In a previous post we put together a set of custom Pipe classes to create the ability to chain together functions to execute a deep research agent.

But a deep research report isn't a single pass of an initial keyword. It requires doing multiple steps where the agent is progressively discovering and uncovering new research questions to pursue.

OpenAI discusses how they developed Deep Research and you should definitely check it out https://openai.com/index/introducing-deep-research/

Another very good article on developing domain specific Deep Research workflows is from Google https://research.google/blog/accelerating-scientific-breakthroughs-with-an-ai-co-scientist/

ChemCrow has also attempted to build a Chemistry agent https://github.com/ur-whitelab/chemcrow-public which can be thought of as a Chemistry Deep Research agent.

As you learn more about a Deep Research workflow, you will see that a reasoning model is one of the critical aspects of developing such as system.

The reasoning model applies chain of thought and self-reflection improve its output. Such models like the o* series of models from OpenAI and r* from DeekSeek have been trained in a specific manner to elicit "thinking".

But what if we can use existing GPT class models to develop the Deep Research workflow.

Towards this goal, I decided to run the "loop" a few times with the addition of asking GPT-4o-mini to identify the next keyword to search for based on the report generated by the previous loop.

STEP_COUNT = 1

starting_keyword = "cost of GPUs"
reports = []

while STEP_COUNT < 4:
  final_report = starting_keyword | get_search_result_links  | pipeline | create_report
  reports.append(final_report)
  starting_keyword = generate_new_search_term(final_report)
  print("Next search with:", starting_keyword)
  STEP_COUNT += 1

The output is very large to reproduce here but let me share the next search terms that the LLM generated.

Performing search on: cost of GPUs

Next search with: "GPU performance comparison between AMD, Nvidia, and Intel"

Next search with: "Future GPU and CPU release predictions 2024"

Next search with: "Upcoming GPU Technologies and Their Impact on Gaming Performance"

The deep research report is provided below for your reference. The way the deep research report below was generated was to send all the reports generated at the end of each step to an LLM and prompt it to create a final detailed report.

Comprehensive Analysis of the GPU and CPU Market Landscape (2025)

Executive Summary

The GPU and CPU markets in 2025 have transitioned into a period of stabilization following dramatic fluctuations in previous years, driven by both consumer demands and technological advancements. This consolidated report integrates findings from multiple research sources, summarizing critical developments, current trends, performance benchmarks, and market predictions that influence consumer choices. The findings provide insights into the competitive dynamics between leading manufacturers—Nvidia, AMD, and Intel—along with recommendations for potential buyers navigating the landscape.

Market Stabilization and Pricing Trends

Current Landscape

The graphics card market is characterized by a more stable pricing environment than in prior years. The significant decline of prices following post-pandemic supply shortages has led to a more favorable atmosphere for consumers looking to purchase new GPUs. Major brands such as Nvidia and AMD remain leaders, with Intel entering the fray through its budget-friendly Arc series. Price segmentation has become a key aspect of the market, clearly distinguishing products into high, mid, and budget categories.

Price Segmentation

• High-end GPUs: Priced above $650, represented by Nvidia's GeForce RTX 4090.
• Mid-range GPUs: Approximately $450, this includes models like the RTX 4070 and RX 7700 XT.
• Budget GPUs: Under $300, featuring offerings such as the RTX 4060 and RX 7600.

These categories allow consumers to match their performance needs with budget constraints, enhancing accessibility.

Consumer Insights and Market Reception

Recent consumer feedback suggests widespread dissatisfaction with the value proposition of new generations of GPUs, cited as overpriced relative to performance. Despite the resolution of supply chain issues, prices have not reverted to pre-pandemic levels, resulting in hesitance among potential buyers.

Performance Analysis: GPU and CPU Differentiation

Comparative Performance Metrics

The analysis emphasizes the different strengths of Intel and AMD CPUs in conjunction with Nvidia and AMD GPUs:

• Intel CPUs outperform in single-threaded applications and gaming scenarios, making them favorable for most gamers.
• AMD CPUs shine in multi-threaded tasks and offer better performance while being more cost-effective in productivity applications.

When focusing on GPUs, Nvidia maintains a lead with its RTX 4090 model, noted for stellar performance coupled with advanced features like ray tracing. AMD's RX 7900 XTX, while a formidable opponent, is often perceived as less capable compared to Nvidia's top-end offerings.

Benchmarking and Performance Evaluation

Performance benchmarks are crucial for potential buyers. Nvidia's leadership in ray tracing provides distinct advantages in certain gaming scenarios, while AMD strives to deliver competitive offerings in terms of value and power consumption. Detailed benchmarking at multiple resolutions reveals differing performance levels across manufacturers, emphasizing the need for consumers to conduct thorough research tailored to their specific gaming and productivity requirements.

Future Predictions: Trends and Expectations in the GPU Market

Anticipated Releases and New Technologies

The report identifies critical future announcements expected in 2025:

• Nvidia is projected to launch the RTX 50 series, anticipated to integrate GDDR7 memory technology, enhancing overall bandwidth and performance.
• AMD’s zenith may be represented by its Ryzen 9000 series along with Threadripper CPUs, allowing for significant advancements in computational power.

Despite optimistic predictions, analysts exhibit a level of skepticism regarding remarkable improvements in price-to-performance ratios for these upcoming products, echoing sentiments from the prior year.

Recommendations for Consumers

Given the current pricing consistency and upcoming hardware launches, consumers are encouraged to make informed buying decisions:

1. Evaluate the necessity of performance upgrades against current hardware.
2. Consider waiting for next-gen releases, particularly in high-end categories, to maximize value.
3. Explore emerging budget options, particularly from Intel’s Arc series, for 1080p gaming without significant financial outlay.

Conclusion

The GPU and CPU markets in 2025 represent a transformative landscape shaped by stabilizing prices, multi-level performance capabilities, and competitive offerings from leading manufacturers. While both gamers and PC builders have access to a diverse range of products, careful consideration of market dynamics, upcoming innovations, and personal usage needs will be crucial for savvy investments. Stakeholders should remain vigilant in tracking technological advancements and market readiness as 2025 unfolds, ensuring that they leverage the best options available that align with their performance expectations and budgetary constraints.

There you have it.

Some of the improvements you can make;

• Use a reasoning model to identify the next set of search terms.
• For each report generated, summarize the key points and pass it to the next report generation step so that each successive report identifies mutually exclusive topics from the previous ones.

This paper is a great overview of Agentic reasoning.

A very interesting topic discussed in the paper is that of developing a Mind Map. Zooming into the mind map from the above figure you can see that each node in the map is a task that needs to be resolved.

What the task type looks like is to be determined by many factors, but creating a collection of tasks with relationships and sequences is pretty much what a workflow is all about.

In our case we would like to find better keyword to continue our search process.

Keyword generation is a very important problem to solve for anyone who has tried to create campaigns using Google Adwords.

The figure below is from a paper "OKG: On-the-Fly Keyword Generation in Sponsored Search Advertising" which describes an On-the-fly Keyword Generation (OKG) agent.

I like the idea of an Agent to generate keywords.

Before diving into Agents and LLMs try to take a more simpler approach. Use Topic models such as https://radimrehurek.com/gensim/index.html on a domain specific dataset.

Explore what keywords cluster together as shown in this great notebook https://www.kaggle.com/code/ingledarshan/gensim-word2vec-tutorial with a Simpsons focus.

Deep Research is a product and technology that brings together so many critical building blocks that we are being presented with a great opportunity to re-examine search, keywords, retrieval, summarisation, report format and layout and so on.

Very exciting.