AI architecture that delivers reliably correct results. The result was APT, or Agentic Pretrained Transformer, as their flagship model is called.
Beyond the model, Scaled Cognition has also built a platform for enterprise AI deployment that includes agentic tooling, live agent monitoring and simulation and evaluation frameworks, the company said.
It is targeting customer experience as a first market, but Genesys is already using APT within its Genesys Cloud platform for agentic virtual agent capabilities.
Roth said large language models rely on token-by-token text prediction, optimizing for linguistic plausibility. The model is detached from external reality and lacks an inherent understanding of whether the output is correct. Scaled Cognition addresses AI hallucination with a model that predicts structured objects, such as programs and system queries, in addition to token streams.
The concept is especially difficult to apply to general models, Roth said, but lends itself to narrow-domain enterprise applications.
“We take advantage of the fact that the places that we apply our models, which are in these enterprise settings, have more limited domains. And when you narrow the scope of what you’re asking the model to predict, you can take advantage of these techniques. If you were to try to apply this to everything, these techniques don’t really work,” Roth said.
A router for AI models
Scaled Cognition’s architecture directs different portions of a query to the most appropriate system, depending upon variables such as the need for extreme reliability, according to venture capital investor Vinod Khosla, founding partner of Khosla Ventures.
“It is a separate model for those parts of the system that need real reliability and cannot be subject to hallucination,” he said.
The need for such high levels of reliability can stress the capabilities of AI architectures and humans alike, according to computer scientist and Databricks co-founder Ion Stoica.
Stoica, like Klein, is a professor of computer science at the University of California, Berkeley. While he was briefed on the company’s work, he said he has no commercial, investment or other involvement with the firm.
“While a human can easily verify a few suggested lines of code, verifying hundreds of thousands of AI-generated lines is practically impossible,” Stoica said. “This makes programmatic reliability an absolute necessity for enterprise systems.”
Scaled Cognition, an AI lab focused on reliability, said Thursday it has raised $100 million in a Series A funding led by Khosla Ventures.
Genesys, a provider of cloud-based AI customer experience technology, also invested in the round, which valued the company at $750 million.
Scaled Cognition will use the funds to expand its research team and accelerate enterprise deployments, the company said. The Mountain View, Calif.-based startup was founded by Chief Executive Dan Roth and Chief Technology Officer Dan Klein, a natural language processing researcher and professor of AI at the University of California, Berkeley.
They sold their prior startup, Semantic Machines, to Microsoft in 2018.
“These frontier models that are out there are amazing—they’re intelligent in so many different ways—but they’re sort of like schizophrenic geniuses,” said Roth. “They can create incredible answers, and then you can ask them the same question a second time and get a completely different answer that … might not even be correct.
“We really believe that for these systems to really be useful, you have to be able to trust them. And in order for you to trust them, they have to be provably reliable,” he said.
A single error can have disastrous consequences, Roth said. An automated healthcare agent that processes a routine prescription refill can’t afford to hallucinate so much as a single digit in a prescription number, lest the patient receive an incorrect medication with potentially harmful effects.
Adding more structure
Roth and Klein set out to design an alternative AI architecture that delivers reliably correct results. The result was APT, or Agentic Pretrained Transformer, as their flagship model is called.
Beyond the model, Scaled Cognition has also built a platform for enterprise AI deployment that includes agentic tooling, live agent monitoring and simulation and evaluation frameworks, the company said.
It is targeting customer experience as a first market, but Genesys is already using APT within its Genesys Cloud platform for agentic virtual agent capabilities.
Roth said large language models rely on token-by-token text prediction, optimizing for linguistic plausibility. The model is detached from external reality and lacks an inherent understanding of whether the output is correct. Scaled Cognition addresses AI hallucination with a model that predicts structured objects, such as programs and system queries, in addition to token streams.
The concept is especially difficult to apply to general models, Roth said, but lends itself to narrow-domain enterprise applications.
“We take advantage of the fact that the places that we apply our models, which are in these enterprise settings, have more limited domains. And when you narrow the scope of what you’re asking the model to predict, you can take advantage of these techniques. If you were to try to apply this to everything, these techniques don’t really work,” Roth said.
A router for AI models
Scaled Cognition’s architecture directs different portions of a query to the most appropriate system, depending upon variables such as the need for extreme reliability, according to venture capital investor Vinod Khosla, founding partner of Khosla Ventures.
“It is a separate model for those parts of the system that need real reliability and cannot be subject to hallucination,” he said.
The need for such high levels of reliability can stress the capabilities of AI architectures and humans alike, according to computer scientist and Databricks co-founder Ion Stoica.
Stoica, like Klein, is a professor of computer science at the University of California, Berkeley. While he was briefed on the company’s work, he said he has no commercial, investment or other involvement with the firm.
“While a human can easily verify a few suggested lines of code, verifying hundreds of thousands of AI-generated lines is practically impossible,” Stoica said. “This makes programmatic reliability an absolute necessity for enterprise systems.”
Scaled Cognition, an AI lab focused on reliability, said Thursday it has raised $100 million in a Series A funding led by Khosla Ventures.
Genesys, a provider of cloud-based AI customer experience technology, also invested in the round, which valued the company at $750 million.
Scaled Cognition will use the funds to expand its research team and accelerate enterprise deployments, the company said. The Mountain View, Calif.-based startup was founded by Chief Executive Dan Roth and Chief Technology Officer Dan Klein, a natural language processing researcher and professor of AI at the University of California, Berkeley.
They sold their prior startup, Semantic Machines, to Microsoft in 2018.
“These frontier models that are out there are amazing—they’re intelligent in so many different ways—but they’re sort of like schizophrenic geniuses,” said Roth. “They can create incredible answers, and then you can ask them the same question a second time and get a completely different answer that … might not even be correct.
“We really believe that for these systems to really be useful, you have to be able to trust them. And in order for you to trust them, they have to be provably reliable,” he said.
A single error can have disastrous consequences, Roth said. An automated healthcare agent that processes a routine prescription refill can’t afford to hallucinate so much as a single digit in a prescription number, lest the patient receive an incorrect medication with potentially harmful effects.
Adding more structure
Roth and Klein set out to design an alternative AI architecture that delivers reliably correct results. The result was APT, or Agentic Pretrained Transformer, as their flagship model is called.
Beyond the model, Scaled Cognition has also built a platform for enterprise AI deployment that includes agentic tooling, live agent monitoring and simulation and evaluation frameworks, the company said.
It is targeting customer experience as a first market, but Genesys is already using APT within its Genesys Cloud platform for agentic virtual agent capabilities.
Roth said large language models rely on token-by-token text prediction, optimizing for linguistic plausibility. The model is detached from external reality and lacks an inherent understanding of whether the output is correct. Scaled Cognition addresses AI hallucination with a model that predicts structured objects, such as programs and system queries, in addition to token streams.
The concept is especially difficult to apply to general models, Roth said, but lends itself to narrow-domain enterprise applications.
“We take advantage of the fact that the places that we apply our models, which are in these enterprise settings, have more limited domains. And when you narrow the scope of what you’re asking the model to predict, you can take advantage of these techniques. If you were to try to apply this to everything, these techniques don’t really work,” Roth said.
A router for AI models
Scaled Cognition’s architecture directs different portions of a query to the most appropriate system, depending upon variables such as the need for extreme reliability, according to venture capital investor Vinod Khosla, founding partner of Khosla Ventures.
“It is a separate model for those parts of the system that need real reliability and cannot be subject to hallucination,” he said.
The need for such high levels of reliability can stress the capabilities of AI architectures and humans alike, according to computer scientist and Databricks co-founder Ion Stoica.
Stoica, like Klein, is a professor of computer science at the University of California, Berkeley. While he was briefed on the company’s work, he said he has no commercial, investment or other involvement with the firm.
“While a human can easily verify a few suggested lines of code, verifying hundreds of thousands of AI-generated lines is practically impossible,” Stoica said. “This makes programmatic reliability an absolute necessity for enterprise systems.”
Scaled Cognition, an AI lab focused on reliability, said Thursday it has raised $100 million in a Series A funding led by Khosla Ventures.
Genesys, a provider of cloud-based AI customer experience technology, also invested in the round, which valued the company at $750 million.
Scaled Cognition will use the funds to expand its research team and accelerate enterprise deployments, the company said. The Mountain View, Calif.-based startup was founded by Chief Executive Dan Roth and Chief Technology Officer Dan Klein, a natural language processing researcher and professor of AI at the University of California, Berkeley.
They sold their prior startup, Semantic Machines, to Microsoft in 2018.
“These frontier models that are out there are amazing—they’re intelligent in so many different ways—but they’re sort of like schizophrenic geniuses,” said Roth. “They can create incredible answers, and then you can ask them the same question a second time and get a completely different answer that … might not even be correct.
“We really believe that for these systems to really be useful, you have to be able to trust them. And in order for you to trust them, they have to be provably reliable,” he said.
A single error can have disastrous consequences, Roth said. An automated healthcare agent that processes a routine prescription refill can’t afford to hallucinate so much as a single digit in a prescription number, lest the patient receive an incorrect medication with potentially harmful effects.
Adding more structure
Roth and Klein set out to design an alternative AI architecture that delivers reliably correct results. The result was APT, or Agentic Pretrained Transformer, as their flagship model is called.
Beyond the model, Scaled Cognition has also built a platform for enterprise AI deployment that includes agentic tooling, live agent monitoring and simulation and evaluation frameworks, the company said.
It is targeting customer experience as a first market, but Genesys is already using APT within its Genesys Cloud platform for agentic virtual agent capabilities.
Roth said large language models rely on token-by-token text prediction, optimizing for linguistic plausibility. The model is detached from external reality and lacks an inherent understanding of whether the output is correct. Scaled Cognition addresses AI hallucination with a model that predicts structured objects, such as programs and system queries, in addition to token streams.
The concept is especially difficult to apply to general models, Roth said, but lends itself to narrow-domain enterprise applications.
“We take advantage of the fact that the places that we apply our models, which are in these enterprise settings, have more limited domains. And when you narrow the scope of what you’re asking the model to predict, you can take advantage of these techniques. If you were to try to apply this to everything, these techniques don’t really work,” Roth said.
A router for AI models
Scaled Cognition’s architecture directs different portions of a query to the most appropriate system, depending upon variables such as the need for extreme reliability, according to venture capital investor Vinod Khosla, founding partner of Khosla Ventures.
“It is a separate model for those parts of the system that need real reliability and cannot be subject to hallucination,” he said.
The need for such high levels of reliability can stress the capabilities of AI architectures and humans alike, according to computer scientist and Databricks co-founder Ion Stoica.
Stoica, like Klein, is a professor of computer science at the University of California, Berkeley. While he was briefed on the company’s work, he said he has no commercial, investment or other involvement with the firm.
“While a human can easily verify a few suggested lines of code, verifying hundreds of thousands of AI-generated lines is practically impossible,” Stoica said. “This makes programmatic reliability an absolute necessity for enterprise systems.”
.svg)
.svg){kind=icon}
.jpg)
.png)
.png)
.svg){kind=small}
