L
How We Built It: Writer - Fireside Chat with Harrison Chase & Waseem Alshikh
@ LangChain
03/27/25
Takeaways
HC
Harrison Chase
03/27/25
@ LangChain
I'm Harrison, co-founder and CEO of LangChain, where we build developer tools to simplify application development with large language models (LLMs).
WA
Waseem Alshikh
03/27/25
@ LangChain
I'm Waseem, co-founder and CTO of Writer, a full-stack generative AI platform for enterprises, focusing on sectors like financial services and healthcare.
WA
Waseem Alshikh
03/27/25
@ LangChain
Writer started in 2020, before GPT-2 existed, focusing on statistical models and later transitioning to transformer models after the 2019 paper on transformers was published.
WA
Waseem Alshikh
03/27/25
@ LangChain
Since the release of ChatGPT, we've seen a shift in enterprise interest from how generative AI works to how it can solve specific business problems.
WA
Waseem Alshikh
03/27/25
@ LangChain
Enterprises are now more focused on the value AI can deliver rather than the specific models or tools being used, seeking stable and reliable solutions.
WA
Waseem Alshikh
03/27/25
@ LangChain
We're observing a trend where enterprises prefer integrated solutions over a collection of disparate tools, aiming for transparency and scalability in AI deployment.
WA
Waseem Alshikh
03/27/25
@ LangChain
Current enterprise use cases for AI include optimizing workflows, particularly in healthcare for claim processing, and in financial services for portfolio risk management.
WA
Waseem Alshikh
03/27/25
@ LangChain
The user experience for AI workflows is evolving from chat interfaces to more integrated systems that run in the background, with a focus on monitoring and performance guarantees.
WA
Waseem Alshikh
03/27/25
@ LangChain
Business users are increasingly taking ownership of AI tools, monitoring their performance and ensuring they align with business processes, rather than relying solely on IT departments.
WA
Waseem Alshikh
03/27/25
@ LangChain
Successful AI implementations in enterprises involve collaboration between business users and IT, ensuring that both the technical and operational aspects are addressed effectively.
L
LangChain Cast
03/27/25
@ LangChain
The trend in building applications has shifted from solely engineers to include product managers and subject matter experts, indicating a broader definition of who can be a builder.
L
LangChain Cast
03/27/25
@ LangChain
People prefer to describe their workflows in natural language rather than using prompts, allowing for simple visualizations of their processes.
L
LangChain Cast
03/27/25
@ LangChain
Larger companies tend to have a smaller, more limited tech stack, while smaller companies often utilize a wider variety of tools, with a common reliance on open-source solutions.
L
LangChain Cast
03/27/25
@ LangChain
AI is considered core to many businesses, leading them to prefer open-source models and tools to maintain control over their implementations.
L
LangChain Cast
03/27/25
@ LangChain
Companies are moving away from renting AI solutions and prefer to own and control their models, seeking flexibility and the ability to implement complex systems.
L
LangChain Cast
03/27/25
@ LangChain
The current AI models still struggle with basic tasks, indicating a need for further optimization and development within the industry.
L
LangChain Cast
03/27/25
@ LangChain
There is a significant gap in the market for frameworks that clearly define how to scale and build AI agents, highlighting an opportunity for innovation.
L
LangChain Cast
03/27/25
@ LangChain
Performance issues in graph-based systems often arise when scaling to enterprise levels, affecting both accuracy and system reliability.
L
LangChain Cast
03/27/25
@ LangChain
Customization for different customers is essential, as user behavior varies significantly; some prefer simple keyword searches while others require complex queries.
L
LangChain Cast
03/27/25
@ LangChain
Companies are increasingly focused on the business value of AI agents rather than just technical benchmarks, emphasizing the importance of measurable impact.
WA
Waseem Alshikh
03/27/25
@ LangChain
The biggest factor that excites our customers is the potential for time savings when deploying AI agents, often quantified as a 10-30% reduction in time across teams.
WA
Waseem Alshikh
03/27/25
@ LangChain
In healthcare, we measure the average time to process claims with and without AI, providing clear metrics for our clients.
WA
Waseem Alshikh
03/27/25
@ LangChain
Enterprise clients often have established metrics and performance indicators, allowing us to track how much time is spent on tasks and the impact of AI adoption.
WA
Waseem Alshikh
03/27/25
@ LangChain
We encourage our customers to take ownership of tracking their metrics, as self-measurement leads to a greater realization of AI's value compared to relying solely on our platform's dashboards.
WA
Waseem Alshikh
03/27/25
@ LangChain
We have developed an internal evaluation system to measure our models, as relying on external benchmarks often leads to data contamination and unreliable results.
WA
Waseem Alshikh
03/27/25
@ LangChain
Our focus is on optimizing for reasoning and behavior of models rather than just their scores on traditional benchmarks like MMLU.
WA
Waseem Alshikh
03/27/25
@ LangChain
We are excited about self-evolving models that learn and adapt in real-time without external systems, enhancing their performance based on user interactions.
WA
Waseem Alshikh
03/27/25
@ LangChain
The goal of self-evolving systems is to improve accuracy in multi-step workflows, aiming for 99% accuracy without human intervention.
WA
Waseem Alshikh
03/27/25
@ LangChain
We find that human feedback is often unreliable, so our self-evolving models are designed to learn from their own mistakes without needing human input.
WA
Waseem Alshikh
03/27/25
@ LangChain
If you have an idea, just go start your own company. It's challenging but also the most enjoyable experience you can have.