Overview of Linear Digressions: Reasoning Models and Why They’re More Than “Fancy Autocomplete”
In this episode, Ben Jaffe and Katie Malone explore reasoning models—LLMs that generate an intermediate “chain of thought” before answering—and explain why they were seen as a major leap beyond standard next-token prediction. Using DeepSeek R1 as the main case study, they break down how reasoning models are trained, why their behavior is surprising, and why the story has technical, economic, and geopolitical implications.
What Are Reasoning Models?
Reasoning models are designed to produce:
- an intermediate thinking trace or chain of thought
- a final answer based on that reasoning
The hosts contrast them with traditional LLMs, which are often described as predicting the next token or acting like “fancy autocomplete.” Reasoning models appeared to do something qualitatively different: solve multi-step problems that require logic, decomposition, and validation.
Why They Felt Like a Breakthrough
Examples discussed include:
- Math problems, especially with large numbers that aren’t likely memorized
- Cipher/encoding puzzles, where the model must infer a pattern step by step
These tasks showed that LLMs could do more than recall patterns—they could appear to reason through unfamiliar problems.
How Reasoning Models Are Built
The episode emphasizes that there is no hidden magical component that makes reasoning models fundamentally different from regular LLMs. Instead, reasoning seems to emerge from combining familiar ingredients in a specific training pipeline.
1) Start with a Pretrained LLM
DeepSeek began with a standard pretrained model, DeepSeek V3, before alignment. This served as the base model for generating candidate reasoning outputs.
2) Create “Cold Start” Reasoning Data
The model was prompted on verifiable tasks such as:
- math
- logic puzzles
- other questions with a known correct answer
It was instructed to:
- put its thinking inside a
<thinking>tag - then provide a final answer
At first, the system graded mostly on:
- whether the format was correct
- whether the final answer was right
Importantly, it did not need the model’s reasoning trace to be “true” or even fully faithful—just useful enough to correlate with correct answers.
3) Use Reinforcement Learning
The outputs from the first stage were used as training data for reinforcement learning. The model was rewarded for:
- correct answers
- proper formatting
- better reasoning traces
The hosts note that some outputs were manually curated so that the next training stage wouldn’t be polluted with garbage.
4) Mix in General Knowledge
To ensure the model could do more than just solve puzzles, training also included some general chat / knowledge examples. The final mixture was roughly:
- 75% reasoning Q&A
- 25% general chat
A Key Insight: The Reasoning Trace Isn’t Necessarily “True”
One of the episode’s most important points is that the chain of thought shown to users may not be the exact internal reasoning the model used to arrive at the answer.
Why That Matters
The hosts stress that:
- the reasoning trace and the final answer can drift apart
- the model may learn to produce a useful-looking explanation without that explanation being the real causal mechanism
- nevertheless, this setup still improves performance on many tasks
This is why reasoning models can feel both impressive and a little mysterious: they produce something that looks like human reasoning, but the system isn’t guaranteed to “think” the way humans do.
Why DeepSeek R1 Stood Out
DeepSeek R1 was notable for several reasons:
- It was very capable
- It was released as open weights, so users could download and run it locally
- DeepSeek published detailed methodology, including a paper in Nature
That openness helped researchers and practitioners better understand how a reasoning model might be trained, unlike earlier closed models from OpenAI.
Distillation
The episode also discusses distillation in two senses:
- A larger reasoning model distilled into smaller models for release
- Possible controversy over whether DeepSeek’s base model, DeepSeek V3, may itself have been distilled from other frontier models
This second point is presented as an allegation, not a settled fact.
Cost, Skepticism, and the Bigger Picture
DeepSeek claimed the model cost only about $5–6 million to build, which the hosts treat skeptically given:
- the usual cost of frontier models
- the scale of compute required
- the possibility that some training work or bootstrap advantages were not fully reflected in the number
They also note that these debates are hard to verify conclusively and overlap with issues of:
- model fingerprinting
- data provenance
- IP and distillation disputes
- international competition in AI
Main Takeaways
1. Reasoning models are not just marketing
They genuinely improved on tasks requiring multi-step logic, not just token prediction.
2. Their success comes from training structure, not a mysterious new mechanism
A mix of:
- supervised examples
- verifiable tasks
- reinforcement learning
- careful curation
can produce reasoning-like behavior.
3. The chain of thought is useful, but not necessarily faithful
The visible reasoning trace may help performance without being a perfect record of the model’s internal process.
4. DeepSeek R1 helped make reasoning models mainstream
Its open-weight release and public methodology made it especially influential.
5. The story is also political and economic
Questions about cost, distillation, and openness aren’t just technical—they reflect broader competition between labs and countries.
Notable Closing Point
The hosts end by noting that reasoning models have become a standard frontier capability, but the fact that one of the earliest and most influential deep dives came from a Chinese lab rather than OpenAI adds a geopolitical dimension to the story.
Practical Implications
For listeners using AI tools, the episode suggests:
- reasoning models are especially useful for coding, math, logic, and structured problem-solving
- the “thinking” output can be helpful, but shouldn’t be treated as fully trustworthy explanation
- open-weight models like DeepSeek R1 offer a chance to experiment locally, though at a performance and speed cost