As Intelligence Approaches Zero, Value Comes From Somewhere Else

Hello hello! I forgot I set up a Beehiiv newsletter…until I checked my statement the other day and realized I'm getting billed $100/month to keep this going. So, I figure I might as well start writing.

Things have been much busier than expected recently. Content-wise, I started a new YouTube show for Benzinga, and we've so far had guests like Norm AI CEO John Nay, Bessemer's Charles Birnbaum, and Hummingbird's cofounder & CEO Joe Robinson. We have a lot of cool guests lined up over the next few months. We're also getting ready for the annual Benzinga Fintech Conference in October in NYC—I've been slammed with that and we have folks like Softbank and Perplexity speaking. If you're doing stuff in AI and finance, please reach out about the show or conference! Also looking to feature more companies (and if you're interested in sponsoring at 500k email newsletter or 250k YouTube subscriber audience, let me know about that too as I have some discounts available.)

Consulting-wise, things have been pretty hectic there too. I recently signed a few new clients and have been working on agentic consumer banking, investment/quant investing, and fraud. Honestly the progress here has been startling, even to me. With the right guardrails and precision, AI Agents can be a lot more functional than I expected. There's still a bit of patchwork required—there's no platform that works out of the box and you need to add certain tools and rules for different types of tasks, but it's been a lot more successful than I expected.

As Intelligence Approaches Zero, Value Comes From Somewhere Else

The investment fund project I'm working on has been a masterclass in how quickly AI can change what we think of as defensible work.

Five years ago a bespoke quant stack felt like proprietary magic; today GPT o3 turns the same work into a 75-cent, three-minute query. That isn't unique to finance: whenever intelligence commoditizes, advantage migrates up-stack—going from clever math to orchestration, workflow ownership, and compounding data loops.

This became even more clear to me recently on a project—it demonstrates not just how real this change in value is but also how fast it happens too.

In December 2024, a fund manager friend reached out to get my thoughts on the AI Agent space. While catching up, I explained how AI Agents could improve and simplify quant investing—gathering research, conducting analysis, even executing trades. Since automating research was top of mind for finance professionals, we decided to get started on a small project around that.

I spent considerable time researching different quant strategies and alpha discovery methods—incorporating options flow, social sentiment data, alternative datasets. The goal was giving AI agents access to comprehensive data sources and letting them determine what to use for each analysis.

The quantitative analysis capabilities already exist inside most AI models; it's more about extracting that information efficiently. A model can conduct a discounted cash flow analysis for Tesla with the right data, but providing the DCF formula saves time and tokens (aka costs). This optimization approach lets the model's intelligence shine through structured workflows.

My solution was straightforward—I built various analysis types and algorithms useful for stock evaluation. It required a different mindset: "If I had a quant who could execute any mathematical calculation effortlessly, what would I request?" (That’s a lie, I actually asked myself if I was Ryan Gosling in The Big Short and had a quant that can do whatever I wanted, what would I tell him to do?) This led to Monte Carlo simulations, scenario generators, and portfolio optimizers.

All this sounds like months of work. It wasn't. The entire system was built in 1.5 weeks of focused brainstorming/strategizing and a few days of hardcore development. AI models acted as thought partners during ideation, making it easier to brainstorm and execute rapidly.

But here's where the story gets interesting. Between starting my project and now, AI models saw dramatic improvements. OpenAI launched Deep Research in February 2025 (built to search, read, and synthesize sources into fully cited briefs) alongside o3-series models that reason step-by-step and autonomously deploy tools—web search, Python, vision—whenever additional data or computation is needed. Add releases from Anthropic, Google, Meta, Deepseek, and others.

So I decided to test something. I gave the newly available o3-mini model a simple prompt: find unique data sources, analyze the market, report back on companies likely to move this week, and suggest specific trades.

I didn't expect much. Within three minutes, it delivered comprehensive analysis with specific stock recommendations and trade suggestions.

All my work on quant calculations and investing strategies was suddenly obsolete. The reasoning model handled that analysis better within its step-by-step process. Instead of running multiple prompts to gather formula components and synthesize results, it executed everything natively. On top of that, the tool-calling ability let it search for data it needed in real-time and answer math questions with Python when necessary. All of it was built-in: I just had to come up with a detailed and well crafted prompt (which is honestly a whole other hurdle in and of itself.)

This revealed something crucial: as models rapidly approach or exceed our intelligence levels, the value of pure cognitive work plummets. Value migrates elsewhere.

The Economics of Intelligence Collapse

The core driver here is cost. Intelligence is becoming more accessible because it's getting exponentially cheaper.

Despite news headlines about expensive AI training runs, the unit economics tell a different story. In 2020, GPT-3's most capable model (Davinci) cost approximately $60 per million tokens. Today, GPT-4o mini charges 15 cents for input and 60 cents for output—a 400× price crash on input and 100× on output. End-to-end, that's roughly $120 → $0.75 for equally complex questions. Context windows have expanded dramatically—from GPT-3's 2-4K tokens to current models supporting 128K+ tokens, with reasoning models like o3 reaching 200K+ tokens—representing a 50-100× expansion in working memory.

For perspective, a quantitative intern can make up to $70,000+ in a summer. Their work can now be done for under $5 in API calls, and exponentially faster too.

Where NOT to Build Your Moat

Skeptics argue every serious company needs proprietary models—custom LLMs or heavy fine-tuning—because that's where defensibility lives. This is mostly wrong.

Custom or heavily fine-tuned models only pay off when two conditions align:

1. You've extracted every drop of value from commodity models

2. Privacy, latency, or domain jargon truly breaks generic performance

Until then, it's over-engineering. Privacy concerns? That’s a routing problem, not a moat—just route it to a small language model or run it locally within your system. Worried about proprietary data leaking (a major concern for financial institutions)? Use a local RAG database to keep info to yourself and send only anonymized embeddings to models if necessary.

Either way, lasting advantage accrues in the workflows and feedback loops you build on top of that data—not in owning another set of weights.

A Practical Playbook

If making products smarter is now an API call, how should you approach AI product development? Here's the framework that's worked across my consulting engagements:

Start with Jobs-To-Be-Done Analysis

AI forces you deeper than traditional user problem-solving. Since successful AI tools require deep business operation integration, think about AI projects through a Jobs-To-Be-Done lens.

Jobs-To-Be-Done works like B2B product development—map all tasks users need to complete from starting point A to end goal Z. It forces end-to-end thinking about problems—perfect for thinking of workflows to automate too.

Take my investment fund project. The core problem: they spend heavily on analysts and want equivalent quality through software. Breaking this down further, an analyst's job involves:

• Finding and gathering information

• Analyzing that information

• Presenting the information

Each represents a specific job requiring execution. What's the information gathering process? How do they analyze data? What deliverable creation workflows exist?

Much of AI's value lies in converting long workflows into simple prompts or software interfaces. Uncovering these workflows is deceptively difficult, but reframing the approach helps.

Build Model Orchestration Infrastructure

When generative AI gained popularity, banks immediately built internal layers integrating various AI models for developer access. Initially, I assumed this was compliance-driven, but it also created orchestration layer foundations.

Models will remain similar yet differentiated—by functionality and user preference. I prefer Claude for writing edits, Gemini Pro or o3 for coding, and o3 for project planning. Some models statistically outperform others at specific tasks, and personal preferences matter too.

Companies need rapid model-swapping capabilities. Beyond functionality and preference, cost and latency create the most important model differentiators. Different tasks require different models, and spending should vary by question complexity.

This suggests either building model orchestration internally or partnering with orchestration-focused startups (vendor relationships are preferable unless compliance requirements mandate internal development).

Historical Patterns: When Technology Gets Commoditized

A lot of this may seem obvious. You can read all about how when technology gets commoditized, the value moves to the app layer from Twitter thought leaders and VC associates. And it is, but it also is different seeing it in practice. It’s not as simple as “AI models are getting cheaper so you need to build apps on top.” That’s better advice for founders starting new projects vs employees building AI into existing products.

When hard technology gets bundled by incumbents, durable value shifts to whoever layers workflow insights or community on top.

GPS Data: Mapping data became widely accessible compared to the early 2000s, with Apple and Google embedding it in smartphones. Yet Waze sold for $1 billion by crowdsourcing real-time traffic data. The raw GPS capability commoditized; value moved to community-generated insights.

Speech Recognition: Premium Nuance licenses once consumed entire IT budgets. Today, Apple, Google, and Microsoft embed free speech-to-text in every device. Yet Otter.ai, Gong, Granola and similar tools built nine-figure businesses layering meeting intelligence, workflow hooks, and team knowledge graphs on commodity transcripts. Raw voice-to-text became cheap; durable value shifted to insights, collaboration, and data flywheels.

What This Means for Your Business

The pattern is clear but the timing varies by industry. The question isn't whether intelligence in your domain will commoditize—it's when, and whether you'll be ready.

Legal research, medical diagnosis, financial analysis, code generation—all following similar trajectories at different speeds. The companies that survive and thrive will be those that recognize the transition early and rebuild their value proposition around orchestration, workflow optimization, and data network effects rather than raw analytical capability.

The intelligence is becoming free. The real value lies in knowing what to do with it.

Note: Specific API pricing and model capabilities are current as of mid-2025 and subject to change as the market evolves rapidly.