AAPG Deep Learning - Machine Learning TIG Discussions

Read This First!

Bogdan Michka — Tue, 09 May 2017 09:37:47 GMT

Welcome to AAPG N.E.T., an online space where you can Network, Engage and Talk.

Please complete the following steps before you begin interacting with the discussion boards:

1. Take time to review your profile and privacy settings, paying attention to your display name and photo, as they will be used to identify your posts.

2. Read the Discussion Boards FAQ page to learn the rules and get familiar with the features: http://broken.aapg.donfick.com/career/aapg-net/discussion-boards-faq — paying special attention to "What are the guidelines for interacting with AAPG's discussion boards used by TIGs and SIGs?" http://broken.aapg.donfick.com/help/faq/articleid/39907/what-are-the-guidelines-for-interacting-with-aapgs-discussion-boards-used-by-tigs-and-sigs and "What is not allowed in AAPG's discussion boards used by TIGs and SIGs?” http://broken.aapg.donfick.com/help/faq/articleid/39906/what-is-not-allowed-in-aapgs-discussion-boards-used-by-tigs-and-sigs

3. If you haven't done so, join any Technical Interest Group (TIG) http://broken.aapg.donfick.com/tigs or Special Interest Group (SIG) http://broken.aapg.donfick.com/sigs that interests you and get involved in a conversation — or start your own topic! Remember to be considerate of other users and enjoy the AAPG N.E.T. communities!

LLMs Are Not Large Numeric Models

Patrick Ng — Fri, 09 May 2025 05:21:55 GMT

Mentioned in a post that "LLMs are not large numeric models" [1], for completeness, see visual examples from prompting LLM on Texas Field Production Reports. [2]

Lesson Learned from Exploration: The Need for Calibration

Beware the peril of using LLMs for oil and gas production volumes without proper calibration. Just as in exploration, performing seismic-well ties is a must before sanctioning or drilling another well.

Ensuring data verification and proper contextual alignment is critical when handling numerical assessments in LLM-driven workflows.

Reference links:

[1] https://www.linkedin.com/pulse/insights-from-aapg-doe-geothermal-workshop-shake-off-patrick-ng-htmnc

[2] https://www.linkedin.com/pulse/llms-large-numeric-models-patrick-ng-a1gac/

AI Vibe Coding: Building Apps Without Writing a Single Line of Code

Patrick Ng — Mon, 21 Apr 2025 10:01:14 GMT

What is Vibe Coding?

Imagine developing a full-fledged application—without ever touching a code editor.

Google has quietly launched Firebase Studio, a cloud-based AI-powered integrated development environment that lets you build full-fledged apps using natural language prompts.

Getting hands on? Here is my step-by-step test drive*.

1) Go to https://studio.firebase.google.com/
2) Enter prompt to describe your app, e.g., "An app that takes user input prompt on places of interest, for example, theaters in London, England and returns the locations in a table."
3) Review AI interpretation of #2 for a prototype: look ok, proceed (otherwise, make changes).
4) AI generate app - complete with a user interface for entering places of interest and location details. For example, "Theaters in West End, London"

The app will list a table displaying West End and London in (latitude, longitude). This test drive aims to illustrate the ease of experimentation with AI beyond Chat-GPT and 10x productivity enhancement potential.

Food for thought - This new era of code-free app creation is democratizing programming, making it more accessible and intuitive for everyone—from seasoned developers and curious geoscientists to creative minds with little or no programming experience. Give it a try and share your journey too!

*requires a Google Account.

Bitnets AI: A Legacy of Efficiency Rooted in Geoscience

Patrick Ng — Fri, 18 Apr 2025 03:46:26 GMT

Bitnets take a minimalist approach to neural network computation, quantizing weights into just three values: -1, 0, and 1. This dramatically reduces memory and computational requirements compared to traditional models, making Bitnets far more efficient—especially for deployment on resource-constrained hardware. Microsoft researchers, for example, have developed a hyper-efficient AI model that can run on CPUs [1].

Is it a coincidence that geoscience played a role in the development of Bitnets AI?

Absolutely not. Innovation is rarely linear, and history is filled with examples of breakthroughs born from unexpected connections.

The drive for efficiency and performance has long been a cornerstone of technological advancement. In the 1970s, geoscientists grappled with severe memory and CPU constraints while processing multichannel seismic data. The same principles that optimized computation back then—such as sign-bit data processing—now echo in the design of AI models hungry for efficiency amid the explosive growth of multibillion-parameter architectures. The lessons learned from geophysics continue to shape cutting-edge AI, as detailed in Maximize LLM Throughput: A Reminiscence of Sign-Bit Data Processing [2].

Takeaway - Bitnets are a testament to the power of interdisciplinary inspiration, proving that foundational ideas can resurface in new contexts, reshaping the future of AI.

Reference links:

[1] https://techcrunch.com/2025/04/16/microsoft-researchers-say-theyve-developed-a-hyper-efficient-ai-model-that-can-run-on-cpus/

[2] https://broken.aapg.donfick.com/career/aapg-net/activity/aft/666

A Closer Look at Calibrated Structured Learning: AAPG-DOE Geothermal Workshop

Patrick Ng — Tue, 15 Apr 2025 16:15:05 GMT

Following up on our earlier post of AI for Education (https://broken.aapg.donfick.com/career/aapg-net/activity/aft/679/groupid/979), let's explore further calibrated structured learning in action.

Join us for an exciting AAPG-DOE hosted virtual workshop:

🗓 Geothermal Subsurface Characterization: Geology & Geophysics
🔗 Register: https://www.geode.energy/resources/geothermal-subsurface-characterization-workshop
📍 April 29-30, 2025

This workshop is designed to equip geoscientists, engineers, and energy professionals with cutting-edge methodologies in subsurface characterization. Under GEODE, the focus is geothermal—but the calibrated structured learning approach used applies just as effectively to oil & gas exploration as well. That discussion, of course, is for another day.

Want to stay ahead of the curve in subsurface technology? Don’t miss this opportunity!

AI for Education - An Experiment

Patrick Ng — Thu, 03 Apr 2025 05:09:52 GMT

AI is transforming education, empowering students to learn in innovative ways. A recent VentureBeat article highlights Anthropic’s groundbreaking contributions, such as their “Claude Learning Mode,” which encourages critical thinking by prompting students to actively engage with the material. Meanwhile, Northeastern University is embracing AI across 13 global campuses, weaving it into their academic fabric to prepare learners for the future. You can read more about Anthropic's work [1].

Experiment: Microbes and Energy – A Glimpse into AI’s Domain-Specific Expertise

To explore the breadth of AI’s capabilities, I ran a simple experiment comparing insights from Copilot and domain-specific Enspired-GPT on the latest advancements in microbial energy.

Prompt: “What is the latest on microbes use in energy?”

Here’s Copilot's response:

1) Recycling Lithium from Batteries: University of Surrey researchers have developed a microbial-based method that recovers up to 95% of lithium from used batteries. This eco-friendly approach avoids toxic chemicals and high energy consumption, potentially revolutionizing the battery recycling process.

2) Hydrogen Production: Marine bacteria, such as those from the Vibrionaceae family, are capable of producing hydrogen gas by breaking down formate, presenting opportunities for clean hydrogen energy solutions.

3) Microbial Fuel Cells (MFCs): These innovative devices leverage bacteria to convert organic material into electricity. Recent efforts focus on scaling MFCs for deep-sea sensor power, reducing reliance on conventional batteries.

4) Carbon Fixation: A newly discovered microbe transforms carbon dioxide into energy-rich chemicals through a unique metabolic pathway, paving the way for advances in biofuel production.

Then repeat on domain-specific Enspired-GPT [2], here is the response:

A Denver-based company is utilizing microbes. Recovery rates of metals from ores are sometimes as low as 1 percent. University of Waterloo researchers in Canada are altering DNA in bacteria so that the bacteria can break down plastics. These bacteria already exist in water systems to clean up microplastics. Dr. Marc Aucoin, a professor in the Department of Chemical Engineering, describes these bacteria as bio-robots that can be programmed to get the job done. The bacteria target polyethylene terephthalate (PET), a common plastic found in carpet, clothing, and food and beverages containers that can take hundreds of years to degrade. PETs typically break down into microplastics, which are pieces of plastic less than 5 millimeters long. These microplastics enter the food chain and contain chemicals that are associated with insulin resistance, cancer, and decreased reproductive health. Current studies are investigating safety concerns and how well the plastic-eating trait in this bacterium will be passed along when the bacteria multiply. Aaron Yip, a PhD candidate in the Department of Chemical Engineering, said that microplastic degradation in wastewater treatment plants is a safer application to target. Many of these facilities are already designed to neutralize bacteria in wastewater, which would kill any engineered bacteria prior to discharging water back into the environment. Geoscientists are trained to understand and research nutrient and chemical transport in water through a natural system. They can measure the bacteria’s effectiveness, and might be the first ones to know if the bacteria were to ever get out.

Assessment - AI indeed has the potential to usher in a more calibrated structured learning.

Stay curious! Exploring the endless possibilities of learning with AI is just the beginning. Start your journey with the DeepLearning TIG Blog and let us know about your own experiments and discoveries. Together, we can enspire and innovate!

Reference links:

[1] https://venturebeat.com/ai/anthropic-flips-the-script-on-ai-in-education-claude-learning-mode-makes-students-do-the-thinking/

[2] https://enspired-gpt.com/ (refer to TIG Feb 4th 2025 post https://broken.aapg.donfick.com/career/aapg-net/activity/aft/674/groupid/979)

What is the latest on geothermal drilling?

Patrick Ng — Tue, 11 Mar 2025 02:49:55 GMT

When I put that question to Enspired-GPT[1], here is the response:

"Geothermal energy is attractive because it can provide reliable, continuous power with little to no pollution by extracting heat from the ground. However, fully realizing this potential is challenging because it's difficult to find subsurface water that is hot and clean enough, and can be accessed and utilized. The high cost of drilling geothermal wells and the logistics of extracting heat from subsurface fluids are also challenges.

One of the hardest parts about accessing geothermal heat sources is the extreme heat, as drill bits are generally comprised of materials that don’t do well in such conditions. The unique difficulties experienced when drilling geothermal wells, such as higher temperatures and harder crystalline rocks, require thinking and solutions outside of the typical soft-rock toolbox.

Quaise plans to surpass the Soviet's deepest hole on Earth, which took two decades to reach 7.6 miles, by using conventional drilling to go 1–2 miles down, then switching to its one-megawatt gyrotron, which will be able to vaporize 11.5 feet of rock per hour. Quaise Energy is also using microwaves to drill into deeper and hotter rocks than a typical bit could withstand.

Gradient Geothermal in Denver sidesteps the requirements of drilling geothermal wells and extracting heat from subsurface fluids by utilizing heat from fluids already being produced by oil and gas companies. They use their HXC System fit-for-purpose organic Rankine cycle equipment to cool produced fluids to meet pipeline specifications, as well as generate onsite.

Sage Geosystems is using geopressured geothermal systems (GGS) by drilling deep, hydraulically fracturing the rocks, and instead of pulling anything out of the fracs, they pump water into them, let it heat up, and then pull that hot water out. Fervo Energy uses an injection well to pump hot water in and another well to pull it out, using two parallel wells instead of vertical wells.

Longer laterals in geothermal drilling mean bigger cones of uncertainty and increased difficulty with bit control while maintaining ROP. Advances in geothermal drilling equipment could help facilitate drilling stability for oil and gas. The oil and gas industry's enthusiasm and expertise in testing and adopting new technologies could help quickly propel geothermal forward in a manner similar to the Shale Boom. The tech that helped during the Shale Boom is quite similar to the tech geothermal needs. However, oil and gas rarely drill as deep and hot as geothermal needs, and the typical sensors used to guide drilling won’t work, or work very poorly, at temperatures useful to geothermal."

Then I ask a follow up question "How far can geothermal wells extend?", see attached.

Tell us what you think and whether Enspired-GPT inspires further geothermal exploration.

Reference:
[1] https://enspired-gpt.com/

Who needs a CEO when AI can make strategic decisions?

Patrick Ng — Mon, 03 Mar 2025 03:56:09 GMT

"AI takes on humans at an even higher-level business function: strategic decision-making." [1]

Possibility? Imagine a future where the rapid advancements in AI technology lead to an office environment straight out of a science fiction movie like The Matrix. As AI penetrates the C-suite, it transforms the very fabric of corporate strategy, forever altering the way executives make critical decisions.

Takeaway [2] -

The article discusses the growing role of AI tools in assisting small businesses with their strategic planning processes. According to a study cited, these AI tools perform as well as or better than human experts when generating and evaluating business plans. They can create plans based on input data and assess them using metrics like ROI. The benefits include time and cost savings, consistency across teams, and accessibility for all sizes of businesses.

However, the article also highlights challenges such as the inability to account for unique local conditions or cultural nuances specific to different regions. Additionally, there's a risk of over-reliance on AI without proper human oversight, potentially leading to decisions that lack necessary human intuition and adaptability.

In conclusion, while AI tools offer significant advantages in enhancing small business planning efficiency and accuracy, they should be used as a supplement rather than a replacement for traditional methods, emphasizing the importance of combining AI with human expertise.

Reference:

[1] https://www.msn.com/en-us/money/smallbusiness/ai-tool-can-write-and-evaluate-business-plans-as-well-as-or-better-than-humans-can-research-indicates/ar-AA1A99Ee?ocid=msedgntp&pc=U531&cvid=e10706f3ecce42eb9c6cb8036753732e&ei=20

[2] summary by AI / DeepSeek

Step by Step - refer to https://broken.aapg.donfick.com/career/aapg-net/activity/aft/672

Open Command Prompt from your Desktop (Windows)

Enter command: ollama run deepseek-r1

Enter prompt: Summarize the article from website: https://www.msn.com/en-us/money/smallbusiness/ai-tool-can-write-and-evaluate-business-plans-as-well-as-or-better-than-humans-can-research-indicates/ar-AA1A99Ee?ocid=msedgntp&pc=U531&cvid=e10706f3ecce42eb9c6cb8036753732e&ei=20

Keep learning.

Enspired Productivity Boost

Patrick Ng — Tue, 04 Feb 2025 03:27:44 GMT

Read AAPG Enspired lately?

A few of us wished for a more efficient way to review Sarah's newsletters, beyond capturing snippets in PowerPoints or saving Words, PDFs for reference, especially during meetings or conference calls? Introducing Enspired-GPT.

https://enspired-gpt.com/

Motivation? AAPG's Enspired newsletter, being succinct and curated, inspired us to develop this AI pilot tool. Imagine asking a question about something you read but can't recall the exact details— Enspired-GPT is designed to address this challenge head-on. See attached.

Seek and we shall find, or just ask Enspired-GPT.

Safe AVO Quantitative - Snapshot on AI Coding

Patrick Ng — Sat, 08 Feb 2025 08:54:22 GMT

Update 02.09.2025 - add two more AI models:

Llama3 (Meta)
Falcon3 (TII)

*** Complete benchmark on modeling AVO out of a total of seven, six models executed the code successfully on the first attempt. However, only three produced meaningful AVO results, with one utilizing the Zoeppritz equations, another Shuey’s approximation and a third Aki-Richards equation. ***

Assessing the potential success of an oil and gas prospect is a critical task for any geoscientist. Historical data from an area where producing wells exhibited Class III AVO responses can offer valuable insights. If a new proposed well demonstrates the same Class III AVO response, does this indicate a high probability of success[1]?

In this blog, we delve into this question and extend our exploration further by examining quantitative AVO responses. This additional calibration point provides valuable insights for ensuring safe AI practice [2] and reliable AI applications to AVO modeling.

AI Going Quantitative

Prompt - "AVO is defined as amplitude-versus-offset. Write python code to perform modeling AVO."

We experiment with five LLMs on this AI coding exercise, namely Gemini (Google), Phi4 (Microsoft), Mistral (Mistral AI), Qwen (Alibaba) and DeepSeek (HighFlyer).

Out of these five, four models generated python codes that ran successfully on the first attempt. However, only two produced meaningful AVO results (one uses the Zoeppritz equations and the other Shuey’s approximation).

Quick look at AI-generated plots: https://www.linkedin.com/pulse/safe-ai-avo-patrick-ng-zjtbc/

Takeaway

While LLMs excel in language (qualitative) tasks, their strength lies in providing a cost-effective starting point for quantitative analysis, augmented by geophysics domain expertise. Without any specialized training of the AI models, this exercise demonstrates the potential of AI as a powerful tool for geoscientists, streamlining initial AVO modeling efforts from ideation to experimentation.

Reference to relevant blog posts:

[1] Prompt for a qualitative AVO response: https://broken.aapg.donfick.com/career/aapg-net/activity/aft/671/groupid/979

[2] Safe AVO mentioned in passing: https://broken.aapg.donfick.com/career/aapg-net/activity/aft/673/groupid/979

Safe DeepSeek or AI use for that matter

Patrick Ng — Sat, 01 Feb 2025 02:58:53 GMT

Premise - studying and learning with AI beats AI alone!

In one week, DeepSeek becomes headline news in the media and creates much angst.

Then a question came up on the last post whether "Use DeepSeek on your Desktop" is safe. Here is a comprehensive answer from Notre Dame [1].

"Safe to Use: Local Open Source Model Use."

All in all, we shall practice safe AI (just as safe AVO in seismic).

Reference:

[1] https://ai.nd.edu/news/deepseek-explained-what-is-it-and-is-it-safe-to-use/

Use DeepSeek and Llama on your Desktop

Patrick Ng — Sat, 25 Jan 2025 18:09:59 GMT

Follow up to request from the post yesterday, here are the steps to get you up and running, with hands on experience.

Step 1: Download tool Ollama from https://ollama.com/

Note - click on Models (top left corner), you'd see a plethora of AI models.

Step 2: Install Ollama (for Windows) - clicking the downloaded file and follow the on-screen instructions during the installation process.

Step 3: Open the Command Prompt window on your PC

Step 4: Type the command to check the version of Ollama installed -- double -

ollama --version

If that works, proceed.

Step 5: Type this command to get the model llama3 (by Meta), if this is the first time, will download the model

ollama run llama3

Step 6: Test drive the model with a prompt using the same command as in Step 5. If the model has been downloaded, we shall see >>>, enter your prompt. Example -

What is amplitude-versus-offset?

If you get a response, you are well on your way to download DeepSeek and compare. Exit current model llama3. Enter next to >>> the command shorthand to exit.

/bye

Step 7 (optional): Fetch for DeepSeek with Command Prompt (download will start)

ollama run deepseek-r1

Step 8 (optional): Test drive DeepSeek (just as Step 5 above)

ollama run deepseek-r1

Step 9 (optional): Compare DeepSeek v Llama3, enter the same prompt after >>>

What is amplitude-versus-offset?

Now you have it, feel free to share your observations and preferences (if any) between DeepSeek and Llama3.

Example, use prompt 'How does Llama3 compare with DeepSeek-R1?', I get this from Llama3.

Llama3 and DeepSeek-R1 are both AI-powered language models designed for various natural language processing (NLP) tasks. Here's a comparison of the two:

Llama3:
Developed by Meta AI
Released in 2022
A transformer-based model with 7.5 billion parameters
Pre-trained on a massive dataset (Colossal Clean Corpus, CC-20) containing over 550 GB of text data
Achieved state-of-the-art results in various NLP benchmarks, such as GLUE, SuperGLUE, and Common Sense

DeepSeek-R1:
Developed by Google Research
Released in 2020 (as part of the BERT-Base model)
A transformer-based model with 12 layers, 768 hidden states, and 110 million parameters
Pre-trained on a massive dataset (Wikipedia, BookCorpus, and Common Crawl) containing over 16 GB of text data
Initially designed for question answering tasks but has since been adapted for various NLP applications

Key differences:
Size: Llama3 has many more parameters than DeepSeek-R1 (~7.5 billion vs ~110 million).
Training dataset: Llama3 was pre-trained on a larger, more diverse dataset (CC-20) compared to DeepSeek-R1's dataset combination.
Architecture: Both models are transformer-based, but Llama3 has a different architecture with more layers and parameters.
Release date: Llama3 is relatively newer than DeepSeek-R1, which was initially released as part of the BERT-Base model in 2020.
Applications: While both models have been used for various NLP tasks, Llama3 has achieved state-of-the-art results on multiple benchmarks, whereas DeepSeek-R1 is often used for question answering and related applications.

In summary, Llama3 appears to be a more powerful and versatile language model compared to DeepSeek-R1, with its larger size, diverse training dataset, and better performance on various NLP benchmarks. However, both models have their strengths and can be used in different contexts depending on the specific application and requirements.

New AI Kid in Town - DeepSeek

Patrick Ng — Fri, 24 Jan 2025 10:54:09 GMT

Why the Buzz about DeepSeek?

Lately, the tech world is abuzz with excitement over DeepSeek[1], and for good reason! Imagine a tool that's not just smaller, but also faster and smarter, all while being more cost-effective. That's exactly what DeepSeek promises—and it delivers, according to recent benchmarks.

Taking DeepSeek for a Spin

Try this prompt: "As a geoscientist, you are tasked to assess an oil and gas prospect. In an area where successful producing wells all showed Class III AVO response, if a new proposed well also shows Class III AVO response, would it be a successful well?"

In essence, "While a Class III AVO response increases confidence in the prospect, it is not a standalone guarantee of success. A comprehensive evaluation integrating seismic, geological, petrophysical, and engineering data is necessary to mitigate risks and validate the prospect’s viability. The new well has a higher chance of success if all supporting factors align with the proven play framework."

Testing AI in the Ring: DeepSeek vs. the Rest [2]

Of course, no test drive would be complete without a little friendly competition. So, I pitted DeepSeek against two of the well-known AI companions: Copilot and Gemini. How did they stack up? They are able to answer in their own way and on par.

Experiment with your favorite AI: Claude, Copilot, Gemini, Grok, OpenAI, HuggingFace, etc, and share your experience. More we play, more we learn.

Suggested Reading:

[1] https://www.nature.com/articles/d41586-025-00229-6

[2] https://www.techopedia.com/can-deepseek-r1-take-on-openai-o1

Triangulation of AI by Real Intelligence - Lesson Learned from Recent Events

Patrick Ng — Fri, 13 Dec 2024 03:54:36 GMT

This is a tieback to the last 2 posts and presents a TIG perspective.

Oil & Gas IQ Operational Excellence (Nov 7 2024) - AI Ethics & Bias for Business [1]

Commercial Alignment: During the breakout group discussions, participants from tech and finance sectors of a major industry player deliberated and summarized a crucial insight: we want AI to perform the roles traditionally handled by consultants—analyzing data, reporting findings, and making strategic recommendations. This realization highlights a significant headwind for consultants and advisors as AI innovation advances.

Implications: Notably absent from this operation-focused event were geoscientists. This absence underscores a critical gap. Imagine the boardroom discussions among the CEO, Chief AI Officer, and CFO without a unified definition of AI value. Each company must bridge this chasm to define AI’s true value proposition. Without alignment, outcomes can be unpredictable and volatile.

AI Subsurface Energy (Dec 10-11 2024) [2] - Current State of Data Bowl: The industry faces an overwhelming volume of data growing exponentially by the month, more than we currently know how to manage effectively. Wrangling this data requires careful planning to avoid a daunting and strenuous effort -- or risk disappointment down the road.

Insight: "After all, for a human, it’s enough to observe just a couple of examples to learn something new [3]," as geologist Wallace Pratt famously noted, "Oil is first found in the mind of men," (AAPG Bulletin, 1952) “and women (interview 1976).” Geologists often extrapolate from one well to field-scale development by synthesizing multi-domain information through cross collaboration and drawing analogs from similar geological plays across multi-dimension geography (time and space).

Observation: It’s the value of data that matters most, not merely its volume, variety, veracity, or velocity. Value can evolve, just as business objectives shift over time. AI can play a pivotal role by analyzing, summarizing, and recommending actions. Just as what we heard from the Oil & Gas IQ's operation excellence breakout group.

At TIG, we serve food for thought over the Holidays and invite all to chime in 2025!

Reference:

[1] https://www.oilandgasiq.com/events-opexinoilandgas
[2] https://broken.aapg.donfick.com/career/training/in-person/workshops/workshop-details/articleid/67704#program
[3] https://venturebeat.com/ai/what-happens-when-we-run-out-of-data-for-ai-models/

GenAI a real game changer as EGS?

Patrick Ng — Fri, 06 Dec 2024 10:06:42 GMT

Enhanced geothermal systems (EGS) are not just new but a real game changer.

See Enhanced Geothermal Systems | Department of Energy [1]

Think of EGS as the perfect high-impact technology transfer from the best of Shale Revolution - drilling horizontal and hydraulic fracturing, to open up new areas for geothermal resources development.

With Google-Fervo and Meta-Sage leading the way for a geothermal renaissance in America, we shall take a closer look at the implication for GenAI development in subsurface energy (from the perspective of DeepLearning TIG).

Analog of Super Computing

GenAI as large language models LLMs are the monoliths (model with billions and more parameters and burning GPUs and energy in unprecedented scale). If the end game is artificial general intelligence AGI, monolith may become Gigantopithecus just as IBM Mainframe was at the zenith of enterprise computing (including seismic data processing), until a “seismic” shift to distributed parallel processing and server farms became the norm.

Switching to Microservices

Next week I'd be presenting "AI Application to Geothermal Development" [2] and highlight a different approach to de-risking geothermal development. It is serverless implementation (the opposite of monolith). We're transforming actionable insights by connecting data, leveraging serverless architecture, and unlocking collective intelligence: Real intelligence / human + AI.

Stay tuned, or better yet, register and be there to share your perspective on using AI and delivering a managed and predictable energy transition.

[1] Department of Energy https://www.energy.gov/eere/geothermal/enhanced-geothermal-systems

[2] program: https://broken.aapg.donfick.com/career/training/in-person/workshops/workshop-details/articleid/67704/generative-ai-machine-learning-and-analytics-for-subsurface-energy?srsltid=AfmBOopbCs3qfKyXJ5ja3s0JVwIMCUtrAqcnOtGE3uKYgbGPNhsSfVFV

The Ethics of AI Development and Addressing Cognitive Bias in Data

Patrick Ng — Wed, 23 Oct 2024 14:39:26 GMT

In the rapidly evolving landscape of AI and Large Language Models (LLMs), a significant challenge is recognizing data bias and mitigating its influence across the entire value chain. There is indeed one upcoming event hosted by Oil & Gas IQ.

https://www.oilandgasiq.com/events-opexinoilandgas?utm_source=LinkedIn&utm_medium=Social&utm_campaign=19378.015_OPEXOG_Speaker_Marketing&utm_term=&utm_content=&disc=&extTreatId=7604127

Mark your calendar and register if you can.

Interactive Discussion Groups (IDG M): The Ethics of AI Development and Addressing Cognitive Bias in Data

https://www.oilandgasiq.com/events-opexinoilandgas/agenda/day-3

In two weeks, I will be moderating this session, providing an opportunity to gather collective insights on how operational experience in Oil & Gas can benefit from AI. My thoughts on the discussion threads / topics are as follows:

1) The Ethics of AI Development and Applications.
2) Anticipating and Recognizing Bias - When and How.
3) Best Practices to Manage and Reduce Bias.

Feel free to share your thoughts. Rationale? Bring our collective insights as a community, and make the IDG session a productive endeavor for the IDG audience.

Together we thrive—whether through AI or human ingenuity!

Let's learn fast and have fun doing so.

Maximize LLM throughput - reminiscence of Sign-bit data processing

Patrick Ng — Mon, 26 Aug 2024 07:27:06 GMT

Update - Microsoft AI Research 02.05.2025 reported

Low-bit quantization reduces model size, enabling efficient execution on edge devices (phones) and reducing energy consumption (battery life) ... 2-bit 7B Llama.

https://www.microsoft.com/en-us/research/blog/advances-to-low-bit-quantization-enable-llms-on-edge-devices/?msockid=19de996e3353661d39138dab32f86767

Next up? Microsoft has squeezed LLM down to 2-bit. If Kroger's "Lower than low" commercial had its say, we'd soon be looking at 1-bit (sign bit) LLM—a retro nod to the sign-bit processing of 1973, making a futuristic comeback!

Bits and Pieces history - recall IBM-360 / 370 mainframe used to process seismic data had 16-bit; not until IBM-390, we could finally handle full 32-bit floating-point data. For true amplitudes and amplitude-versus-offset analysis to infer rock properties, that is a must.

LLM quantization v sign-bit* processing Comp - Can't help to make the following comparison when it comes to the number of bits to make LLM run faster and cheaper today, just as seismic data processing in 1973.

A) On one hand, here is a post on LLM infrastructure optimization 08.23.2024 from Google.

https://cloud.google.com/blog/products/ai-machine-learning/selecting-gpus-for-llm-serving-on-gke

Quantization uses fewer bits, "newer model checkpoints are already published in 16-bit precision" and even drop down to 4-bit. "Recommendation - 1: Use quantization to save memory and cost. If you use less than 8-bit precision, do so only after evaluating model accuracy."

B) On the other hand, paper on using sign bits in GEOPHYSICS vol. 38, Issue 6, December 1973,

https://library.seg.org/doi/10.1190/1.1440394

"The statistical properties of sign‐bit semblance were such that this system could do a velocity analysis and an interpretation with no human intervention. In this mode of operation it yielded state‐of‐the‐art accuracy at greatly increased speed and with greatly reduced storage requirements."

Q: so can we use sign-bit for LLM quantization in all matters of GenAI training?

That will be one for the IMAGE audience to take on.

*Copilot assist - sign bit is a specific use of a one-bit representation, but not all one-bit representations necessarily serve as sign bits.

Corroborated reading - "Lack of bits and bytes is what could make or break your AI/ML project" {1} may find one bit [2] an enticing experiment.

[1] https://www.techradar.com/pro/storage-not-gpus-is-the-biggest-challenge-to-ai-says-influential-report-lack-of-bits-and-bytes-is-what-could-make-or-break-your-aiml-project

[2] Samuel Allen - https://wiki.seg.org/wiki/Samuel_Allen

Sam Allen (past president of SEG) co-founded a new company, Geophysical Systems Corp., whose existence is based on an entirely new recording concept, sign bit recording. This revolutionary technique, conceived and implemented by Sam and his company, is making 1000-channel recording practical.

Geothermal Energy Renaissance - Fact or Fiction?

Patrick Ng — Fri, 23 Aug 2024 07:43:52 GMT

In the June 2024 issue of AAPG Explorer, the article "Could There Be a Geothermal Energy Renaissance in the United States?" and in particular, its closing remarks resonated deeply with me: “The current geothermal energy renaissance in the United States might only reach its full potential if energy professionals, regulators, legislators, and stakeholders collaborate to overcome the challenges facing geothermal energy.”

https://explorer.aapg.org/story/articleid/67289/could-there-be-a-geothermal-energy-renaissance-in-the-united-states

One State at a Time

But what if we shift our focus from the entire Lower 48 to the underestimated geothermal giant—the Lone Star State? According to “The Future of Geothermal in Texas,” Texas holds immense untapped potential.

https://energy.utexas.edu/research/geothermal-texas

Here’s our approach, build on post GenAI Experiment on Geothermal (March 2, 2024) https://broken.aapg.donfick.com/career/aapg-net/activity/aft/663/groupid/979 and conduct two complementary experiments (more concretely webapps we can all interact with):

1) AI-First Perspective: Explore geothermal opportunities from a high-level standpoint at Geothermal-GPT. https://geothermal-gpt.com/

2) Getting Quantitative: Gut check project economics and feasibility at Geothermal-TEA. https://geothermal-tea.com/

Take a test drive and share your experience with us! Let’s unlock Texas’ geothermal future together.

Gen AI Experiment on Geothermal

Patrick Ng — Fri, 01 Mar 2024 18:08:12 GMT

We often ponder if an idea can make a play early in the exploration cycle.

Often it is not just geology but the legal framework regarding surface ownership and mineral right that can determine the outcome - one off or at scale.

So what if we can ask Gen AI and start crystalizing idea from there?

Here is a post with short video to illustrate just that possibility.

https://www.shaleforce.com/blog/gen-ai-experiment-on-geothermal-exploration-and-development

Feel free to share your own experience with Gen AI so we can all learn fast and furious. Start the geothermal revolution indeed in the Lunar New Year of the Dragon.

Gen AI working on Map

Patrick Ng — Wed, 13 Dec 2023 15:40:12 GMT

AAPG Webinar on upcycling wells for geothermal in Oklahoma opens new play ideas. (Hey if geothermal can work in OK, it has got to work in TX, right?)

Video recording (courtesy of Susan Nash) https://www.youtube.com/watch?v=VjPezOAQcu0

What if we can use AI to look at a map (image) and generate ideas on where to play?

Here is one exercise using Google Gen AI (from workshop today).

https://www.linkedin.com/pulse/gen-ai-getting-closer-geothermal-explorer-patrick-ng-acrjf

Q: how many other areas that we have maps and couldn't get around to them, now Gen AI may?

Food for thought - expand application of Gen AI across geothermal, critical minerals and hydrogen exploration.