Overcome Failing Document Ingestion & RAG Strategies with Agentic Knowledge Distillation

Stay Ahead, Stay ONMINE

Overcome Failing Document Ingestion & RAG Strategies with Agentic Knowledge Distillation

Introduction Many generative AI use cases still revolve around Retrieval Augmented Generation (RAG), yet consistently fall short of user expectations. Despite the growing body of research on RAG improvements and even adding Agents into the process, many solutions still fail to return exhaustive results, miss information that is critical but infrequently mentioned in the documents, require multiple search iterations, and generally struggle to reconcile key themes across multiple documents. To top it all off, many implementations still rely on cramming as much “relevant” information as possible into the model’s context window alongside detailed system and user prompts. Reconciling all this information often exceeds the model’s cognitive capacity and compromises response quality and consistency. This is where our Agentic Knowledge Distillation + Pyramid Search Approach comes into play. Instead of chasing the best chunking strategy, retrieval algorithm, or inference-time reasoning method, my team, Jim Brown, Mason Sawtell, Sandi Besen, and I, take an agentic approach to document ingestion. We leverage the full capability of the model at ingestion time to focus exclusively on distilling and preserving the most meaningful information from the document dataset. This fundamentally simplifies the RAG process by allowing the model to direct its reasoning abilities toward addressing the user/system instructions rather than struggling to understand formatting and disparate information across document chunks. We specifically target high-value questions that are often difficult to evaluate because they have multiple correct answers or solution paths. These cases are where traditional RAG solutions struggle most and existing RAG evaluation datasets are largely insufficient for testing this problem space. For our research implementation, we downloaded annual and quarterly reports from the last year for the 30 companies in the DOW Jones Industrial Average. These documents can be found through the SEC EDGAR website. The information on EDGAR is accessible and able to be downloaded for free or can be queried through EDGAR public searches. See the SEC privacy policy for additional details, information on the SEC website is “considered public information and may be copied or further distributed by users of the web site without the SEC’s permission”. We selected this dataset for two key reasons: first, it falls outside the knowledge cutoff for the models evaluated, ensuring that the models cannot respond to questions based on their knowledge from pre-training; second, it’s a close approximation for real-world business problems while allowing us to discuss and share our findings using publicly available data. While typical RAG solutions excel at factual retrieval where the answer is easily identified in the document dataset (e.g., “When did Apple’s annual shareholder’s meeting occur?”), they struggle with nuanced questions that require a deeper understanding of concepts across documents (e.g., “Which of the DOW companies has the most promising AI strategy?”). Our Agentic Knowledge Distillation + Pyramid Search Approach addresses these types of questions with much greater success compared to other standard approaches we tested and overcomes limitations associated with using knowledge graphs in RAG systems. In this article, we’ll cover how our knowledge distillation process works, key benefits of this approach, examples, and an open discussion on the best way to evaluate these types of systems where, in many cases, there is no singular “right” answer. Building the pyramid: How Agentic Knowledge Distillation works Image by author and team depicting pyramid structure for document ingestion. Robots meant to represent agents building the pyramid. Overview Our knowledge distillation process creates a multi-tiered pyramid of information from the raw source documents. Our approach is inspired by the pyramids used in deep learning computer vision-based tasks, which allow a model to analyze an image at multiple scales. We take the contents of the raw document, convert it to markdown, and distill the content into a list of atomic insights, related concepts, document abstracts, and general recollections/memories. During retrieval it’s possible to access any or all levels of the pyramid to respond to the user request. How to distill documents and build the pyramid: Convert documents to Markdown: Convert all raw source documents to Markdown. We’ve found models process markdown best for this task compared to other formats like JSON and it is more token efficient. We used Azure Document Intelligence to generate the markdown for each page of the document, but there are many other open-source libraries like MarkItDown which do the same thing. Our dataset included 331 documents and 16,601 pages. Extract atomic insights from each page: We process documents using a two-page sliding window, which allows each page to be analyzed twice. This gives the agent the opportunity to correct any potential mistakes when processing the page initially. We instruct the model to create a numbered list of insights that grows as it processes the pages in the document. The agent can overwrite insights from the previous page if they were incorrect since it sees each page twice. We instruct the model to extract insights in simple sentences following the subject-verb-object (SVO) format and to write sentences as if English is the second language of the user. This significantly improves performance by encouraging clarity and precision. Rolling over each page multiple times and using the SVO format also solves the disambiguation problem, which is a huge challenge for knowledge graphs. The insight generation step is also particularly helpful for extracting information from tables since the model captures the facts from the table in clear, succinct sentences. Our dataset produced 216,931 total insights, about 13 insights per page and 655 insights per document. Distilling concepts from insights: From the detailed list of insights, we identify higher-level concepts that connect related information about the document. This step significantly reduces noise and redundant information in the document while preserving essential information and themes. Our dataset produced 14,824 total concepts, about 1 concept per page and 45 concepts per document. Creating abstracts from concepts: Given the insights and concepts in the document, the LLM writes an abstract that appears both better than any abstract a human would write and more information-dense than any abstract present in the original document. The LLM generated abstract provides incredibly comprehensive knowledge about the document with a small token density that carries a significant amount of information. We produce one abstract per document, 331 total. Storing recollections/memories across documents: At the top of the pyramid we store critical information that is useful across all tasks. This can be information that the user shares about the task or information the agent learns about the dataset over time by researching and responding to tasks. For example, we can store the current 30 companies in the DOW as a recollection since this list is different from the 30 companies in the DOW at the time of the model’s knowledge cutoff. As we conduct more and more research tasks, we can continuously improve our recollections and maintain an audit trail of which documents these recollections originated from. For example, we can keep track of AI strategies across companies, where companies are making major investments, etc. These high-level connections are super important since they reveal relationships and information that are not apparent in a single page or document. Sample subset of insights extracted from IBM 10Q, Q3 2024 (page 4) We store the text and embeddings for each layer of the pyramid (pages and up) in Azure PostgreSQL. We originally used Azure AI Search, but switched to PostgreSQL for cost reasons. This required us to write our own hybrid search function since PostgreSQL doesn’t yet natively support this feature. This implementation would work with any vector database or vector index of your choosing. The key requirement is to store and efficiently retrieve both text and vector embeddings at any level of the pyramid. This approach essentially creates the essence of a knowledge graph, but stores information in natural language, the way an LLM natively wants to interact with it, and is more efficient on token retrieval. We also let the LLM pick the terms used to categorize each level of the pyramid, this seemed to let the model decide for itself the best way to describe and differentiate between the information stored at each level. For example, the LLM preferred “insights” to “facts” as the label for the first level of distilled knowledge. Our goal in doing this was to better understand how an LLM thinks about the process by letting it decide how to store and group related information. Using the pyramid: How it works with RAG & Agents At inference time, both traditional RAG and agentic approaches benefit from the pre-processed, distilled information ingested in our knowledge pyramid. The pyramid structure allows for efficient retrieval in both the traditional RAG case, where only the top X related pieces of information are retrieved or in the Agentic case, where the Agent iteratively plans, retrieves, and evaluates information before returning a final response. The benefit of the pyramid approach is that information at any and all levels of the pyramid can be used during inference. For our implementation, we used PydanticAI to create a search agent that takes in the user request, generates search terms, explores ideas related to the request, and keeps track of information relevant to the request. Once the search agent determines there’s sufficient information to address the user request, the results are re-ranked and sent back to the LLM to generate a final reply. Our implementation allows a search agent to traverse the information in the pyramid as it gathers details about a concept/search term. This is similar to walking a knowledge graph, but in a way that’s more natural for the LLM since all the information in the pyramid is stored in natural language. Depending on the use case, the Agent could access information at all levels of the pyramid or only at specific levels (e.g. only retrieve information from the concepts). For our experiments, we did not retrieve raw page-level data since we wanted to focus on token efficiency and found the LLM-generated information for the insights, concepts, abstracts, and recollections was sufficient for completing our tasks. In theory, the Agent could also have access to the page data; this would provide additional opportunities for the agent to re-examine the original document text; however, it would also significantly increase the total tokens used. Here is a high-level visualization of our Agentic approach to responding to user requests: Image created by author and team providing an overview of the agentic research & response process Results from the pyramid: Real-world examples To evaluate the effectiveness of our approach, we tested it against a variety of question categories, including typical fact-finding questions and complex cross-document research and analysis tasks. Fact-finding (spear fishing): These tasks require identifying specific information or facts that are buried in a document. These are the types of questions typical RAG solutions target but often require many searches and consume lots of tokens to answer correctly. Example task: “What was IBM’s total revenue in the latest financial reporting?” Example response using pyramid approach: “IBM’s total revenue for the third quarter of 2024 was $14.968 billion [ibm-10q-q3-2024.pdf, pg. 4] Total tokens used to research and generate response This result is correct (human-validated) and was generated using only 9,994 total tokens, with 1,240 tokens in the generated final response. Complex research and analysis: These tasks involve researching and understanding multiple concepts to gain a broader understanding of the documents and make inferences and informed assumptions based on the gathered facts. Example task: “Analyze the investments Microsoft and NVIDIA are making in AI and how they are positioning themselves in the market. The report should be clearly formatted.” Example response: Response generated by the agent analyzing AI investments and positioning for Microsoft and NVIDIA. The result is a comprehensive report that executed quickly and contains detailed information about each of the companies. 26,802 total tokens were used to research and respond to the request with a significant percentage of them used for the final response (2,893 tokens or ~11%). These results were also reviewed by a human to verify their validity. Snippet indicating total token usage for the task Example task: “Create a report on analyzing the risks disclosed by the various financial companies in the DOW. Indicate which risks are shared and unique.” Example response: Part 1 of response generated by the agent on disclosed risks. Part 2 of response generated by the agent on disclosed risks. Similarly, this task was completed in 42.7 seconds and used 31,685 total tokens, with 3,116 tokens used to generate the final report. Snippet indicating total token usage for the task These results for both fact-finding and complex analysis tasks demonstrate that the pyramid approach efficiently creates detailed reports with low latency using a minimal amount of tokens. The tokens used for the tasks carry dense meaning with little noise allowing for high-quality, thorough responses across tasks. Benefits of the pyramid: Why use it? Overall, we found that our pyramid approach provided a significant boost in response quality and overall performance for high-value questions. Some of the key benefits we observed include: Reduced model’s cognitive load: When the agent receives the user task, it retrieves pre-processed, distilled information rather than the raw, inconsistently formatted, disparate document chunks. This fundamentally improves the retrieval process since the model doesn’t waste its cognitive capacity on trying to break down the page/chunk text for the first time. Superior table processing: By breaking down table information and storing it in concise but descriptive sentences, the pyramid approach makes it easier to retrieve relevant information at inference time through natural language queries. This was particularly important for our dataset since financial reports contain lots of critical information in tables. Improved response quality to many types of requests: The pyramid enables more comprehensive context-aware responses to both precise, fact-finding questions and broad analysis based tasks that involve many themes across numerous documents. Preservation of critical context: Since the distillation process identifies and keeps track of key facts, important information that might appear only once in the document is easier to maintain. For example, noting that all tables are represented in millions of dollars or in a particular currency. Traditional chunking methods often cause this type of information to slip through the cracks. Optimized token usage, memory, and speed: By distilling information at ingestion time, we significantly reduce the number of tokens required during inference, are able to maximize the value of information put in the context window, and improve memory use. Scalability: Many solutions struggle to perform as the size of the document dataset grows. This approach provides a much more efficient way to manage a large volume of text by only preserving critical information. This also allows for a more efficient use of the LLMs context window by only sending it useful, clear information. Efficient concept exploration: The pyramid enables the agent to explore related information similar to navigating a knowledge graph, but does not require ever generating or maintaining relationships in the graph. The agent can use natural language exclusively and keep track of important facts related to the concepts it’s exploring in a highly token-efficient and fluid way. Emergent dataset understanding: An unexpected benefit of this approach emerged during our testing. When asking questions like “what can you tell me about this dataset?” or “what types of questions can I ask?”, the system is able to respond and suggest productive search topics because it has a more robust understanding of the dataset context by accessing higher levels in the pyramid like the abstracts and recollections. Beyond the pyramid: Evaluation challenges & future directions Challenges While the results we’ve observed when using the pyramid search approach have been nothing short of amazing, finding ways to establish meaningful metrics to evaluate the entire system both at ingestion time and during information retrieval is challenging. Traditional RAG and Agent evaluation frameworks often fail to address nuanced questions and analytical responses where many different responses are valid. Our team plans to write a research paper on this approach in the future, and we are open to any thoughts and feedback from the community, especially when it comes to evaluation metrics. Many of the existing datasets we found were focused on evaluating RAG use cases within one document or precise information retrieval across multiple documents rather than robust concept and theme analysis across documents and domains. The main use cases we are interested in relate to broader questions that are representative of how businesses actually want to interact with GenAI systems. For example, “tell me everything I need to know about customer X” or “how do the behaviors of Customer A and B differ? Which am I more likely to have a successful meeting with?”. These types of questions require a deep understanding of information across many sources. The answers to these questions typically require a person to synthesize data from multiple areas of the business and think critically about it. As a result, the answers to these questions are rarely written or saved anywhere which makes it impossible to simply store and retrieve them through a vector index in a typical RAG process. Another consideration is that many real-world use cases involve dynamic datasets where documents are consistently being added, edited, and deleted. This makes it difficult to evaluate and track what a “correct” response is since the answer will evolve as the available information changes. Future directions In the future, we believe that the pyramid approach can address some of these challenges by enabling more effective processing of dense documents and storing learned information as recollections. However, tracking and evaluating the validity of the recollections over time will be critical to the system’s overall success and remains a key focus area for our ongoing work. When applying this approach to organizational data, the pyramid process could also be used to identify and assess discrepancies across areas of the business. For example, uploading all of a company’s sales pitch decks could surface where certain products or services are being positioned inconsistently. It could also be used to compare insights extracted from various line of business data to help understand if and where teams have developed conflicting understandings of topics or different priorities. This application goes beyond pure information retrieval use cases and would allow the pyramid to serve as an organizational alignment tool that helps identify divergences in messaging, terminology, and overall communication. Conclusion: Key takeaways and why the pyramid approach matters The knowledge distillation pyramid approach is significant because it leverages the full power of the LLM at both ingestion and retrieval time. Our approach allows you to store dense information in fewer tokens which has the added benefit of reducing noise in the dataset at inference. Our approach also runs very quickly and is incredibly token efficient, we are able to generate responses within seconds, explore potentially hundreds of searches, and on average use

Introduction

Many generative AI use cases still revolve around Retrieval Augmented Generation (RAG), yet consistently fall short of user expectations. Despite the growing body of research on RAG improvements and even adding Agents into the process, many solutions still fail to return exhaustive results, miss information that is critical but infrequently mentioned in the documents, require multiple search iterations, and generally struggle to reconcile key themes across multiple documents. To top it all off, many implementations still rely on cramming as much “relevant” information as possible into the model’s context window alongside detailed system and user prompts. Reconciling all this information often exceeds the model’s cognitive capacity and compromises response quality and consistency.

This is where our Agentic Knowledge Distillation + Pyramid Search Approach comes into play. Instead of chasing the best chunking strategy, retrieval algorithm, or inference-time reasoning method, my team, Jim Brown, Mason Sawtell, Sandi Besen, and I, take an agentic approach to document ingestion.

We leverage the full capability of the model at ingestion time to focus exclusively on distilling and preserving the most meaningful information from the document dataset. This fundamentally simplifies the RAG process by allowing the model to direct its reasoning abilities toward addressing the user/system instructions rather than struggling to understand formatting and disparate information across document chunks.

We specifically target high-value questions that are often difficult to evaluate because they have multiple correct answers or solution paths. These cases are where traditional RAG solutions struggle most and existing RAG evaluation datasets are largely insufficient for testing this problem space. For our research implementation, we downloaded annual and quarterly reports from the last year for the 30 companies in the DOW Jones Industrial Average. These documents can be found through the SEC EDGAR website. The information on EDGAR is accessible and able to be downloaded for free or can be queried through EDGAR public searches. See the SEC privacy policy for additional details, information on the SEC website is “considered public information and may be copied or further distributed by users of the web site without the SEC’s permission”. We selected this dataset for two key reasons: first, it falls outside the knowledge cutoff for the models evaluated, ensuring that the models cannot respond to questions based on their knowledge from pre-training; second, it’s a close approximation for real-world business problems while allowing us to discuss and share our findings using publicly available data.

While typical RAG solutions excel at factual retrieval where the answer is easily identified in the document dataset (e.g., “When did Apple’s annual shareholder’s meeting occur?”), they struggle with nuanced questions that require a deeper understanding of concepts across documents (e.g., “Which of the DOW companies has the most promising AI strategy?”). Our Agentic Knowledge Distillation + Pyramid Search Approach addresses these types of questions with much greater success compared to other standard approaches we tested and overcomes limitations associated with using knowledge graphs in RAG systems.

In this article, we’ll cover how our knowledge distillation process works, key benefits of this approach, examples, and an open discussion on the best way to evaluate these types of systems where, in many cases, there is no singular “right” answer.

Building the pyramid: How Agentic Knowledge Distillation works

AI-generated image showing a pyramid structure for document ingestion with labelled sections. — Image by author and team depicting pyramid structure for document ingestion. Robots meant to represent agents building the pyramid.

Overview

Our knowledge distillation process creates a multi-tiered pyramid of information from the raw source documents. Our approach is inspired by the pyramids used in deep learning computer vision-based tasks, which allow a model to analyze an image at multiple scales. We take the contents of the raw document, convert it to markdown, and distill the content into a list of atomic insights, related concepts, document abstracts, and general recollections/memories. During retrieval it’s possible to access any or all levels of the pyramid to respond to the user request.

How to distill documents and build the pyramid:

Convert documents to Markdown: Convert all raw source documents to Markdown. We’ve found models process markdown best for this task compared to other formats like JSON and it is more token efficient. We used Azure Document Intelligence to generate the markdown for each page of the document, but there are many other open-source libraries like MarkItDown which do the same thing. Our dataset included 331 documents and 16,601 pages.
Extract atomic insights from each page: We process documents using a two-page sliding window, which allows each page to be analyzed twice. This gives the agent the opportunity to correct any potential mistakes when processing the page initially. We instruct the model to create a numbered list of insights that grows as it processes the pages in the document. The agent can overwrite insights from the previous page if they were incorrect since it sees each page twice. We instruct the model to extract insights in simple sentences following the subject-verb-object (SVO) format and to write sentences as if English is the second language of the user. This significantly improves performance by encouraging clarity and precision. Rolling over each page multiple times and using the SVO format also solves the disambiguation problem, which is a huge challenge for knowledge graphs. The insight generation step is also particularly helpful for extracting information from tables since the model captures the facts from the table in clear, succinct sentences. Our dataset produced 216,931 total insights, about 13 insights per page and 655 insights per document.
Distilling concepts from insights: From the detailed list of insights, we identify higher-level concepts that connect related information about the document. This step significantly reduces noise and redundant information in the document while preserving essential information and themes. Our dataset produced 14,824 total concepts, about 1 concept per page and 45 concepts per document.
Creating abstracts from concepts: Given the insights and concepts in the document, the LLM writes an abstract that appears both better than any abstract a human would write and more information-dense than any abstract present in the original document. The LLM generated abstract provides incredibly comprehensive knowledge about the document with a small token density that carries a significant amount of information. We produce one abstract per document, 331 total.
Storing recollections/memories across documents: At the top of the pyramid we store critical information that is useful across all tasks. This can be information that the user shares about the task or information the agent learns about the dataset over time by researching and responding to tasks. For example, we can store the current 30 companies in the DOW as a recollection since this list is different from the 30 companies in the DOW at the time of the model’s knowledge cutoff. As we conduct more and more research tasks, we can continuously improve our recollections and maintain an audit trail of which documents these recollections originated from. For example, we can keep track of AI strategies across companies, where companies are making major investments, etc. These high-level connections are super important since they reveal relationships and information that are not apparent in a single page or document.

Sample subset of insights extracted from IBM 10Q, Q3 2024 (page 4)

We store the text and embeddings for each layer of the pyramid (pages and up) in Azure PostgreSQL. We originally used Azure AI Search, but switched to PostgreSQL for cost reasons. This required us to write our own hybrid search function since PostgreSQL doesn’t yet natively support this feature. This implementation would work with any vector database or vector index of your choosing. The key requirement is to store and efficiently retrieve both text and vector embeddings at any level of the pyramid.

This approach essentially creates the essence of a knowledge graph, but stores information in natural language, the way an LLM natively wants to interact with it, and is more efficient on token retrieval. We also let the LLM pick the terms used to categorize each level of the pyramid, this seemed to let the model decide for itself the best way to describe and differentiate between the information stored at each level. For example, the LLM preferred “insights” to “facts” as the label for the first level of distilled knowledge. Our goal in doing this was to better understand how an LLM thinks about the process by letting it decide how to store and group related information.

Using the pyramid: How it works with RAG & Agents

At inference time, both traditional RAG and agentic approaches benefit from the pre-processed, distilled information ingested in our knowledge pyramid. The pyramid structure allows for efficient retrieval in both the traditional RAG case, where only the top X related pieces of information are retrieved or in the Agentic case, where the Agent iteratively plans, retrieves, and evaluates information before returning a final response.

The benefit of the pyramid approach is that information at any and all levels of the pyramid can be used during inference. For our implementation, we used PydanticAI to create a search agent that takes in the user request, generates search terms, explores ideas related to the request, and keeps track of information relevant to the request. Once the search agent determines there’s sufficient information to address the user request, the results are re-ranked and sent back to the LLM to generate a final reply. Our implementation allows a search agent to traverse the information in the pyramid as it gathers details about a concept/search term. This is similar to walking a knowledge graph, but in a way that’s more natural for the LLM since all the information in the pyramid is stored in natural language.

Depending on the use case, the Agent could access information at all levels of the pyramid or only at specific levels (e.g. only retrieve information from the concepts). For our experiments, we did not retrieve raw page-level data since we wanted to focus on token efficiency and found the LLM-generated information for the insights, concepts, abstracts, and recollections was sufficient for completing our tasks. In theory, the Agent could also have access to the page data; this would provide additional opportunities for the agent to re-examine the original document text; however, it would also significantly increase the total tokens used.

Here is a high-level visualization of our Agentic approach to responding to user requests:

Overview of the agentic research & response process — Image created by author and team providing an overview of the agentic research & response process

Results from the pyramid: Real-world examples

To evaluate the effectiveness of our approach, we tested it against a variety of question categories, including typical fact-finding questions and complex cross-document research and analysis tasks.

Fact-finding (spear fishing):

These tasks require identifying specific information or facts that are buried in a document. These are the types of questions typical RAG solutions target but often require many searches and consume lots of tokens to answer correctly.

Example task: “What was IBM’s total revenue in the latest financial reporting?”

Example response using pyramid approach: “IBM’s total revenue for the third quarter of 2024 was $14.968 billion [ibm-10q-q3-2024.pdf, pg. 4]

Screenshot of total tokens used to research and generate response — Total tokens used to research and generate response

This result is correct (human-validated) and was generated using only 9,994 total tokens, with 1,240 tokens in the generated final response.

Complex research and analysis:

These tasks involve researching and understanding multiple concepts to gain a broader understanding of the documents and make inferences and informed assumptions based on the gathered facts.

Example task: “Analyze the investments Microsoft and NVIDIA are making in AI and how they are positioning themselves in the market. The report should be clearly formatted.”

Example response:

Screenshot of the response generated by the agent analyzing AI investments and positioning for Microsoft and NVIDIA. — Response generated by the agent analyzing AI investments and positioning for Microsoft and NVIDIA.

The result is a comprehensive report that executed quickly and contains detailed information about each of the companies. 26,802 total tokens were used to research and respond to the request with a significant percentage of them used for the final response (2,893 tokens or ~11%). These results were also reviewed by a human to verify their validity.

Screenshot of snippet indicating total token usage for the task — Snippet indicating total token usage for the task

Example task: “Create a report on analyzing the risks disclosed by the various financial companies in the DOW. Indicate which risks are shared and unique.”

Example response:

Screenshot of part 1 of a response generated by the agent on disclosed risks. — Part 1 of response generated by the agent on disclosed risks.

Screenshot of part 2 of a response generated by the agent on disclosed risks. — Part 2 of response generated by the agent on disclosed risks.

Similarly, this task was completed in 42.7 seconds and used 31,685 total tokens, with 3,116 tokens used to generate the final report.

Screenshot of a snippet indicating total token usage for the task — Snippet indicating total token usage for the task

These results for both fact-finding and complex analysis tasks demonstrate that the pyramid approach efficiently creates detailed reports with low latency using a minimal amount of tokens. The tokens used for the tasks carry dense meaning with little noise allowing for high-quality, thorough responses across tasks.

Benefits of the pyramid: Why use it?

Overall, we found that our pyramid approach provided a significant boost in response quality and overall performance for high-value questions.

Some of the key benefits we observed include:

Reduced model’s cognitive load: When the agent receives the user task, it retrieves pre-processed, distilled information rather than the raw, inconsistently formatted, disparate document chunks. This fundamentally improves the retrieval process since the model doesn’t waste its cognitive capacity on trying to break down the page/chunk text for the first time.
Superior table processing: By breaking down table information and storing it in concise but descriptive sentences, the pyramid approach makes it easier to retrieve relevant information at inference time through natural language queries. This was particularly important for our dataset since financial reports contain lots of critical information in tables.
Improved response quality to many types of requests: The pyramid enables more comprehensive context-aware responses to both precise, fact-finding questions and broad analysis based tasks that involve many themes across numerous documents.
Preservation of critical context: Since the distillation process identifies and keeps track of key facts, important information that might appear only once in the document is easier to maintain. For example, noting that all tables are represented in millions of dollars or in a particular currency. Traditional chunking methods often cause this type of information to slip through the cracks.
Optimized token usage, memory, and speed: By distilling information at ingestion time, we significantly reduce the number of tokens required during inference, are able to maximize the value of information put in the context window, and improve memory use.
Scalability: Many solutions struggle to perform as the size of the document dataset grows. This approach provides a much more efficient way to manage a large volume of text by only preserving critical information. This also allows for a more efficient use of the LLMs context window by only sending it useful, clear information.
Efficient concept exploration: The pyramid enables the agent to explore related information similar to navigating a knowledge graph, but does not require ever generating or maintaining relationships in the graph. The agent can use natural language exclusively and keep track of important facts related to the concepts it’s exploring in a highly token-efficient and fluid way.
Emergent dataset understanding: An unexpected benefit of this approach emerged during our testing. When asking questions like “what can you tell me about this dataset?” or “what types of questions can I ask?”, the system is able to respond and suggest productive search topics because it has a more robust understanding of the dataset context by accessing higher levels in the pyramid like the abstracts and recollections.

Beyond the pyramid: Evaluation challenges & future directions

Challenges

While the results we’ve observed when using the pyramid search approach have been nothing short of amazing, finding ways to establish meaningful metrics to evaluate the entire system both at ingestion time and during information retrieval is challenging. Traditional RAG and Agent evaluation frameworks often fail to address nuanced questions and analytical responses where many different responses are valid.

Our team plans to write a research paper on this approach in the future, and we are open to any thoughts and feedback from the community, especially when it comes to evaluation metrics. Many of the existing datasets we found were focused on evaluating RAG use cases within one document or precise information retrieval across multiple documents rather than robust concept and theme analysis across documents and domains.

The main use cases we are interested in relate to broader questions that are representative of how businesses actually want to interact with GenAI systems. For example, “tell me everything I need to know about customer X” or “how do the behaviors of Customer A and B differ? Which am I more likely to have a successful meeting with?”. These types of questions require a deep understanding of information across many sources. The answers to these questions typically require a person to synthesize data from multiple areas of the business and think critically about it. As a result, the answers to these questions are rarely written or saved anywhere which makes it impossible to simply store and retrieve them through a vector index in a typical RAG process.

Another consideration is that many real-world use cases involve dynamic datasets where documents are consistently being added, edited, and deleted. This makes it difficult to evaluate and track what a “correct” response is since the answer will evolve as the available information changes.

Future directions

In the future, we believe that the pyramid approach can address some of these challenges by enabling more effective processing of dense documents and storing learned information as recollections. However, tracking and evaluating the validity of the recollections over time will be critical to the system’s overall success and remains a key focus area for our ongoing work.

When applying this approach to organizational data, the pyramid process could also be used to identify and assess discrepancies across areas of the business. For example, uploading all of a company’s sales pitch decks could surface where certain products or services are being positioned inconsistently. It could also be used to compare insights extracted from various line of business data to help understand if and where teams have developed conflicting understandings of topics or different priorities. This application goes beyond pure information retrieval use cases and would allow the pyramid to serve as an organizational alignment tool that helps identify divergences in messaging, terminology, and overall communication.

Conclusion: Key takeaways and why the pyramid approach matters

The knowledge distillation pyramid approach is significant because it leverages the full power of the LLM at both ingestion and retrieval time. Our approach allows you to store dense information in fewer tokens which has the added benefit of reducing noise in the dataset at inference. Our approach also runs very quickly and is incredibly token efficient, we are able to generate responses within seconds, explore potentially hundreds of searches, and on average use (this includes all the search iterations!).

We find that the LLM is much better at writing atomic insights as sentences and that these insights effectively distill information from both text-based and tabular data. This distilled information written in natural language is very easy for the LLM to understand and navigate at inference since it does not have to expend unnecessary energy reasoning about and breaking down document formatting or filtering through noise.

The ability to retrieve and aggregate information at any level of the pyramid also provides significant flexibility to address a variety of query types. This approach offers promising performance for large datasets and enables high-value use cases that require nuanced information retrieval and analysis.

Note: The opinions expressed in this article are solely my own and do not necessarily reflect the views or policies of my employer.

Interested in discussing further or collaborating? Reach out on LinkedIn!

Stay Ahead

Explore More Insights

Stay ahead with more perspectives on cutting-edge power, infrastructure, energy, bitcoin and AI solutions. Explore these articles to uncover strategies and insights shaping the future of industries.

Practical quantum computers are over a decade away, says NEC

A practical, commercial quantum computer is over a decade away, executives at Japanese IT services company NEC are reported as saying. That’s why, according to Japanese news publication The Mainichi, company has pulled the plug on its plans to develop a quantum computer — although it will still continue research

Huawei aims to deliver faster AI chips, faster

Huawei is accelerating its AI chips development, bringing forward the release of the next two models in the family powering its AI computing clusters by three to nine months. Its Ascend 960 chip family is a major component of supercomputing portfolio. It now plans to release the Ascend 960DT in

The amount of e-waste caused by AI is underestimated: we can’t only include the servers

“IT leaders buying AI infrastructure should treat lifecycle and end-of-life impacts as an architectural and procurement requirement, tracking and reporting not only compute equipment, but also supporting systems, batteries, refrigerants, suppression agents, reuse potential, and responsible recovery or disposal,” he said. Luthra said that the fix for this problem is

Cisco patches max-severity ISE flaw, the second critical zero-day this week

Cisco released patches for an actively exploited authentication bypass vulnerability in its Cisco Identity Services Engine (ISE) platform, which is used for enterprise network access control and policy enforcement. This is the second zero-day flaw Cisco has been forced to release emergency patches for this week, after fixing a critical

Trump Administration Moves to Keep Indiana Coal Plants Operating to Support Grid Reliability

WASHINGTON—U.S. Secretary of Energy Chris Wright issued emergency orders to keep two Indiana coal plants operational to ensure Americans in the Midwest region of the United States have continued access to affordable, reliable, and secure electricity. The orders direct the Northern Indiana Public Service Company (NIPSCO), CenterPoint Energy, and the Midcontinent Independent System Operator, Inc. (MISO) to take all measures necessary to ensure specified generation units at both the R.M. Schahfer and F.B. Culley generating stations in Indiana are available to operate. Certain generation units at these coal plants were scheduled to shut down at the end of 2025. The orders will minimize the risk of unnecessary blackouts for the American people. Since the U.S. Department of Energy’s (DOE) original orders were issued on December 23, 2025, the Schahfer and Culley coal plants have proven critical to MISO’s operations, operating during periods of high energy demand and low levels of intermittent energy production, including during Winter Storm Fern. “Forcing reliable, dispatchable coal generation off the grid would compromise energy reliability and needlessly raises energy costs for Americans,” said Energy Secretary Wright. “Midwestern families should not be forced to pay the price for the misguided energy subtraction policies of the past. They deserve affordable, reliable, and secure energy, regardless of the wind blowing or the sun shining.” Thanks to President Trump’s leadership, coal generating plants across the country are being saved from premature retirement. For example, in 2025, more than 17 gigawatts of coal power electricity generation were saved from going offline. The availability of R.M. Schahfer and F.B. Culley generating stations to operate will continue to be an asset to maintain reliability in the MISO region and is necessary to address elevated reliability risks in that region during extreme weather and reduce the risk of power outages that could threaten public health and safety. As

Energy Secretary Secures Carolinas’ Grid Amidst Hot Weather Conditions

WASHINGTON—The U.S. Department of Energy (DOE) issued an emergency order to mitigate the risk of blackouts in the Carolinas amid hot weather conditions. Issued pursuant to Section 202(c) of the Federal Power Act, the order authorizes Duke Energy Carolinas, LLC (Duke) to dispatch specified resources and to order their operation as needed to maintain reliability. The order also authorizes Duke, in collaboration with its Transmission Owners, to direct backup generation resources to operate as a last resort before declaring an Energy Emergency Alert (EEA) 3 or during an EEA 3. This order was issued pursuant to an application from Duke submitted on September 18, 2026. “Today’s order will help secure reliable electricity access for millions of American families and businesses across North and South Carolina by making additional power generation, including backup power, available to use as needed,” said U.S. Secretary of Energy Chris Wright. “It should come as no surprise that during the end of summer and early fall, there are fewer hours of daylight—and therefore, less power generation from solar power. The North American Electric Reliability Corporation and others have warned of the potential dangers late summer temperature spikes can pose to the grid when leaders prematurely retire reliable power sources. While past leaders’ energy subtraction policies have made the grid more vulnerable to blackouts when the sun doesn’t shine or the wind doesn’t blow, this administration remains committed to using every available tool to prevent blackouts.” DOE estimates more than 35 GW of unused backup generation remains available nationwide. The order is in effect upon issuance on September 18, 2026, through September 21, 2026.

Enbridge launches open season for West Texas Express natural gas pipeline

Enbridge has launched a non-binding open season for its proposed 2-bcfd West Texas Express (WTX) natural gas pipeline project, designed to transport Permian basin supply west from the Waha area to markets in and around El Paso, Tex. The proposed project responds to growing demand for reliable natural gas supplies from proposed power generation, utilities, generators, and industrial customers such as data centers across west Texas and downstream markets in Mexico, New Mexico, and Arizona, Enbridge said. WTX is currently expected to include more than 150 miles of new 42-in. OD pipeline. The project could also include laterals serving Hudspeth County, Tex., and delivery points at the US-Mexico border. Enbridge said WTX can be designed to connect with existing pipeline infrastructure based on customer requirements identified through the open season. Final capacity, routing, receipt and delivery points, and system design will be informed by market interest. Subject to securing sufficient commercial support and obtaining required approvals, Enbridge is targeting a fourth-quarter 2029 in-service date. The open season will close at 5 p.m. CDT, Sept. 25, 2026. Enbridge last week agreed to acquire Tallgrass Energy LP’s crude oil business for $2.55 billion in cash.

Editorial: Let’s make a deal

The Trump administration announced Aug. 28, 2026, that the US and Venezuela had agreed to give the US majority control over development of more than 65 billion bbl of Venezuelan proven oil reserves. The agreement presents an extraordinary opportunity to the US oil industry, but also a great deal of risk, at least some of which should seem familiar. Before private capital follows Washington into Venezuela, the industry needs answers to some fundamental questions about the deal’s legal durability, political risk, commercial structure, and ultimate purpose. The agreement covers 17 fields and roughly 20% of Venezuela’s proved reserves. Development would be led by Barbados-based North American Blue Energy Partners (NABEP)—controlled by Venezuelan businessman Alejandro Betancourt López—under what the White House described as a 100-year concession. It’s a huge deal. But its timeline alone stretches credulity. A typical international concession agreement would last for 20-30 years, a term consistent with both in-country media reports and outside analysis. As noted by the Center for Strategic & International Studies, Venezuela’s Organic Hydrocarbon Law, passed in January 2026 after Nicolás Maduro’s ouster, only allows “production participation contracts” to private companies, not concessions of any duration.1 Venezuela’s constitution also creates questions about the agreement. Article 150 requires National Assembly approval of “public interest” contracts to entities based outside Venezuela while Article 302 reserves the petroleum industry to the State. Beyond the deal itself Looking beyond legal and structural technicalities, large questions remain regarding both stable governance in Venezuela and the viability of any agreements struck in its absence. There has been no meaningful progress toward establishing a functional democracy in Venezuela since the US captured Maduro. Both Acting President (and former VP) Delcy Rodríguez and Betancourt owe much of their political and personal fortunes to Maduro and his predecessor, Hugo Chávez. Rodríguez has done a

US sanctions bill targets Russian energy but gives Trump broad discretion

US President Donald Trump is poised to sign legislation aimed at increasing economic pressure on Russia over its war in Ukraine by targeting Russian oil and gas revenues and countries that continue to buy Russian energy. While the measure mandates broad sanctions, it gives Trump wide discretion over implementation, including which countries face tariffs, the tariff rates imposed, and whether sanctions provisions are waived. The Lindsey O. Graham Sanctioning Russia and Iran Act of 2026, named for the late South Carolina senator who championed the legislation, passed the House Sept. 16 by a vote of 262-159 after clearing the Senate 86-11 in August. The measure now awaits Trump’s signature. The White House has said the administration supports the legislation and would recommend that Trump sign it into law. The legislation directs the president to impose broad sanctions and tariff measures targeting Russian energy exports and countries that facilitate sanctions evasion. However, Trump “may waive the application” of sanctions provisions, restrictions, or duties if he certifies to Congress that doing so is “in the national interest of the United States” and explains the basis for the decision. While the law mandates sanctions, it leaves key implementation decisions to the administration. Tariff provisions Within 30 days of enactment, the act requires the president to impose duties of up to 100% on goods imported from countries that fall within specified categories involving Russian oil and gas purchases or sanctions evasion. The covered countries include those among the five largest importers of Russian-origin crude oil or natural gas by total volume during the 12 months preceding enactment, as well as countries that meet separate criteria for facilitating Russian sanctions evasion. The administration must reassess those countries every 180 days. A country is exempt from the gas-related duties if its Russian gas imports accounted for

California Resources unloads Uinta assets

The leaders of California Resources Corp., Long Beach, have sold the company’s Uinta basin assets for about $90 million to an undisclosed buyer. The deal has an effective date of July 1 and is expected to close by yearend. “Today’s transaction strengthens our business,” said Francisco Leon, CRC president and chief executive officer. “This transaction enhances our capital allocation flexibility, allowing us to invest in higher-return opportunities within the Golden State, and supports our shareholder return strategy.” CRC had come to own the Uinta assets, which span about 100,000 net acres, after it acquired Berry Corp. in December of last year for $709 million. But the operation accounts for a small part of CRC’s business–2.5% of oil production and 8% of natural gas production in the second quarter–and Leon last month told analysts “it’s hard to see allocating a lot of dollars back into the Uinta” as his team focuses on building out its California network of assets. “It requires a pretty significant amount of capital to develop the scale that we need for a second asset,” Leon said Aug. 10 after CRC reported its second-quarter results. “So as we do a side-by-side and we compare the Uinta assets with California, Uinta has higher capital intensity, higher break-evens, lower crude quality [and] higher transportation and operating costs and steeper declines.” In the deal announcement, Leon said the Uinta sale also offsets the price CRC will pay for a set of midstream assets in California it plans to buy from CorEnergy Infrastructure Trust. The purchase of those pipelines and other operations is expected to close later this month. Shares of CRC (Ticker: CRC) were down slightly to $54.24 in late-morning trading Sept. 17. They have lost about 15% of their value over the past 6 months, trimming the company’s market capitalization

Local AI is getting small enough to make every app multilingual

On-device translation used to mean a separate model for every language you wanted to support. English to French, English to German, and so on. However, that becomes unsustainable at a global scale when you’re talking about thousands of possible language pairs. Add to that the fact that most developers have to either send translation requests to the cloud to get fast, accurate results, or keep it local with restricted language support. Tether’s AI Research team has developed a family of multilingual translation models, TranslatePsy-EuroNano, that each support nine European languages, with deployment built around a pair of multilingual models rather than separate bilingual models for every language pair. What makes this possible Supporting a full European market on-device has previously meant bundling dozens of separate model files, but this is impractical for mobile apps and those building them. Tether AI’s multilingual open‑source edge translation models set the standard for efficiency, quality, and speed. For developers, the possibilities are endless. Using English as a pivot, the models remain comparable to Mozilla Firefox’s Bergamot-based translation system while dramatically reducing the size of on-device translation. At its smallest tier, Tether’s deployment is 17.6 times smaller while maintaining comparable translation quality. Tether’s deployment takes up 36MB to 89MB, depending on the tier you use. By comparison, the equivalent Firefox setup requires 18 separate bilingual models totaling 633MB to provide the same language coverage. The models are small enough to run efficiently on edge devices while supporting nine European languages from a single multilingual deployment, making multilingual experiences practical for a much wider range of software. Potential applications include travel and navigation apps, educational platforms that present lessons and resources on-device. The models are also designed for academics and researchers. Because the weights are openly available, researchers can fine-tune them for specialized domains, like customer

AI Infrastructure Is Redrawing the Data Center Services Landscape

For gigawatt-scale AI developments, the developer may be involved with substations, transmission interconnections, generation plants, batteries or other behind-the-meter infrastructure long before servers arrive. Solaris now describes its overall portfolio as including generation, distribution, installation and commissioning, aftermarket support, and operations and maintenance. The arrival of companies with roots in energy and heavy industrial services suggests that the data center supplier base itself is changing as projects begin to resemble large industrial infrastructure developments. The Pattern Extends Across the Services Stack The transactions involving T5, Limbach, JK Technology Services and Solaris are hardly isolated. A wider wave of acquisitions and partnerships is pushing equipment manufacturers, contractors, engineering firms and specialist service providers toward broader roles across the data center lifecycle. Vertiv provided perhaps the clearest parallel in September, announcing an agreement to acquire UtilityInnovation Group for approximately $1.45 billion in cash, with additional consideration tied to performance. UIG brings microgrid controls, onsite-generation orchestration, specialized switchgear and behind-the-meter power architecture. The deal also extends a broader 2026 acquisition push by Vertiv that has added liquid-cooling specialist Strategic Thermal Labs, chiller manufacturer ThermoKey and prefabricated infrastructure provider Bmarko as the company builds out more of the AI data center infrastructure stack. Vertiv described the move as extending its portfolio upstream from the critical power and cooling systems inside the facility toward the grid interconnection and onsite generation itself — effectively creating a path from power source to chip. Days later, Flex announced a $4.4 billion agreement to acquire EPC Power, adding grid-forming and power-conversion technology designed for data centers, utility-scale energy storage and microgrids. EPC Power’s platform includes rectifiers and DC-DC conversion for emerging 800-volt data center architectures, with solid-state transformer development also planned. The company says it has more than 15 GW deployed across 62 countries and expects its annual U.S.

From Announcements to Delivery: What Separates Real AI Data Center Projects From the Rest

The AI infrastructure market has become very good at announcing gigawatts. Delivering them is another matter. That distinction framed one of the closing sessions of Day 1 at the Data Center Frontier Trends Summit 2026 (Aug. 4-6), where Sean Farney, vice president of data center strategy at JLL and a member of the Data Center Frontier Editorial Advisory Board, moderated a discussion on why some AI data center projects advance from concept to construction while others remain little more than ambitious site plans. Farney was joined by Lawrence Vo, vice president of M&A and capex at Csquare; John Day, chief commercial officer at CleanArc Data Centers; Justin Loth, executive director of power development at Provident Data Centers; and Roshan Shah, co-founder and CEO of Decimal Digital. The question Farney put before the group was straightforward: amid a market moving at what he called “the speed of light,” what separates the developers that actually get projects done from those that do not? The answers repeatedly came back to the same point. In the current market, land, capital and an announcement are no longer enough. Developers have to prove that power is deliverable, infrastructure is ready, regulatory processes are moving, communities are receptive, talent is available and the commercial model can withstand changing conditions. A Gigawatt on Paper Is Not a Gigawatt of Capacity For Loth, who spent roughly 15 years on the utility side before joining Provident, the scale of current data center proposals alone should force the industry to think differently about what constitutes a credible project. Before the hyperscale and AI expansion, he noted, gigawatts were a measure more commonly associated with cities than individual loads. “A 3.5 gigawatt campus,” Loth said, is roughly equivalent to the native load of Austin or San Antonio. That scale makes the distinction

The Future of Data Centers: Biomimicry and Community-Centric Design

As a result, Microsoft has said six additional data centers planned in the region are being designed around biomimicry principles rather than treating landscaping as something added after the engineering work is finished. The change, from landscaping as decoration to ecology as a design input, is now being applied elsewhere. There is already a significant US example, set in Mecklenburg County, Virginia, where Microsoft originally announced the Chase City Conservancy in 2022, as part of a data center development south of Chase City. The completed project, which opened in April 2025, protects more than 230 acres from development. It includes more than eight acres of wetlands, over 16,300 linear feet of restored streams, 185 acres of native pollinator habitat, more than 25,000 planted trees and over three miles of publicly accessible walking trails. Local environmental organizations helped shift the design away from what the company describes as a more conventional recreational area toward biodiversity and habitat conservation illustrating the community-engagement side of Microsoft’s model, which, given the current temperature of such relationships, can’t be understated. For data center developers, that may be as important as the ecological results. Community impact is no longer being evaluated on just tax revenue and jobs. Turning portions of a site into protected wetlands, forests, trails or habitat potentially creates a visible local benefit in ways that renewable-energy contracts hundreds of miles away cannot. Microsoft’s commitment to the local community has been led by their Community First AI Infrastructure Plan announced in January 2026. Wetlands in Wisconsin, Screening in Georgia At Microsoft’s massive Mount Pleasant, Wisconsin, AI data center development, the company is working with the Root-Pike Watershed Initiative Network on restoration projects involving wetlands, native prairie and forested riparian buffers. One element involves returning previously straightened streams to more natural, winding channels, improving aquatic

Axelera Europa targets enterprise data centers with far more efficient AI

Software is still the gatekeeper Axelera In terms of software enablement, Axelera’s Voyager SDK spans its existing Metis products and the new Europa architecture, providing a common environment across embedded, edge and server deployments, with support for a multitude of computer vision models, LLMs, VLMs, diffusion models, speech and other AI workloads. To automate setup, Axelera’s Voyager Wingman uses natural-language prompts to help developers create or port inference pipelines, while AxeleraScript, or AxScript, provides a Python-enabled domain-specific language with lower-level AIPU control for custom operators and transformer models. This could prove every bit as important as Europa’s performance and efficiency. Enterprises already have models, development environments and application stacks. Extensive rewriting or specialized expertise adds development and operational costs that can quickly undermine savings on hardware and power.

Scott Bergs, CEO of Kirkwood IG: Fiber and the AI Data Center Buildout

For years, fiber was one of the more forgiving elements of data center site selection. Developers could secure land, line up power, begin planning the facility and then work with carriers to establish the connectivity required by tenants. In a traditional multi-tenant data center, that model generally worked. At AI scale, Scott Bergs says it increasingly does not. “The architecture of those original communications service provider networks just don’t meet the latency and/or capacity needs” of today’s high-density compute environments, said Bergs, CEO of Kirkwood Infrastructure Group, during a recent episode of the Data Center Frontier Show. The result is a significant change in the data center development stack: network infrastructure can no longer be treated as something that gets solved after the site is chosen. For hyperscalers and neo-cloud providers, fiber route diversity, latency, physical security and future capacity increasingly need to enter the conversation alongside power and land. And as data center campuses follow available power farther from established digital infrastructure hubs, the scale of the network challenge is expanding with them. A connection between data center campuses that might once have extended two or 30 miles can now stretch 250 miles or more, Bergs said. What would traditionally have been considered a long-haul fiber route is increasingly becoming another piece of inter-campus infrastructure. That change is helping drive Kirkwood’s own expansion. From DF&I to Kirkwood Bergs previously led DF&I, a dark-fiber infrastructure platform concentrated in Northern Virginia and Maryland. Kirkwood Infrastructure Group is not simply DF&I under a new name, he said. Rather, it represents what Bergs described as a second phase in a broader infrastructure investment strategy developed originally through IPI Partners. IPI, an investment platform focused on digital infrastructure, backed DF&I after identifying communications infrastructure serving dense compute environments as an area requiring greater direct

Microsoft will invest $80B in AI data centers in fiscal 2025

And Microsoft isn’t the only one that is ramping up its investments into AI-enabled data centers. Rival cloud service providers are all investing in either upgrading or opening new data centers to capture a larger chunk of business from developers and users of large language models (LLMs). In a report published in October 2024, Bloomberg Intelligence estimated that demand for generative AI would push Microsoft, AWS, Google, Oracle, Meta, and Apple would between them devote $200 billion to capex in 2025, up from $110 billion in 2023. Microsoft is one of the biggest spenders, followed closely by Google and AWS, Bloomberg Intelligence said. Its estimate of Microsoft’s capital spending on AI, at $62.4 billion for calendar 2025, is lower than Smith’s claim that the company will invest $80 billion in the fiscal year to June 30, 2025. Both figures, though, are way higher than Microsoft’s 2020 capital expenditure of “just” $17.6 billion. The majority of the increased spending is tied to cloud services and the expansion of AI infrastructure needed to provide compute capacity for OpenAI workloads. Separately, last October Amazon CEO Andy Jassy said his company planned total capex spend of $75 billion in 2024 and even more in 2025, with much of it going to AWS, its cloud computing division.

John Deere unveils more autonomous farm machines to address skill labor shortage

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More Self-driving tractors might be the path to self-driving cars. John Deere has revealed a new line of autonomous machines and tech across agriculture, construction and commercial landscaping. The Moline, Illinois-based John Deere has been in business for 187 years, yet it’s been a regular as a non-tech company showing off technology at the big tech trade show in Las Vegas and is back at CES 2025 with more autonomous tractors and other vehicles. This is not something we usually cover, but John Deere has a lot of data that is interesting in the big picture of tech. The message from the company is that there aren’t enough skilled farm laborers to do the work that its customers need. It’s been a challenge for most of the last two decades, said Jahmy Hindman, CTO at John Deere, in a briefing. Much of the tech will come this fall and after that. He noted that the average farmer in the U.S. is over 58 and works 12 to 18 hours a day to grow food for us. And he said the American Farm Bureau Federation estimates there are roughly 2.4 million farm jobs that need to be filled annually; and the agricultural work force continues to shrink. (This is my hint to the anti-immigration crowd). John Deere’s autonomous 9RX Tractor. Farmers can oversee it using an app. While each of these industries experiences their own set of challenges, a commonality across all is skilled labor availability. In construction, about 80% percent of contractors struggle to find skilled labor. And in commercial landscaping, 86% of landscaping business owners can’t find labor to fill open positions, he said. “They have to figure out how to do

2025 playbook for enterprise AI success, from agents to evals

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More 2025 is poised to be a pivotal year for enterprise AI. The past year has seen rapid innovation, and this year will see the same. This has made it more critical than ever to revisit your AI strategy to stay competitive and create value for your customers. From scaling AI agents to optimizing costs, here are the five critical areas enterprises should prioritize for their AI strategy this year. 1. Agents: the next generation of automation AI agents are no longer theoretical. In 2025, they’re indispensable tools for enterprises looking to streamline operations and enhance customer interactions. Unlike traditional software, agents powered by large language models (LLMs) can make nuanced decisions, navigate complex multi-step tasks, and integrate seamlessly with tools and APIs. At the start of 2024, agents were not ready for prime time, making frustrating mistakes like hallucinating URLs. They started getting better as frontier large language models themselves improved. “Let me put it this way,” said Sam Witteveen, cofounder of Red Dragon, a company that develops agents for companies, and that recently reviewed the 48 agents it built last year. “Interestingly, the ones that we built at the start of the year, a lot of those worked way better at the end of the year just because the models got better.” Witteveen shared this in the video podcast we filmed to discuss these five big trends in detail. Models are getting better and hallucinating less, and they’re also being trained to do agentic tasks. Another feature that the model providers are researching is a way to use the LLM as a judge, and as models get cheaper (something we’ll cover below), companies can use three or more models to

OpenAI’s red teaming innovations define new essentials for security leaders in the AI era

Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI coverage. Learn More OpenAI has taken a more aggressive approach to red teaming than its AI competitors, demonstrating its security teams’ advanced capabilities in two areas: multi-step reinforcement and external red teaming. OpenAI recently released two papers that set a new competitive standard for improving the quality, reliability and safety of AI models in these two techniques and more. The first paper, “OpenAI’s Approach to External Red Teaming for AI Models and Systems,” reports that specialized teams outside the company have proven effective in uncovering vulnerabilities that might otherwise have made it into a released model because in-house testing techniques may have missed them. In the second paper, “Diverse and Effective Red Teaming with Auto-Generated Rewards and Multi-Step Reinforcement Learning,” OpenAI introduces an automated framework that relies on iterative reinforcement learning to generate a broad spectrum of novel, wide-ranging attacks. Going all-in on red teaming pays practical, competitive dividends It’s encouraging to see competitive intensity in red teaming growing among AI companies. When Anthropic released its AI red team guidelines in June of last year, it joined AI providers including Google, Microsoft, Nvidia, OpenAI, and even the U.S.’s National Institute of Standards and Technology (NIST), which all had released red teaming frameworks. Investing heavily in red teaming yields tangible benefits for security leaders in any organization. OpenAI’s paper on external red teaming provides a detailed analysis of how the company strives to create specialized external teams that include cybersecurity and subject matter experts. The goal is to see if knowledgeable external teams can defeat models’ security perimeters and find gaps in their security, biases and controls that prompt-based testing couldn’t find. What makes OpenAI’s recent papers noteworthy is how well they define using human-in-the-middle