New DeepSeek Research - The Future Is Here!
TL;DR
DeepSeek released an 80-page paper providing what may be the first fully open, reproducible recipe for creating ChatGPT-level AI. Five key techniques — from teacher-free training to distillation — together produce small, cheap models that dramatically outperform older frontier models, and the entire framework is free and public. ---
Key Concepts
GRPO (Group Relative Policy Optimization)
tap to reveal ↩
Training method where the model generates multiple answers to one question and the answers are graded against each other — no second "teacher" model required
Reinforcement Learning from scratch
tap to reveal ↩
Training AI purely by self-play against rules, with no human examples provided
Cold start / flashlight nudge
tap to reveal ↩
Providing just a few examples at the start of RL training to stabilize language and direction without constraining reasoning ability
Distillation
tap to reveal ↩
Using a large, capable model to generate a training dataset (800K examples) that teaches smaller models to replicate its reasoning style
Emergent "aha" moment
tap to reveal ↩
AI spontaneously learning to pause, reconsider, and think longer — without being explicitly taught to do so
Notes
§The Openness Problem with Existing Labs
- OpenAI's GPT-4 paper explicitly states it contains "no further details about the architecture, hardware, training, compute, dataset construction, or training method"
- DeepSeek's new paper is 80 pages (up from 20 pages a year prior) and contains substantive, reproducible methodology
- Author frames this as a meaningful step toward science being open and reproducible
§Technique 1 — Group Relative Policy Optimization (GRPO)
- Traditional training uses a second equally large AI as a "teacher" to grade every sentence → expensive and slow
- GRPO eliminates the teacher model entirely
- Instead: model generates 16 different answers to one question
- Answers are graded against each other using objective criteria (e.g., did the code run? was the answer correct?)
- Best answer is rewarded; poor answers are discarded
- Cheap enough to run at massive scale
§Technique 2 — Emergent Thinking (The "Aha Moment")
- Model was observed spontaneously generating phrases like "wait" or "let me recalculate"
- Over time it learned that longer thinking → higher scores, so it extended its own reasoning without being told to
- No human explicitly taught this behavior — it emerged from the training process
- First documented case of an AI naturally learning to think before responding
§Technique 3 — Pure Reinforcement Learning (No Human Examples)
- Analogy: learning chess by playing millions of games vs. reading a textbook — self-play has no ceiling
- DeepSeek proved you can train reasoning ability with zero human demonstration examples
- Model started at ~15% success rate on competition math; reached ~80% on its own
- Discovered novel strategies not present in any training data
§Technique 4 — The Flashlight Nudge (Cold Start)
- Starting from zero is possible but causes instability: gibberish output, random language switching
- Providing a small number of examples at the start fixes direction immediately
- Effect on math: minimal improvement (only ~2 percentage points, sometimes negative) — math is language-agnostic
- Effect on natural language tasks (AlpacaEval): more than tripled performance — language coherence matters here
- Takeaway: a small guided start is high value for language quality, low cost to reasoning quality
§Technique 5 — Distillation (Learn from Giants)
- DeepSeek R1 (large model) was used to generate 800,000 reasoning examples — essentially a textbook
- Small, cheap models are then trained on this textbook
- Result: a 7-billion parameter model beats the previous GPT-4o on competition math by nearly 6×
- 7B models can run on many laptops today; likely on phones within a few years
- The predecessor required billions of dollars to train; the distilled version is free
Actionable Takeaways
- Generate multiple options before committing — produce 5+ solutions to a problem and grade them against each other before picking one
- Pause deliberately when facing a hard question — say "wait" out loud, double-check logic; the extra time consistently pays off
- Prioritize practice over theory — learn fundamentals minimally, then do the task and self-correct through failure rather than consuming endless tutorials
Quotes Worth Keeping
“
"Given the competitive landscape, this report contains no further details about the architecture, hardware, training, compute, dataset construction, or training method." — OpenAI, GPT-4 technical report
“
It started generating words like 'wait' or 'let me recalculate.' And over time it realizes that spending more time thinking leads to a higher score. So it started thinking longer and longer by itself.
“
You need the genius to write it, but not to read it.
“
We are not maximizing money here. We are maximizing meaning.