I started this newsletter almost two years ago. I started it to teach myself and reinforce what I learned about AI. Last year, at this time, the AIE had just this newsletter and about 2,000 subscribers. Today we have over 250,000 subscribers and a free training program. If you are interested in a broader overview of AI, here’s

☕️ AI Tangle - AI Tangle is your trusted source for AI news, twice a week.

🎯 The AI Marketing Advantage -  For sales and marketers wanting to get smarter with AI.

💡 AI CIO - Written by one of the Godfathers of DevOps, John Willis, it’s a newsletter for technical leaders looking to integrate AI into their operations.

📚 The AIOS - This is an evolving project. I started with a 14-day free AI email course to get smart on AI. But the next evolution will be a ChatGPT Super-user Course and a course on How to Build AI Agents. As readers of this newsletter, you will get early access when it’s ready.

AI’s Next Big Leap: Thinking Smarter, Scaling Sustainably

How AI’s evolving intelligence is shaping the future while raising sustainability challenges

OpenAI’s latest models, such as o3-preview and o1-mini, represent a fundamental shift in AI performance scaling. Instead of focusing solely on increasing model size, OpenAI has introduced new scaling laws for inference—the computational "thinking" process that occurs post-training (University of Pennsylvania, Professor Ethan Mollick has a great explanation of this).

This approach enables models to perform multiple internal reasoning steps, generating "hidden thinking tokens" to methodically solve problems before producing an output. Research into scaling laws for language model inference indicates that extended AI "thinking" leads to more accurate and contextually relevant responses. However, these improvements follow a law of diminishing returns, requiring exponentially greater computational resources for incremental gains.

A new set of much more challenging evaluations has emerged in response to these advancements, developed by companies, nonprofits, and governments. Even on the most advanced evals, AI systems are making astonishing progress.

In November, the nonprofit research institute Epoch AI introduced a set of exceptionally challenging math problems called FrontierMath, created in collaboration with leading mathematicians. On release, currently available models scored only 2%. However, just one month later, OpenAI’s newly announced o3 model achieved a remarkable score of 25.2%. According to Epoch’s director, Jaime Sevilla, this improvement was “far better than our team expected so soon after release.”

While not true cognition, these leaps in performance mirror human problem-solving capabilities, unlocking more advanced AI applications. Yet, this progress entails significant trade-offs, particularly concerning energy consumption.

A Paradigm Shift in AI Performance

OpenAI’s latest models, such as o3-preview and o1-mini, represent a fundamental shift in AI performance scaling. Instead of focusing solely on increasing model size, OpenAI has introduced new scaling laws for inference—the computational "thinking" process that occurs post-training.

This approach enables models to perform multiple internal reasoning steps, generating "hidden thinking tokens" to methodically solve problems before producing an output. Research into scaling laws for language model inference indicates that extended AI "thinking" leads to more accurate and contextually relevant responses. However, these improvements follow a law of diminishing returns, requiring exponentially greater computational resources for incremental gains.

While not true cognition, this method mirrors human problem-solving, unlocking more advanced AI capabilities. Yet, this progress entails significant trade-offs, particularly concerning energy consumption.

The Hidden Cost of Smarter Thinking

Extended inference times demand substantial computational power, resulting in increased energy consumption. ChatGPT’s daily energy use is equal to 180,000 U.S. households—daily. That’s according to a recent study shared by Forbes.

A single AI query can use up to ten times as much energy as a traditional search, as noted in research on AI energy demands. Concentrated computing clusters are already causing power shortages and voltage fluctuations in certain regions, with studies showing these clusters often exceed local grid capacities.

The rapid expansion of artificial intelligence is placing unprecedented strain on the U.S. power grid, with significant consequences for electricity quality and reliability. A growing issue, particularly near large-scale data centers, is the emergence of harmonic distortions—deviations in the electrical wave patterns that power appliances and homes.

These distortions disrupt the steady flow of electricity, causing appliances to overheat, motors to malfunction, and systems to degrade more quickly. The problem extends beyond immediate appliance damage, as poor power quality increases the risk of electrical fires, voltage surges, and other hazards that threaten public safety and infrastructure stability.

Data centers, the backbone of the AI revolution, are driving these distortions due to their immense energy consumption and proximity to population centers. In areas with heavy data center activity, such as Northern Virginia, power quality readings frequently exceed safety thresholds, impacting millions of nearby residents. These facilities, often the size of small cities, consume electricity at levels exponentially higher than residential or commercial properties, creating localized stress on power grids ill-equipped to handle such concentrated demand. The effects are not limited to urban areas; rural regions near data centers also experience measurable declines in power quality, showing the pervasive nature of the issue.

The U.S. power grid, already grappling with aging infrastructure and rising electrification demands, faces a critical inflection point. Without significant investment in modernization and grid capacity, harmonic distortions and broader power quality issues will only intensify as the number of data centers grows.

These challenges underscore the urgency for coordinated action among utilities, regulators, and policymakers to ensure the grid can meet the demands of AI-driven energy consumption without sacrificing reliability or safety. As AI continues to reshape industries, its energy demands must be met with infrastructure capable of sustaining its growth responsibly.

Competitive Innovations: Efficiency Over Size

While U.S. companies explore longer "thinking" times, Chinese AI startup DeepSeek is adopting a different strategy. Its DeepSeek-V3 model employs a mixture-of-experts architecture, activating only a fraction of its parameters for specific tasks.

This selective activation significantly reduces energy consumption while maintaining performance comparable to leading models. This selective parameter activation allows the model to process information at 60 tokens per second, three times faster than its previous versions. In benchmark tests, DeepSeek-V3 outperforms Meta's Llama 3.1 and other open-source models, matches or exceeds GPT-4o on most tests, and shows particular strength in Chinese language and mathematics tasks.

Only Anthropic's Claude 3.5 Sonnet consistently outperforms it on certain specialized tasks. The company reports spending $5.57 million on training through hardware and algorithmic optimizations, compared to the estimated $500 million spent training Llama-3.1.

DeepSeek’s efficiency-focused approach highlights an important consideration: bigger isn’t always better. As energy constraints become more pressing, innovations that prioritize cost-effectiveness and sustainability are poised to shape the future of AI.

AI Moving Forward: Balancing Innovation and Responsibility

The dual scaling of AI—both in training and inference—represents a remarkable leap in capability, but it also places unprecedented demands on resources. As models "think" longer and tackle more complex problems, the environmental and infrastructural stakes grow higher. Yet, this challenge is also an opportunity.

Future advancements in AI will likely hinge on a combination of factors: increasing efficiency through architectural innovations, expanding access to renewable energy, and rethinking how we deploy and manage computational resources. As demonstrated by models like DeepSeek-V3, prioritizing intelligent design over sheer size can achieve meaningful gains while conserving energy. This shift in focus from brute force to strategic scaling could define the next era of AI development.

Looking ahead, the AI community must adopt a holistic approach. Policymakers, energy providers, and industry leaders will need to collaborate to create infrastructures capable of supporting AI’s growth without exacerbating its environmental impact. At the same time, researchers should continue exploring novel scaling strategies and sustainable practices, ensuring that technological progress remains aligned with global sustainability goals.

Also, as these models get more capable there’s another concern raised by Geoffrey Hinton, the British-Canadian computer scientist often referred to as a “godfather” of artificial intelligence, who has raised fresh concerns about the existential risks posed by AI. Hinton, who recently received the Nobel Prize in Physics for his groundbreaking contributions to AI, now estimates a 10% to 20% chance that the technology could lead to human extinction within the next 30 years.

This marks a stark shift from Hinton’s earlier prediction, where he assigned a 10% probability to catastrophic outcomes for humanity. He also noted that the pace of AI advancement is accelerating much faster than previously anticipated, intensifying concerns about its long-term impact on society.

Ultimately, the success of AI will not only be measured by its ability to think longer or more efficiently but also by the broader impact it has on society. The tools we build today are laying the foundation for tomorrow’s innovations—innovations that must balance ambition with responsibility. By embracing this dual mandate, we can ensure that the AI of the future is as conscientious as it is powerful.

Aimfox - Launch AI-personalized outbound campaigns and streamline lead management on LinkedIn. Automate unlimited LinkedIn accounts, unify their conversations and sync their connections within a single Dashboard.

AISmartCube - Low-code platform to build, automate, and enhance AI tools and assistants. Access a rich library of ready-made solutions, streamline workflows, and integrate public knowledge bases to create smarter tools effortlessly.

Plus AI for PowerPoint - Create PowerPoint decks with AI and edit slides with AI. Plus AI works where you do, so you don't have to learn how to use a new app.
(I like this approach better than another editor that generates slides outside of PowerPoint or Google Slides).

AI Productivity Tip: Think Longer, Not Harder

As I mentioned we are seeing improvements to the models in general. I see two ways to help get better results from today’s models. The first tactic is to break down complex tasks into manageable steps. For example, my team of agents helped me write this letter by breaking the goal of writing a newsletter into tasks.

The biggest problem I see with many AI users today is that they give the LLM or agent too broad a task. So I break things into manageable tasks and for each task, I give several examples.

First, I break it down to do research on my topic and find the latest articles on a topic I supply.

Second, I ask it to outline an article based on the research.

Third, I ask it to create a draft of the article.

Fourth, I then ask it to write the lead.

Finally, I give it some additional tasks like creating the title, subtitle, preview text, and meta information for SEO. Each is a separate task.

Then I edit, revise, and write the final parts of the newsletter.

The next tip is to make the LLM think longer. Here’s a technique you can use to make this happen.

Try including the following in your prompt to maximize results and invoke reasoning. By breaking the task into steps it keeps thinking about

Break down the following task into three logical steps. For each step, evaluate the potential outcomes and provide a concise summary before proceeding to the next.

To analyze a trend in sales effectively, try including the following expanded prompt to maximize results and invoke reasoning. In this case, you could upload your sales reports for the quarter.

# Role 
You are an AI sales analyst. 

# Goal
Break down the task of analyzing the given sales trend into three logical steps. 

For each step:

1. Identify key factors driving the trend (e.g., seasonal changes, promotions, market conditions).
2. Evaluate the potential implications of these factors (e.g., increased revenue, stock depletion, shifts in customer preferences).
3. Recommend actionable strategies based on the analysis, summarizing outcomes for each proposed action.

This structured approach ensures that the AI evaluates trends comprehensively, highlights influencing factors, and suggests actionable insights to improve decision-making. Whether used for identifying growth opportunities, addressing challenges, or forecasting future performance, this method will help you make the most of AI's reasoning capabilities.

Here are some additional ways to get ChatGPT to take its time answering, and use prompts that encourage depth, analysis, and step-by-step reasoning. Here's how:

1. Request Processes: "Explain step by step how this solution works."

2. Comparative Analysis: "Compare two approaches and justify which is better."

3. Support with Data: "Provide evidence or examples to back your answer."

4. Explore Open-Ended Topics: "Discuss how AI might transform creativity over the next decade."

5. Use Scenarios: "Plan an AI strategy for a Fortune 500 company, addressing challenges and ROI."

6. Prioritize Depth: "Provide a detailed, multi-perspective response."

7. Simulate Debates: "Write a debate between AI advocates and skeptics."

8. Iterative Steps: "Design a solution and justify each phase before moving on."

9. Reflect on Limits: "Explain potential flaws or limitations in this approach."

10. Persona Requests: "Act as a slow-thinking philosopher to analyze this issue."

Explicitly ask for thoroughness and reflection to slow responses.

Incorporating these structured prompting techniques can significantly enhance AI performance, leading to more efficient and effective outcomes.

Additionally, for an overview of how new AI advancements are pushing reasoning capabilities further and insights on evolving prompt engineering.