Understanding how the AI you use most often actually works is a challenge for all of us. I recently read Ethan Mollick’s new book (Co-Existence) that offered an observation I found interesting, and I think I was able to test. Mollick suggested that an AI tool could think either fast or slow. The way to explore the difference is what I want to describe here.
You may have encountered this distinction and not known what the options implied. For example, I use Readwise to archive the highlights and notes I create while reading Kindle and Apple Books. Readwise includes an AI chat feature that lets you interact with the stored content. The image below shows the interface for this feature. You can toggle between “Fast” and “Thinking” as alternate ways you want AI to respond to a prompt. This option didn’t make sense to me until I understood it as one way of acting on the distinction Mollick identified.
Fast and Slow Thinking in Humans
Daniel Kahneman wrote an influential book on this topic.
One take away from Kahneman’s Fast and Slow Thinking is that we have a tendency to be cognitively lazy. The way we use AI may be similar. With AI, we often accept prompt replies as fine, but we could demand better and deeper responses. We accept lazy responses from AI and it could do better.
Kahneman’s fast thinking is automatic, intuitive, and quick involving mostly pattern recognition without the purposeful use of effortful cognition. Slow thinking is deliberate, effortful, and analytical. For example, seeing a familiar word and reading it without effort is fast thinking. Carefully working through a complex arithmetic problem is slow thinking. Fast thinking can create problems when it involves the application of stereotypes or problems such as confirmation bias – seeing what we want to see. Slow thinking, when unnecessary, wastes time and attention.
Our metacognition (cognitive oversight and control) offers the opportunity to override the tendency for fast thinking. For example, I have written on several occasions that research findings show taking notes by hand is superior to taking notes on a keyboard, based on the desirable difficulty hypothesis. The claim is that people can type faster than they can write and when typing is available, they tend to record verbatim information rather than use more constructive processes such as summarization. If you consider this claim carefully, you might wonder why, if typing is easier and faster, a person using a keyboard does not take advantage by using these opportunities to also summarize and analyze. Verbatim recording is not required. Sometimes, without insight and training, people may ignore opportunities that are right in front of them. I will claim we may ignore a similar opportunity when prompting AI.
More than retrieval – reasoning with an LLM
It is difficult for most of us to understand how an LLM does what it does. Mollick suggests we may think that responses to a prompt are just an improved version of search. However, you or I understand the abstract nature of the “understanding” embedded in an LLM, Mollick proposes that retrieving answers to prompts (questions) is really the lower limit of what AI can provide. Even when the discovery of patterns in data available for retrieval that we have previously not identified can be quite surprising, but even then is pattern discovery all there is? When we use a word like reasoning, is that different and can LLMs reason?
Mollick proposes that what we call reasoning is a process involving multiple components that go beyond a pattern-recognition summary. LLMs trained with reinforcement learning can take a multi-step approach that resembles the following.
- Identify the question and any assumptions that need examination.
- Break the problem into its major components.
- Consider plausible alternative explanations or interpretations.
- Evaluate the evidence and distinguish established facts from inferences.
- Look for counterexamples, contradictions, and weaknesses in the reasoning.
- Check important factual claims against reliable sources when appropriate.
- Provide a concise conclusion that reflects the strength and limitations of the evidence.
For example, here is how I might prompt AI to address the handwriting vs. keyboarding controversy I use in many of my examples.
Instead of asking:
What does the research say about whether handwritten notes produce better learning than typed notes?
Ask:
Conduct a critical analysis of research comparing handwritten and typed note-taking. Identify the major studies, distinguish immediate recall from delayed learning, examine the quality of the research designs, consider competing explanations, identify limitations, and explain where the evidence supports firm conclusions versus tentative ones. Cite the original research where possible.
The second prompt defines what a satisfactory analysis requires. It encourages deeper work rather than simply a longer response.
Require independent verification.
You might also ask the AI to check its answer against original sources, recalculate numerical results, test competing explanations, or look for counterexamples. For research questions, this is particularly valuable because a convincing summary can conceal we
One final example created in a different way
You could fashion a prompt yourself, but some AI tools will understand a request to think and analyze before responding, doing something similar to what I outlined. I wanted to show the difference between what AI can do when thinking fast and slow. I was using the local AI I have running on my Mac mini and stumbled upon the following demonstration. I first prompted the tool to explain the difference between fast and slow AI functioning. I then followed up by asking the AI to demonstrate the difference it had just described with a request I would specify. Finally, I asked for a description of the difference in what the two approaches involved. The following takes you through this sequence
Prompt: Explain how an AI tool thinks fast and slow.
Let me think about how to explain this simply:
Traditional LLMs: quick, direct answers, like a fast talker
Reasoners: slower, more thoughtful, like someone who thinks before speaking I need to capture the essence: internal reasoning chain vs. immediate generation, better performance on complex tasks, trade-off with speed.
Prompt: Show me what you are doing as you apply each approach.
Demonstrate this difference with the topic – what challenges to the economy would a democratic majority in the upcoming election produce
To illustrate the difference, here is how both approaches tackle a complex, politically nuanced question:
1. Traditional LLM Approach (Direct Pattern Completion)
How it works internally: The model predicts words sequentially based on common correlations in its training data without pausing to deliberate or structure an outline.
Output:
A Democratic majority could introduce several economic challenges, primarily centered around fiscal policy, taxation, and regulatory enforcement. Key concerns often raised by economists and market analysts include:
Corporate and High-Earner Tax Increases: Proposals to raise the federal corporate income tax rate and increase capital gains taxes could reduce corporate net margins, potentially dampening domestic capital investment and research spending.
Regulatory Burdens and Energy Transition Costs: Stricter labor regulations, environmental compliance standards, and aggressive decarbonization mandates can raise compliance overhead and production costs for legacy energy, manufacturing, and transportation sectors.
National Debt and Inflationary Pressure: Substantial federal investments in social infrastructure, climate initiatives, and entitlement expansions could widen fiscal deficits if spending outpaces new revenue, potentially putting upward pressure on interest rates.
2. AI Reasoner Approach (Chain-of-Thought & Critical Deliberation)
What happens during its internal reasoning chain before responding:
[Reasoning Phase – Internal Thinking]
1. Intent & Scope:
– Topic: Economic challenges resulting from a Democratic majority.
– Requirement: Detail potential downsides while remaining analytically objective and grounded in mainstream economic theory.
2. Disentangling Policy Levers:
– Tax side: Higher statutory corporate rates (e.g., 21% -> 28%), top marginal brackets, capital gains revisions.
* Mechanism: Impact on cost of capital, investment decisions, corporate tax incidence (wages vs. returns).
– Regulatory side: FTC/DOJ antitrust posture, labor market rules (PRO Act provisions, overtime expansion), climate mandates.
* Mechanism: Heightened compliance costs, friction in M&A/deal-making, energy transition friction.
– Fiscal/Monetary interaction: High spending priorities vs. deficit targets in an era of higher baseline interest rates.
3. Distinguishing Immediate Market Shocks vs. Structural Headwinds:
– Short-term: Equity market re-pricing due to anticipated margin compressions and regulatory scrutiny.
– Long-term: Potential dampening of private R&D vs. productivity gains from public infrastructure.
4. Filtering & Refinement:
– Ensure arguments represent recognized fiscal and economic perspectives (e.g., CBO, Tax Foundation, macro consensus) rather than purely partisan talking points.
– Synthesize logically from immediate market frictions to long-term macroeconomic impacts.
Output:
If a Democratic majority takes control of both the executive and legislative branches, the primary economic challenges would stem from the structural shift toward higher public-sector intervention, progressive taxation, and stricter market oversight:
>Marginal Cost of Capital and Private Investment: Legislative proposals typically emphasize raising the corporate statutory tax rate (e.g., from 21% toward 28%) and reforming international tax frameworks. From an economic perspective, these measures reduce corporate after-tax returns, which can disincentivize domestic capital expenditure, prompt capital reallocation, and compress corporate profit margins.
Frictional Costs of Decarbonization and Sector-Specific Transitions: Accelerated regulatory enforcement—particularly through tighter EPA standards, limits on fossil fuel extraction, and clean energy mandates—tends to increase near-term input costs for energy, utilities, and transportation. While intended to achieve environmental sustainability, these policies can generate supply-side friction and transitional energy price volatility before renewable capacity fully scales.
> M&A Friction and Heightened Regulatory Oversight: A Democratic administration generally maintains an aggressive antitrust and labor posture (e.g., via the FTC, DOJ, and NLRB). Tighter scrutiny over mergers, stricter labor union protections, and expanded worker classification rules can raise corporate compliance overhead and slow consolidation in sectors like technology, healthcare, and retail.
> Sovereign Debt Burden in a Higher-Yield Environment: Ambitious spending programs aimed at the social safety net, housing subsidies, and healthcare expansion risk exacerbating structural budget deficits. In an economic climate where baseline Treasury yields and borrowing costs are elevated, higher sovereign debt issuance can crowd out private borrowing and constrain future fiscal flexibility.
Show my adjustments the reasoner approach made as it worked
Here is a look “under the hood” at the specific adjustments, course corrections, and self-checks the reasoner makes while formulating the response:
1. Tone Correction: Removing Partisan Rhetoric
First Instinct: The model initially recalls common political debate points, such as *”crippling business growth with anti-corporate penalties”* or *”reckless social spending causing rampant inflation.”*
The Reasoner’s Correction: Recognizes that value-laden words sound biased rather than analytical. It swaps subjective rhetoric for precise economic terminology:
* *Drafted thought:* “Businesses will stop investing because taxes are too high.”
* *Adjusted to:* “Higher corporate statutory rates increase the marginal cost of capital, potentially disincentivizing private capital expenditure.”
2. Disentangling Policy from Economic Mechanisms
**First Instinct:** Simply listing Democratic platform items (e.g., green subsidies, raising the minimum wage, student loan relief).
**The Reasoner’s Correction:** Realizes a list of policies is not an explanation of *economic challenges*. It traces each policy to its secondary transmission channel:
* *Drafted thought:* “They want stricter climate regulations and electric vehicle mandates.”
* *Adjusted to:* Traces the supply-chain friction—identifying how rapid regulatory changes increase near-term input costs for heavy industry and utilities before cleaner alternatives reach full scale.
3. Factoring in Current Macroeconomic Realities
**First Instinct:** Treat government borrowing in the abstract (e.g., “Deficits increase the national debt”).
* **The Reasoner’s Correction:** Grounds the critique in today’s higher interest-rate climate rather than the near-zero-rate environment of previous decades:
* *Drafted thought:* “More spending increases the national debt.”
* *Adjusted to:* Evaluates how deficit-financed programs interact with higher baseline Treasury yields, highlighting the risk of crowding out private credit and escalating debt servicing costs.
4. Structuring for Analytical Depth Rather Than Breadth
**First Instinct:** Try to touch on every possible friction point (healthcare rules, immigration policy, trade tariffs, tech regulation, banking scrutiny).
**The Reasoner’s Correction:** Recognizes that covering ten disparate areas leads to a superficial, shallow list. It clusters related factors into four core structural pillars:
1. *Capital & Investment* (tax structure and incentives)
2. *Supply-Side Transition Costs* (energy and environment)
3. *Operational & Transaction Frictions* (labor rules and antitrust)
4. *Macroeconomic & Fiscal Constraints* (sovereign debt and rates)
Summary of the Shift
“`
[Superficial Surface Association]
“Democrats spend money and raise taxes -> bad for businesses and causes debt.”
?
[Reasoner’s Internal Adjustments]
• Strip partisan spin ? use economic theory (cost of capital, compliance drag)
• Trace indirect effects ? explain the *mechanism* of transition friction
• Account for macro environment ? evaluate debt in an elevated-interest-rate regime
?
[Refined Output]
Structured, objective analysis focused on capital allocation, regulatory friction, and fiscal mechanics.
*****
For a further discussion of Fast and Slow AI try thus blog post from Precision Weave.
You must be logged in to post a comment.