July 20, 2026
Bot drama: now thinking in tiers
Controlling Reasoning Effort in LLMs
AI’s new “thinking harder” button has commenters joking it’s just an old trick in fancy packaging
TLDR: AI makers are adding a “think harder” setting to their chatbots, making step-by-step answering a standard feature. Commenters were amused that this sounds a lot like an old user trick turned into an official product, with equal parts skepticism, nerdy hacks, and book recommendations.
The big news is that AI companies are now openly selling models with different “reasoning effort” levels — basically, versions that can spend more or less time working through a problem before answering. The article walks through how this trend grew after OpenAI’s earlier reasoning-focused models and DeepSeek’s splashy research, which showed that you can train an AI to get better at math and code by rewarding correct answers. In plain English: instead of just blurting things out, these systems now produce a visible step-by-step draft of their thinking before the final answer.
But in the comments, the real mood was: “Wait, isn’t this just the old ‘think step by step’ trick with better branding?” That was the sharpest hot take by far, with one reader laughing that the whole movement feels like a polished, official version of a prompting hack people were already using years ago. Another gem from the peanut gallery? A delightfully nerdy flex about controlling the model’s effort by watching for the exact moment the output suggests the bot thinks it’s done thinking — the kind of comment that makes the whole thread feel half lab notebook, half meme factory.
There wasn’t full-on flame war chaos here, but there was definitely a subtle tension between “this is a meaningful advance” and “congrats on productizing a workaround.” Meanwhile, one supportive voice cut through the snark to recommend the author’s book, proving that even in a skeptical thread, someone is always ready to drop the reading list.
Key Points
- •The article argues that reasoning models are now a standard component of modern LLM releases, citing OpenAI’s o1, DeepSeek-R1, and GPT-5.6.
- •It defines a reasoning model in LLM research as a model that outputs an intermediate reasoning trace rather than literally reasoning like a human.
- •The article says reasoning-task performance can be improved through two approaches: training scaling and inference scaling.
- •DeepSeek-R1 is described as using reinforcement learning with verifiable rewards (RLVR) to train reasoning models in verifiable domains such as math and code.
- •The article states that DeepSeek-R1 found output rewards alone sufficient for learning behaviors such as intermediate explanations, backtracking, and self-correction, while reasoning traces were not ultimately used for training updates.