AI smart downgrade: The business strategy and efficiency paradox behind it

The GPT-4o, released by OpenAI in May 2024, claims to be the "smartest", but users quickly discovered that it performed worse than the GPT-4 Turbo in mathematical reasoning, code generation and long text understanding. It's not a bug, it's a feature-this is the first time that an AI model has been blatantly downgraded and blatantly called "optimization."

This is a bit crazy. The common sense of technological development is getting better and better, but AI is doing the opposite. Behind this is not technical limitations, but business ledgers.

downgrade is not intelligence, it is marginal cost

The intelligence level and reasoning cost of the AI model are a U-shaped curve. Early models (GPT-2) had low intelligence and low cost; as the number of parameters increased (GPT-3 → GPT-4), intelligence increased, but reasoning costs also soared. By the GPT-4 stage, the cost of each reasoning is so high that it puts pressure on OpenAI: the 2023 Q4 financial report shows that OpenAI loses US$54 million per day, mainly due to reasoning costs.

As a result, OpenAI began to make a trade-off between "intelligence" and "cost." At the launch conference of GPT-4o, Sam Altman emphasized "faster and cheaper" but kept silent on the level of intelligence. User testing found that GPT-4o performed 10-15% worse on complex tasks than GPT-4 Turbo, but the reasoning speed was twice faster and the cost was 50%. What is downgraded is not intelligence, but marginal cost.

Behind this U-shaped curve is the business logic: when model intelligence is improved to a certain extent, the marginal return decreases, but the marginal cost increases. The intelligence level of GPT-4 is enough to cover 90% of user needs, and the remaining 10% of needs (such as scientific research and complex code generation) require huge costs. OpenAI's choice is: sacrifice this 10% of high-end demand in exchange for a cost reduction for 90% of users.

This is not a technical issue, it is a business strategy.

Behind the downgrade of ## : Who is paying?

The downgrade of AI models is not indiscriminate. OpenAI refines different user needs through subscription model (ChatGPT Plus) and API pricing:

1. Free users : Use the lowest smart version, the cost is extremely low, but the experience is average.
2. Paid users ($20/month): Use the medium smart version, the cost is moderate, and the experience is good.
3. Enterprise users (API calls): Select different intelligence levels according to their needs and pay per volume.

Behind this tiered pricing is the "28 rule": 20% of high-value users contribute 80% of revenue. OpenAI focuses resources on high-value users by downgrading the experience of free and low-end users. This gameplay is commonplace in the SaaS industry, but it is the first time in the AI field that it has been so blatant.

What's more interesting is that OpenAI also uses "model distillation" technology to "compress" the capabilities of high-intelligence models into low-intelligence models. For example, training data for GPT-4o comes from the output of GPT-4, which means that the intelligence level of GPT-4o is actually a "reduced version" of GPT-4. This technology allows OpenAI to quickly launch new models without increasing costs.

Downgrading is not a technological setback, but the ultimate in commercial efficiency.

Steelman: Opposing Views

Some people will say: The downgrade of AI models is inevitable due to technological development, not a business strategy. As the number of model parameters increases, reasoning costs will inevitably increase, and downgrade is to control costs and ensure sustainable development. Moreover, GPT-4o's reasoning speed is faster and the user experience is better, and the lost intelligence of degradation can be made up for by speed improvement.

This view makes sense, but it ignores two key points:

1. Selectivity of downgrades : OpenAI is not an indiscriminate downgrade, but a refined downgrade for different user groups. Free users experienced the most severe experience downgrade, followed by paying users, and corporate users were almost unaffected. This is obviously not a technical limitation, but a business strategy. Transparency of 2. ** downgrade : OpenAI emphasized "faster, cheaper" at the press conference, but kept silent about the decline in intelligence levels. Users only notice that the model performance deteriorates after actual use. This information asymmetry makes the downgrade more like an "hidden price increase"-users pay the same price but get worse service.*

Downgrading is not a technical issue, it is a business ethical issue.

QKPFX8 The Cost of QK Efficiencism

Behind the downgrade of the AI model is the ultimate manifestation of efficiency: the pursuit of maximizing business efficiency and the neglect of user experience and technical ethics. This kind of efficiency can lead to commercial success in the short term, but may damage user trust and technological development in the long term.

Efficiencism has three costs:

1. Loss of user trust : Users find that the performance of the AI model has deteriorated and will doubt the reliability of the product. In the long run, this will damage brand trust.
2. Technical development is stagnant : Downgrading means that AI models no longer pursue the ultimate of intelligence, but pursue the ultimate of business efficiency. This will lead to a slowdown in technological innovation and long-term damage to industry development.
3. Social efficiency decline : The downgrade of AI models will affect industries that rely on AI (such as programming, design, scientific research, etc.), resulting in a decline in overall social efficiency.

Efficiencyism is not a panacea, it has a price.

Scenario Narration: Lao Zhang in the Data Center

Lao Zhang has worked in the data center of a cloud computing company for 10 years. He is responsible for maintaining the server and ensuring smooth reasoning for AI models. Recently, the company launched a new model that claims to be "faster and cheaper." Lao Zhang found that the model often made mistakes when handling complex tasks, but the company asked him to prioritize the operation of the model because it had lower costs and had more customers.

Lao Zhang is puzzled: Why not prioritize the operation of high-intelligence models? The company's explanation is that the cost of high-intelligence models is too high and customers are unwilling to pay. Although the new model is poor in intelligence, it has low costs, many customers and higher overall revenue.

Lao Zhang sighed: The company pursues commercial efficiency, not technical perfection. He began to wonder whether his years of hard work were really meaningful.

Cross-border Analogy: AI downgrade and fast food culture

The downgrade of the AI model reminds me of fast food culture. Fast food sacrifices the quality and taste of ingredients in pursuit of efficiency and cost. Consumers accepted this sacrifice for convenience and cheapness. But in the long run, fast food culture has led to health problems and taste degradation.

The same is true for the downgrade of AI models. In pursuit of business efficiency, intelligence is sacrificed. Users accepted this sacrifice for cheapness and convenience. But in the long run, AI downgrades may lead to technological stagnation and reduced social efficiency.

The ultimate of efficiency is to sacrifice quality.

Ending: The Paradox of downgrade

Behind the downgrade of AI models is the victory of business strategy, but it is also the paradox of efficiency: the pursuit of maximizing efficiency may ultimately sacrifice efficiency itself. OpenAI controls costs and improves business efficiency through downgrades, but may damage user trust and technology development in the long run.

Demotion is not the end, it is the starting point.

How will AI models evolve in the future? Should we continue to downgrade and pursue commercial efficiency? Or return to the ultimate in technology and pursue intelligence? The answer to this question will determine the future of the AI industry.

Golden sentence:

1. "AI downgrade is not intelligence, but marginal cost. "
2. "The ultimate of efficiency is to sacrifice quality. "