Models Are Getting Dumber on Purpose
Large language models are increasingly being optimized to store less factual knowledge in their weights and instead rely on tools, search, and external knowledge bases, raising questions about how to balance “reasoning engines” with up‑to‑date information. Commenters debate whether specialized, pluggable expert models are desirable or even compatible with current architectures, versus the historical trend that large, general models with broad training tend to outperform hand‑crafted specializations (“the bitter lesson”). Others criticize the referenced article as outdated and likely AI-generated, and stress that moving facts out of weights does not, by itself, solve hallucinations or guarantee trustworthy citations.
Benchmark validity and article freshness
- Several comments say the article leans on outdated benchmarks (SimpleQA stopping in 2025) and mischaracterizes current leaders and capabilities.
- Links are provided to updated SimpleQA-Verified graphs showing better performance and to specific example logs.
- Multiple posters claim the blog post itself is likely AI-generated and out of date, which undermines its authority on “models getting dumber.”
Knowledge vs reasoning separation
- Many like the conceptual split between “reasoning engine” and “knowledge in tools/KBs,” citing small reasoning models (e.g., Needle, VibeThinker) as early examples.
- Others argue reasoning is tightly entangled with world knowledge and language use; you can’t cleanly strip “facts” without degrading reasoning.
- Some raise philosophical objections: reasoning is always embedded in specific language practices.
Specialized / pluggable models vs general-purpose
- A long subthread explores a “pluggable knowledge base” dream: small core reasoning + swappable domain modules (Swift, GIS, frontend, etc.), ideally running locally.
- Critics respond that current transformer LLMs don’t compose this way; cross-domain data (e.g., multiple languages) often improves performance due to shared abstract structures.
- MoE is frequently misunderstood: commenters explain that “experts” are not neat domain modules like “Swift expert,” but abstract pattern subnets.
- The “Bitter Lesson” is invoked: hand-crafted modular architectures are likely to be outperformed by large, general models plus simple prompting/agent harnesses.
Hallucinations, tool use, and search quality
- Several note that moving facts out of weights into tools/RAG does not automatically “solve hallucinations”; models can still fabricate or mis-handle retrieved data.
- Reliable “I don’t know” behavior is seen as crucial and still unsolved/unclear.
- Some report that modern chatbots hallucinate less because of aggressive tool-calling and prompts to verify via web search.
- There is concern that dependence on web search is fragile as public search quality degrades, but curated/domain-specific KBs and internal indices may mitigate this.
Use patterns and agentic systems
- Posters debate real-world usage: many non-developers use LLMs for learning, text processing, and drafting, not coding or heavy agents.
- Others argue token-heavy coding/agent use may dominate compute consumption even if user counts skew elsewhere.
- Multi-agent systems and specialized harnesses are proposed as today’s practical path to “modular intelligence,” rather than rewiring model weights.
Intelligence, knowledge, and trust
- Some stress that LLMs are statistical sequence models, not “intelligent” in a human sense; others note that intelligence itself is ill-defined.
- Several worry that AI-written, lightly-checked essays erode reader trust and force readers to perform their own QA on reasoning and facts.