Back to list
Why AI Models are Getting Dumber on Purpose: The Strategic Trade-off Between Reasoning and Factual Knowledge
Industry NewsArtificial IntelligenceLarge Language ModelsAI Benchmarks

Why AI Models are Getting Dumber on Purpose: The Strategic Trade-off Between Reasoning and Factual Knowledge

A significant shift is occurring in the AI industry where model developers are intentionally prioritizing reasoning capabilities over factual recall. While new models like GLM-5.2 and Qwen3.5 are achieving record-breaking scores on mathematical benchmarks such as AIME with a fraction of the parameters used by older models like GPT-4, their ability to remember specific facts is declining. Research indicates that factual knowledge requires substantial parameter space—approximately two bits per parameter—leading labs to sacrifice "world knowledge" to fit advanced logic into smaller, more efficient architectures. This trend results in high hallucination rates for niche facts in smaller models, even as they dominate in coding and logic tasks, signaling a move toward reasoning-centric AI design.

Hacker News

Key Takeaways

  • Reasoning Efficiency is Skyrocketing: New models are achieving higher scores on math benchmarks (AIME) while using significantly fewer active parameters per token compared to 2023-era models.
  • The Factual Knowledge Sacrifice: There is a deliberate trade-off where labs are removing world knowledge to make room for reasoning skills within limited parameter budgets.
  • High Hallucination Rates in Small Models: Despite their logic, smaller models (under 10B parameters) show hallucination rates as high as 80-82% on factual knowledge benchmarks.
  • The Cost of Facts: Research suggests that factual knowledge occupies roughly two bits of space per parameter, making "encyclopedic" models increasingly expensive to maintain.
  • Benchmark Divergence: Performance on math and coding benchmarks is improving at an absurd rate, while factual recall (SimpleQA) remains a major hurdle for even the largest models.

In-Depth Analysis

The Paradox of Reasoning vs. Recall

The landscape of Large Language Models (LLMs) in 2026 reveals a striking divergence: models are becoming exponentially better at thinking while simultaneously becoming worse at remembering. The data points provided by Walter van der Giessen highlight this shift clearly. GLM-5.2, for instance, has achieved a staggering 99.2% on the AIME 2026 benchmark with only 40 billion active parameters per token. In contrast, GPT-4, which was rumored to utilize approximately 280 billion active parameters in 2023, struggled to solve even basic AIME problems.

This trend continues down the scale. Qwen3.5, with only 17 billion active parameters, scores 91.3% on the same benchmark, and DeepSeek V4-Flash operates with just 13 billion active parameters. At the smallest end of the spectrum, the Qwen3.5 9B model—which can fit into 6GB of VRAM when quantized—roughly doubles the intelligence index score of any other model under 10 billion parameters. If one were to judge progress solely by math and coding benchmarks, the conclusion would be that AI is advancing at an unprecedented rate. However, this progress in logic comes at a steep price: the loss of factual accuracy.

The Physics of Language Models and the Parameter Tax

The reason for this divergence is rooted in the physical constraints of model architecture. According to the "Physics of Language Models" research series, factual knowledge capacity is limited to approximately two bits of information per parameter. This means that for a model to function as a comprehensive encyclopedia—storing the birth years of minor historical figures, the populations of small municipalities, or the specific argument orders of obscure software packages—it must dedicate a massive number of weights to static data.

AI labs have realized that these "facts" take up valuable space that could otherwise be used for reasoning pathways. Consequently, the trade is now deliberate. By stripping away the need to store trillions of niche facts, developers can optimize the remaining parameters for high-level reasoning and problem-solving. This is why, on the SimpleQA benchmark (which tests factual recall without external tools), even the industry leader, Gemini 2.5 Pro, only manages a 53% accuracy rate. The smaller models, such as Qwen3.5 4B and 9B, fail almost entirely in this regard, with hallucination rates between 80% and 82%. When these models encounter a question about a fact they haven't stored, they do not admit ignorance; instead, they generate a confident but incorrect answer.

Industry Impact

This shift from "encyclopedic" models to "reasoning" models marks a fundamental change in how AI will be integrated into professional workflows. For the AI industry, this implies that the era of the "all-knowing" standalone model may be ending. Instead, we are seeing the rise of highly efficient reasoning engines that are designed to be paired with external data sources, such as Retrieval-Augmented Generation (RAG) or search tools, to compensate for their lack of internal factual memory.

Furthermore, the ability to run high-reasoning models on consumer-grade hardware (like the Qwen3.5 9B fitting in 6GB of VRAM) democratizes access to advanced logic. However, it also places a higher burden of verification on the user. As models get "dumber" regarding world facts on purpose, the industry must pivot toward better grounding mechanisms to prevent the confident hallucinations that now characterize small, high-reasoning models.

Frequently Asked Questions

Question: Why are new AI models hallucinating more if they are supposed to be "smarter"?

While models are getting smarter in terms of logic, math, and coding (reasoning), they are being built with fewer parameters to increase speed and reduce costs. Since facts take up significant parameter space (about 2 bits per parameter), labs are intentionally sacrificing factual memory. This leads to higher hallucination rates when the model is asked about specific facts it no longer has the capacity to store.

Question: How does GLM-5.2 compare to GPT-4 in terms of efficiency?

GLM-5.2 is significantly more efficient. It uses only 40 billion active parameters per token to achieve a 99.2% score on the AIME 2026 math benchmark. In comparison, 2023's GPT-4 used an estimated 280 billion active parameters and had very poor performance on the same types of complex mathematical problems.

Question: What is the current limit of factual recall in top-tier AI models?

Even the best models currently available struggle with pure factual recall. Gemini 2.5 Pro, the current leader on the SimpleQA benchmark, only answers 53% of factual questions correctly when not allowed to use external tools or search engines. Smaller models perform much worse, often hallucinating over 80% of the time on knowledge-based tasks.

Related News

How Google Duplex Evolved to Call Businesses on Your Behalf in Gemini
Industry News

How Google Duplex Evolved to Call Businesses on Your Behalf in Gemini

A Google Research Blog entry titled “How Google Duplex evolved to also call businesses for you in Gemini” describes its subject as the evolution of Google Duplex to support calls to businesses on a user’s behalf in Gemini. The entry is dated November 20, 2026, and the supplied article text consists only of the label “Generative AI.” The headline identifies the capability and connects it to Duplex, but the available text does not explain how the system evolved, how users request a call, or where the feature is available. The headline establishes the topic, but the supplied material does not provide a technical account or rollout details.

How AI and Lean Are Helping Automate Formal Verification to Fight Software Slop
Industry News

How AI and Lean Are Helping Automate Formal Verification to Fight Software Slop

In Latent Space, Richard MacManus examines formal verification as an emerging way to fight what the headline calls “slop.” The central claim is that proving software correct with mathematics has traditionally required specialist work, and that AI and Lean are now helping automate that process. The distinction is important: the reported change concerns how the work of proving correctness is done, not what counts as a proof. The supplied excerpt presents automation as assistance, without saying that it replaces specialists or completes proofs independently. It also does not specify which parts of the process AI and Lean help automate, how much work they save, or which software has been verified this way.

AI and Human Claim Solutions to the 167-Year-Old Riemann Hypothesis
Industry News

AI and Human Claim Solutions to the 167-Year-Old Riemann Hypothesis

TechCrunch AI reports that AI claimed a solution to the Riemann hypothesis, a mathematical problem the headline describes as 167 years old. A human also put forward a competing solution. The report, written by Tim Fernholz and published on November 20, 2026, describes mathematicians as astonished by the AI-driven result. The supplied excerpt does not explain either proposed solution, identify the AI system or the human claimant, or say whether either claim has been verified. The distinction between a claimed solution and a confirmed one is central to reading the report: the excerpt establishes that two claims are being reported, but it does not establish that either has resolved the hypothesis.