DecodeLM LogoDecodeLM
Try for free

Philosophy

AI that answers for you
makes you worse at thinking.

Helpfulness and learning are not the same thing. Most AI is designed to be helpful. DecodeLM is designed to make you learn. Here is the argument built entirely from research.

"The distinction between AI systems that enhance learning and those that destroy it is not about the underlying technology. The difference lies entirely in design."

Carl Hendrick, The Learning Dispatch · on Bastani et al., PNAS 2025

Part I

How learning actually works

Decades of cognitive science have established what makes knowledge stick. None of it is controversial. Most AI ignores all of it.

2σ

One-on-one tutoring produces a 2 standard deviation improvement

Bloom, 1984 · Educational Researcher

Benjamin Bloom's landmark 1984 study found that students who received one-on-one tutoring performed two standard deviations better than students in conventional classrooms, meaning the average tutored student outperformed 98% of the classroom group. Bloom called this the '2-sigma problem': how to achieve the results of individual tutoring at scale. For 40 years, no one solved it.

Our response

DecodeLM is designed to approximate the 2-sigma effect. A tutor that knows your background, tracks your knowledge, adapts to your gaps, and makes you do the thinking. Available to everyone, not just those who can afford a private teacher.

30–50%

Retrieval practice beats re-reading

Adesope et al., 2017 meta-analysis · Rowland, 2014

Testing yourself improves long-term retention by 30–50% over passive re-reading. The mechanism is called the testing effect: the act of retrieving information from memory strengthens the neural pathways that store it. Reading or watching passively does not.

Our response

DecodeLM builds retrieval practice into every session. The tutor asks you to recall before it explains, not as a quiz feature, but as the default mode of interaction.

Struggle first

Productive failure accelerates deep learning

Sinha & Kapur, Review of Educational Research 2021

Across 15+ years of research, Manu Kapur found a consistent pattern: students who attempted a problem before being taught the solution learned significantly more deeply than those taught first. Failure activates prior knowledge, surfaces gaps, and creates a readiness to learn that makes the correct answer land differently.

Our response

DecodeLM will not rescue you from difficulty. The tutor helps you navigate struggle, because the gap between 'I don't know' and 'I understand' is where learning actually happens.

FSRS

Memory follows a forgetting curve. Spaced review reverses it.

Ebbinghaus, 1885 · Wozniak, 1987 · Open Spaced Repetition (FSRS v5), 2024

Hermann Ebbinghaus showed in 1885 that memory decays exponentially, and that reviewing material at precisely the right moment before it's forgotten reverses that decay permanently. Piotr Wozniak formalized this into the SM-2 algorithm in 1987. FSRS v5 is the current state of the art: it models each item's stability and difficulty individually, using a neural-network-trained formula to predict exactly when you need to see each concept again.

Our response

Every concept you learn with DecodeLM is tracked using FSRS v5. The tutor knows how stable each idea is in your memory and resurfaces it at the optimal moment. When you quiz yourself, the AI grades your response and updates your schedule automatically.

Part II

What default AI does to learning

When you give students an AI that answers for them, you get the opposite of everything in Part I. The research is recent, controlled, and damning.

18% vs 54%

Students themselves draw a line between using AI and learning.

Pew Research Center, 2025

Pew found 26% of U.S. teens use ChatGPT for schoolwork, double the rate from 2023. Only 18% find it acceptable to use for writing essays, while 54% accept it for researching new topics. Students themselves sense the difference between using a tool and replacing thought. The research on cognitive offloading confirms what they sense: sustained AI assistance produces superficial learning, loss of agency, and over-dependency.

Our response

DecodeLM is not designed to help you finish faster. It is designed to help you understand deeper. Speed is the metric of a tool. Understanding is the metric of a teacher.

−17%

Unrestricted AI conditions the brain to stop thinking

Bastani et al., Wharton / PNAS 2025

In a randomized controlled trial of ~1,000 high school math students, those who used ChatGPT freely scored 48% better during AI-assisted practice, then 17% worse on the unassisted exam than students who never had AI access. Their brains had been conditioned to offload thinking rather than build it. The performance gain was an illusion: the AI was doing the cognitive work, not the student.

Our response

DecodeLM never gives you the answer directly. When you ask a question, the tutor asks one back, guiding you to construct the understanding yourself. The knowledge that results is yours.

Neural degradation

Months of AI use measurably weakens brain connectivity

Kosmyna et al., MIT Media Lab 2025

MIT researchers put EEG monitors on 54 participants across 4 months. ChatGPT users showed systematically weaker brain connectivity than search engine users or no-tools controls. When the tool was removed, they could not recover; their brains had been functionally conditioned to lower cognitive effort. They also couldn't accurately quote essays they had written with AI assistance.

Our response

Every interaction with DecodeLM is designed to require cognitive effort. The tutor asks before it tells. The difficulty is not a bug; it is the mechanism by which learning occurs.

Part III

What happens when AI follows the science

The same studies that documented the harm also tested a different design, one built on the learning science from Part I. The results are not subtle.

127%

Guardrailed AI produces durable gains (same study)

Bastani et al., Wharton / PNAS 2025

In the same Wharton trial, a second group used an AI tutor designed around the learning science: it gave hints instead of answers. They scored 127% better during practice and did not show the learning losses seen in the unguarded group. Same underlying model. Same students. Completely different design. Opposite outcomes. The technology was not the variable. The design was.

Our response

This is the design DecodeLM is built on. Hints, not answers. Questions, not explanations. The tutor makes you do the cognitive work, because that work is the learning.

2×

Socratic AI outperforms the best human classrooms

Kestin et al., Harvard / Scientific Reports 2025

194 physics students were randomized into two groups: AI-tutored vs. active learning classrooms, the gold standard of human teaching. The Socratic AI tutor produced learning gains more than double the classroom group. Students also reported higher engagement and motivation. Other controlled studies have shown that unguided ChatGPT access makes students worse. The technology is the same. The design is the variable.

Our response

The Harvard tutor followed one rule: never reveal the full solution; give one step at a time and make the student try. That is the rule DecodeLM's tutor follows by default.

Our design principles

These are not aspirations. They are hard constraints that every feature is built around.

01

Socratic by default

The tutor asks questions before giving answers. Always. This is not a feature you can turn off. It is the product.

02

Personalized to your knowledge, not your preferences

The tutor builds a model of what you actually understand, not just what topics you say you're interested in. It calibrates difficulty, skips what you know, and focuses on your real gaps.

03

Productive struggle is the mechanism, not a side effect

When you don't know something, the tutor does not give you the answer. It helps you construct it. The difficulty is intentional.

04

Knowledge compounds across sessions

Every session extracts what you learned and stores it. Future sessions build on it. The tutor never re-explains what you already know. Your understanding only moves forward.

05

Spaced retrieval, not re-exposure

DecodeLM uses FSRS v5, the current state of the art in spaced repetition, to schedule when each concept resurfaces. It models the stability and difficulty of every idea you've learned and surfaces it at exactly the right moment. When you answer a quiz question, the AI grades it and updates your schedule automatically.

06

Community knowledge, human-verified

The knowledge base is built by people who understand the field: writers who go through expert peer review. The AI tutors from this knowledge. It does not hallucinate foundational concepts.

See the principles in practice.

The best way to understand the difference is to experience it.

Try it free

Free to start · No credit card needed