TellDear Is a Bet on the Future of AI Economics

In 1965, Gordon Moore observed that the number of transistors on a chip doubled roughly every two years. He didn't predict the personal computer, the smartphone, or the cloud — but he described the economic force that made all of them inevitable. The observation wasn't about technology. It was about cost. And cost, once it starts falling exponentially, reshapes everything it touches.

Something analogous is happening in artificial intelligence right now. Not with transistors, but with inference — the cost of asking an AI model to think. And TellDear is built as a bet on that curve.

The Collapsing Cost of Thinking

In March 2023, analyzing a page of text with GPT-4 cost roughly $0.03 in API fees. By late 2024, equivalent capability cost a tenth of that. By early 2026, models that match or exceed GPT-4's reasoning abilities are available at costs that would have seemed absurd three years ago. The trend isn't slowing — it's accelerating.

This isn't just hardware getting cheaper (though it is). It's a convergence of multiple forces: architectural improvements like mixture-of-experts that activate only the parameters needed for a given query; distillation techniques that compress large models into smaller, faster ones without proportional quality loss; inference optimization through speculative decoding, quantization, and KV-cache tricks; and sheer competition — dozens of companies racing to offer the best model at the lowest price.

The result is a cost curve that looks less like Moore's Law and more like the early days of DNA sequencing after next-gen methods arrived: a cliff, not a slope. What cost a dollar in 2023 costs a penny in 2026. What costs a penny today will cost a fraction of a fraction tomorrow.

This matters because intelligence — artificial intelligence, specifically — is becoming a commodity input. Like electricity. Like bandwidth. Like compute itself. And when an input commoditizes, the value shifts to what you do with it.

What TellDear Actually Costs (Today)

Let's be concrete. TellDear's analysis engine examines text across 452 aspects organized in six dimensions: logical fallacies, cognitive biases, rhetorical techniques, statistical errors, epistemological issues, and structural weaknesses. At the "Thorough" depth level, a full analysis involves multiple passes over the text — identifying patterns, cross-referencing aspects, evaluating argument structure, and generating a detailed report.

Today, a thorough analysis of a typical opinion piece costs roughly $0.15–0.30 in inference. That's not expensive in absolute terms — cheaper than a cup of coffee, cheaper than a single Google ad click. But it adds up. If you wanted to analyze every article in a daily newspaper, you'd spend $15–30 per day. If you wanted real-time analysis of a political debate, you'd need multiple parallel inference calls running continuously for 90 minutes. The cost would be significant.

These economics shape what TellDear can offer today. The analysis lenses are powerful but not instantaneous. The depth levels exist partly because deeper analysis costs more. The system is designed to be thorough and accurate — but it operates within the constraints of current pricing.

Now imagine those constraints dissolving.

The Moore's Law Bet

TellDear is architecturally designed for a future where inference is nearly free. This isn't optimism — it's engineering for the trajectory.

Consider what happens when the cost of a thorough analysis drops from $0.25 to $0.01. Suddenly, analyzing every article in a newspaper costs less than the newspaper itself. At $0.001 per analysis, a browser extension could evaluate every article you read in a day for less than a cent total. At $0.0001, real-time analysis of live speech becomes economically trivial — a running commentary on a political debate, flagging straw man arguments, false dichotomies, and appeals to fear as they happen.

Each order-of-magnitude cost reduction doesn't just make existing features cheaper. It unlocks entirely new use cases that were previously unthinkable.

The capability cascade

At current prices (~$0.15–0.30 per thorough analysis):

• Individual articles, analyzed on demand
• Three depth levels to manage cost-quality tradeoffs
• Batch analysis feasible but not casual

At 10× cheaper (~$0.015–0.03):

• Browser extensions analyzing every article you open
• "Always-on" analysis for news readers
• Classroom tools where every student analyzes every text, every day
• The classroom use cases become free enough to be universal

At 100× cheaper (~$0.0015–0.003):

• Real-time analysis of live speech and video
• API integration for social media platforms (flag misleading posts at ingest)
• Continuous monitoring of news sources for rhetorical shifts
• The kind of journalistic analysis that currently requires deliberate effort becomes ambient

At 1000× cheaper (~$0.00015–0.0003):

• Every email, every tweet, every comment — analyzed passively
• Argumentation quality scores as metadata on all digital text
• TellDear as infrastructure, not an application

We're not speculating about thousand-fold cost reductions over decades. Based on the trajectory from 2023 to 2026, we've already seen roughly 30–50× reduction in effective inference cost for equivalent capability. Another 20× over the next two to three years is conservative, not optimistic.

The Taxonomy Is the Moat

Here's the critical insight: models change, but the 452 aspects don't. Or rather, they change slowly and deliberately — through intellectual work, not through the next model release.

When GPT-4 gives way to GPT-5, or when Claude improves its reasoning, TellDear doesn't have to rebuild. The taxonomy — the structured understanding of what a confirmation bias looks like in text, how a red herring diverts attention, why evasion is structurally different from straw-manning — is durable intellectual infrastructure. Better models mean better detection of the same aspects. The knowledge graph that maps relationships between aspects becomes more powerful automatically, because a smarter model can navigate it more effectively.

This is the opposite of the "wrapper" critique that haunts AI startups. A wrapper adds a UI on top of a model and prays the model doesn't add that UI natively. TellDear adds a structured intellectual framework that no model contains natively — because it was built through years of cataloging, categorizing, and connecting patterns of human reasoning failure. The model is the engine. The taxonomy is the map. You can swap engines. You can't drive without the map.

As we argued in Wrapper Economics, the question isn't whether you use someone else's model — it's what you add on top. Domain expertise, structured knowledge, and intellectual infrastructure survive model transitions. Prettier interfaces don't.

What This Means for the Mission

TellDear's mission is to make critical thinking tools universally accessible. Today, that mission operates under constraint: analysis costs money, which means it either costs users money or requires subsidization. Neither scales to "universal."

But the economic trajectory solves this problem without charity. When inference costs drop by two or three orders of magnitude, offering free critical thinking analysis becomes economically trivial — comparable to serving a web page. The marginal cost of helping one more person think more clearly approaches zero.

This is not a business model pivot. It's a bet that the economics of AI will catch up to the ambition of the mission. Build the taxonomy now. Build the UX now. Build the detection quality now. The cost barrier is temporary. The intellectual infrastructure is permanent.

Compare this to the early days of internet video. In 2000, streaming a single video to a single user was expensive enough to bankrupt most startups. Bandwidth cost real money. YouTube launched in 2005 and hemorrhaged cash on bandwidth bills. But the cost of bandwidth was falling exponentially, and YouTube was betting that it would keep falling. They were right. Within a few years, streaming video went from "ruinously expensive" to "essentially free." YouTube didn't invent cheaper bandwidth. They built the infrastructure that would become valuable once bandwidth was cheap enough.

TellDear is making the same bet, but with intelligence instead of bandwidth.

The Speed Dimension

Cost isn't the only variable on this curve. Speed matters equally. As we explored in Analysis Speed, the time required for a thorough analysis directly shapes the user experience and the possible use cases.

Today, a thorough analysis takes 30–90 seconds depending on text length. That's fine for deliberate, sit-down analysis. It's useless for real-time applications. You can't flag a logical fallacy in a debate if the flag arrives two minutes after the speaker has moved on.

But inference speed is improving even faster than cost is falling. Techniques like speculative decoding, specialized inference chips, and architectural innovations are pushing latency down. When a thorough analysis takes under two seconds, the user experience transforms. Analysis becomes conversational. Interactive. Ambient.

Imagine reading a news article while subtle annotations appear in the margins — not after you click "analyze," but automatically, as your eyes move down the page. A gentle highlight on a paragraph that contains a false equivalence. A tooltip noting that a statistic lacks context (see How Numbers Lie for why this matters). Not intrusive. Not preachy. Just... present. Like a spell-checker for reasoning.

That experience requires both cheap and fast inference. The cost curve gives us cheap. The speed curve gives us fast. Both are moving in the right direction.

The Uncomfortable Present

Betting on the future doesn't make the present comfortable. Today, TellDear faces real constraints:

• Cost limits scale. Every free analysis is a subsidy. Offering unlimited free analysis would be financially unsustainable at current prices.
• Speed limits UX. The most powerful depth level requires patience. Not everyone has it.
• Model limitations limit accuracy. Current models are good at detecting rhetorical patterns, but not perfect. They miss subtleties. They occasionally hallucinate aspects that aren't there. As explored in the AI slop problem, even tools designed to improve thinking must guard against their own models' weaknesses.

These are real limitations. They're also temporary. Every one of them is on a trajectory toward resolution — not through hope, but through documented, measurable trends in AI capability and cost.

The question isn't whether these constraints will ease. It's whether TellDear will have built the right infrastructure when they do.

Why Build It Now?

If the economics are going to get dramatically better, why not wait? Why build at today's costs when tomorrow's will be so much lower?

Because infrastructure takes time to build, and the window for building it is now.

The knowledge graph connecting 452 aspects in six dimensions wasn't built in a weekend. The relationships between a tu quoque fallacy and whataboutism, between anchoring bias and framing effects, between statistical cherry-picking and publication bias — these connections represent thousands of hours of intellectual work. They need to exist before the economics make deployment trivial, not after.

Similarly, the user experience, the pedagogical approach (explored in TellDear in the Classroom), and the detection quality all benefit from iteration. Users today provide feedback that improves the system for users tomorrow. The Fallacy Trainer teaches recognition patterns that are refined with each cohort of learners. This learning loop needs time, and time is the one resource you can't compress.

The companies that win on exponential cost curves are the ones that build the infrastructure during the expensive phase and ride the curve down. The ones that wait for cheap economics find that someone else already built the map.

A Bet, Not a Guarantee

Let's be honest about what this is: a bet. Bets can lose.

The cost curve could plateau. Some fundamental physical or economic limit could slow the decline in inference costs. Regulation could intervene. A major geopolitical disruption could constrain chip supply. The open-source ecosystem could fragment in ways that increase rather than decrease costs.

Or the market could solve the problem differently. A future model might have critical thinking analysis built in natively, making a dedicated taxonomy unnecessary. (We think this is unlikely — structured dimensional analysis is qualitatively different from general reasoning — but it's not impossible.)

TellDear is a bet that the trend continues: that inference gets cheaper, faster, and better; that structured domain knowledge remains valuable even as models improve; and that helping people think clearly is a mission worth building infrastructure for, even before the economics fully support it.

It's the same kind of bet that every durable technology company has made. Build for where the curve is going, not where it is. The economics will either vindicate the bet or they won't. But the alternative — waiting for certainty — is not a strategy. It's a way of ensuring that when the future arrives, you have nothing built for it.

The Endgame

Picture a world where analyzing the reasoning quality of any text is as cheap and fast as checking its spelling. Where a student can paste a political speech into a tool and receive, in under a second, a structured breakdown of every logical fallacy, every cognitive bias appeal, every statistical manipulation — with explanations, examples, and links to learn more. Where a journalist's CMS automatically flags weasel words, appeals to authority, and misleading precision in their own drafts before publication. Where "How strong is this argument, really?" is a question anyone can answer, about anything, instantly, for free.

That world doesn't require a breakthrough. It requires the continuation of a trend that has been consistent for three years and shows no signs of stopping. It requires someone to build the intellectual infrastructure — the taxonomy, the knowledge graph, the detection patterns, the pedagogical framework — so that when the economics arrive, the capability is ready.

That's what TellDear is. Not a product optimized for today's constraints. A bet on tomorrow's economics, built with today's conviction that critical thinking shouldn't be a luxury.

The curve will tell us if we're right.