Josh Sirota: AI models must update frequently for business effectiveness, local hardware enhances data privacy, and proprietary solutions address task inefficiencies | TWIST

2 hours ago 1

Key takeaways

AI models need to be updated frequently to stay effective in business environments.
Proprietary AI solutions can address inefficiencies in current task execution models.
Small language models running on local hardware can enhance data privacy and ownership.
Integrating financial services into platforms like Aragon can empower agents with financial agency.
The evolution of frontier models in sales technology may eliminate the need for vertical sales agents.
Storing data in AI model weights can lead to faster and more cost-effective operations.
Historical data should be instantly accessible within AI models to improve efficiency.
Owning AI intelligence requires training models on proprietary data and retaining the weights.
Competitive pricing in AI token usage can provide strategic advantages in the industry.
The use of cron jobs for data extraction from platforms like Slack can maintain up-to-date knowledge bases.
The shift towards local AI models could redefine data privacy and the role of open-source models.
Financial integration in enterprise software can enhance the capabilities of digital agents.
AI models that store relevant information internally can significantly improve workflow speed.

Guest intro

Josh Sirota is the founder and CEO of Eragon, an AI operating system for work that connects email, Slack, calendar, and financial data into a single agentic layer pre-trained to understand any business. He previously led go-to-market efforts at Salesforce and enterprise software sales at Oracle. After moving to San Francisco in August 2025 with no connections or technical background, he raised a $12M seed round at a $100M valuation less than a year later.

Addressing inefficiencies in current AI models

There is a significant inefficiency in how tasks are executed within companies using current models.

— Josh Sirota
Current AI models often rely on closed-source solutions, which may not be the most efficient approach.
Proprietary AI solutions offer a potential shift towards more efficient task execution.
We just think there’s a huge amount of inefficiency as far as how we’re executing tasks.

— Josh Sirota
Understanding the limitations of current models is crucial for developing more efficient solutions.
The reliance on closed-source models highlights the need for innovation in AI development.
We don’t necessarily think you have to use these frontier kind of closed source models right for everything.

— Josh Sirota
The industry is moving towards proprietary solutions to address these inefficiencies.

The future of AI and local hardware

The future of AI involves creating small language models that can run on local hardware and utilize personal data.

— Josh Sirota
Local AI models could enhance data privacy by keeping personal data on local devices.
This trend may impact how data ownership is perceived and managed.
We’re on a bit of a collision course… small language models can be produced and run on local hardware.

— Josh Sirota
The ability to run AI models locally could redefine the role of open-source models.
Local models may offer more personalized and secure AI solutions.
The shift towards local hardware in AI development is a significant trend.
If they have all your data in it well what becomes the difference between an open claw and all these skills being made.

— Josh Sirota

Challenges in AI model training and updates

Current AI models need frequent updates to remain effective for businesses.
Current AI models need to be able to update themselves frequently based on real-time interactions to be effective for businesses.

— Josh Sirota
Reinforcement learning plays a crucial role in updating AI models.
The challenge lies in developing algorithms with the right reward functions for updates.
What doesn’t exist yet is like to make that vision true is an algorithm that can actually have the right reward functions.

— Josh Sirota
Real-time interaction data is essential for effective AI model updates.
Can this model actually update itself overnight based on the interactions that it has with people.

— Josh Sirota
Businesses need AI models that can adapt quickly to changing environments.

Data extraction and knowledge base creation

The system uses a cron job to regularly extract data from Slack, creating a comprehensive knowledge base.

— Josh Sirota
Regular data extraction ensures that knowledge bases remain up-to-date.
This process involves running a cron job every 15 minutes to gather data.
Basically what we do is we run like a cron job at a frequency of about every 15 minutes.

— Josh Sirota
Creating a knowledge base from Slack data enhances collaborative environments.
The technical process involves reading data from multiple Slack channels.
We create… a recreated knowledge base if you will of our Slack.

— Josh Sirota
Understanding this process is valuable for managing data in collaborative settings.

Financial integration in enterprise platforms

Aragon integrates with financial services to enhance agent capabilities.
Aragon integrates with various financial services to enhance the agency of its agents.

— Josh Sirota
This integration allows agents to have bank accounts, increasing their financial agency.
Financial infrastructure is a key feature of platforms like Aragon.
It also has a financial infrastructure too raj you you partner with slash to allow enterprises to give their agents bank accounts.

— Josh Sirota
Empowering agents with financial tools enhances their capabilities.
Understanding the role of financial services in enterprise platforms is crucial.
This is one of the cases where like the agent sort of has a little bit of financial agency.

— Josh Sirota

The power of frontier models in sales technology

Frontier models in sales technology are becoming increasingly powerful.
The power of frontier models in sales technology is increasing, leading to questions about the future of vertical sales agents.

— Josh Sirota
This trend raises questions about the need for vertical sales agents.
Integrated solutions may replace traditional sales models.
It became pretty clear to me like actually you know from like a venture model everyone wanted to verticalize.

— Josh Sirota
The evolution of these models could reshape sales technology.
Are we going to have such a thing as like a vertical kind of like a sales agent.

— Josh Sirota
Understanding this shift is important for adapting to future sales strategies.

Efficiency gains in AI workflows

AI models can be more cost-effective by storing information in their weights.
Models can be more cost-effective and faster by storing relevant information in their weights instead of retrieving it from external sources.

— Josh Sirota
This approach leads to faster workflow completion.
The workflow will be done much faster right like 10 times faster as well.

— Josh Sirota
Storing data in model weights reduces retrieval time and cost.
It’s way more cost effective on the second piece which is like the speed and performance.

— Josh Sirota
Businesses can optimize costs and performance with this method.
Understanding how AI models process information is crucial for efficiency.

The importance of proprietary data in AI ownership

Owning AI intelligence requires training models on proprietary data.
To truly own your intelligence as an asset, you need to train a model on proprietary data and retain the weights.

— Josh Sirota
Proprietary data is essential for AI model training and ownership.
Our view is basically in order to actually own your intelligence to have an asset that the company owns.

— Josh Sirota
Retaining model weights ensures that the company owns its AI assets.
This principle is fundamental for AI ownership and data management.
The only way to actually do this is to train a model on your proprietary data and then give you the weights right so you own it.

— Josh Sirota
Understanding this concept is vital for companies investing in AI.

Competitive pricing in AI token usage

Eragon offers competitive pricing for AI token usage.
Our pricing model is competitive, charging $5 per million tokens, which is significantly lower than some competitors.

— Josh Sirota
This pricing strategy provides a clear advantage in the industry.
Comparisons with competitors highlight the cost-effectiveness of Eragon’s model.
We charge by tokens we have a $5 cost per million tokens blended.

— Josh Sirota
Understanding the AI token pricing landscape is important for business decisions.
You can comp that with what anthropic charges which is like $15 per million tokens for opus.

— Josh Sirota
Competitive pricing can influence industry dynamics and adoption rates.