AI · Topic 2 of 2
DeepSeek, AI Applications and the Future of AI
Learn DeepSeek-V3 and DeepSeek-R1, Mixture of Experts, MLA and reasoning models, where AI is applied, and the privacy, bias and safety disputes around AI.
- 5 min
- DeepSeek-V3 is an MoE model with MLA; only selected experts run for each token.
- DeepSeek-R1 is a reasoning model trained with large-scale RL on DeepSeek-V3-Base.
- R1-Zero used RL without SFT; R1 added cold-start data to fix readability.
What it is
DeepSeek is a Chinese AI company whose open-weight models, DeepSeek-V3 and DeepSeek-R1, showed that efficient architectures and reinforcement learning can deliver top-level results. They pushed the industry toward open models and lower training and inference cost.
AI applications are the places AI creates value, from phones to factories. Disputes are the ethical, legal and social questions that AI raises as it spreads.
How it works
DeepSeek-V3: an efficient Mixture of Experts model
DeepSeek-V3 is a Mixture of Experts (MoE) language model. Its technical report gives 671B total parameters, with 37B activated for each token.
In an MoE layer, a router (gating network) sends each token to a few expert sub-networks instead of the whole layer. Only the chosen experts compute, so the model has a huge capacity but a much smaller cost per token.
| DeepSeek-V3 technique | What it does |
|---|---|
| DeepSeekMoE | Its MoE architecture of many experts with routing |
| Multi-head Latent Attention (MLA) | Compresses keys and values into a small latent vector, shrinking the KV cache and speeding inference |
| Auxiliary-loss-free load balancing | Keeps experts evenly used without an extra balancing loss term |
| Multi-token prediction (MTP) | A training objective that predicts more than one future token |
| FP8 mixed precision training | Uses 8-bit floating point to cut training memory and compute |
DeepSeek-R1: a reasoning model
DeepSeek-R1 is a reasoning model built on DeepSeek-V3-Base. It produces a long chain of thought before giving the final answer, which helps in math, code and logic.
| Model | How it was trained | Notes |
|---|---|---|
| DeepSeek-R1-Zero | Large-scale reinforcement learning applied directly to the base model, with no supervised fine tuning first | Reasoning emerged from RL alone, but output had repetition, poor readability and language mixing |
| DeepSeek-R1 | Cold-start data before RL, then further training stages | Fixed readability issues; strong reasoning |
| Distilled models | R1 outputs used to fine tune smaller Qwen2.5 and Llama3 based models | Brings reasoning to small, cheap models |
The R1 repository and weights are released under the MIT License, which allows commercial use and distillation.
DeepSeek's influence on AI development
- Open weights at the top level: anyone can download, run and fine tune a leading model.
- Efficiency over brute force: MoE, MLA and FP8 show that architecture and engineering can reduce cost.
- RL for reasoning: R1-Zero showed that reasoning can be learned largely through reinforcement learning.
- Distillation: small models can inherit reasoning from a large teacher, which helps edge and private deployment.
AI applications
| Area | Example applications |
|---|---|
| Smart city and transport | Traffic flow prediction, license plate recognition, public safety video analysis |
| Healthcare | Medical image analysis, drug discovery, health monitoring |
| Finance | Fraud detection, credit scoring, intelligent customer service |
| Manufacturing | Visual quality inspection, predictive maintenance |
| Retail and Internet | Recommendation, search, intelligent marketing |
| Consumer devices | Voice assistants, photo enhancement, smart homes |
| Autonomous driving | Perception, planning and driver assistance |
| Office and coding | LLM writing assistants, code generation, AI agents |
Most of these combine CV, speech and NLP, and many now use foundation models fine tuned for the industry.
Disputes around AI
| Dispute | The concern |
|---|---|
| Privacy | Models are trained on, and can reveal, personal data |
| Bias and fairness | Biased training data produces unfair decisions, for example in hiring or lending |
| Deepfakes and misinformation | Generated images, video and voice make "seeing is believing" unreliable |
| Accountability | Who is responsible when an autonomous car or medical AI makes a harmful decision |
| Explainability | Deep models are "black boxes", hard to audit |
| Jobs | Automation replaces some tasks and changes many roles |
| Security and misuse | Adversarial attacks, prompt injection, malicious use |
| Copyright | Use of protected works in training data and outputs |
The future of AI
The direction is toward more general foundation models, multimodal models, AI agents that plan and use tools, and cheaper inference on edge devices. Governments are adding AI rules, and the industry is investing in trustworthy AI: fairness, transparency, robustness and privacy protection.
A worked example
A hospital wants a chatbot that answers staff questions using internal guidelines, without sending data outside. It downloads a distilled open-weight reasoning model and runs it on its own servers.
This uses DeepSeek's influence (open weights and distillation) for an AI application (healthcare). It must still handle disputes: patient privacy, wrong answers and who is accountable for advice.
How the exam asks about this
- "Which architecture lets DeepSeek-V3 activate only part of its parameters for each token?" → Mixture of Experts.
- "How was DeepSeek-R1-Zero trained?" → With large-scale reinforcement learning on the base model, without supervised fine tuning first.
- "Which AI dispute is raised by fake but realistic videos of real people?" → Deepfakes and misinformation.
What to remember for the exam
- DeepSeek-V3 is an MoE model with MLA; only selected experts run for each token.
- DeepSeek-R1 is a reasoning model trained with large-scale RL on DeepSeek-V3-Base.
- R1-Zero used RL without SFT; R1 added cold-start data to fix readability.
- DeepSeek released open weights, and distilled its reasoning into smaller Qwen and Llama based models.
- Main disputes: privacy, bias, deepfakes, accountability, explainability, jobs and security.
In an MoE model the total parameter count is not the compute per token. Only the activated experts run, so cost follows activated parameters.
Common mistakes
Thinking MoE means several separate models voting. MoE experts are sub-networks inside one model's layers. A router picks a few of them for each token.
Assuming reasoning models need human-written reasoning steps. R1-Zero showed reasoning behavior can emerge from reinforcement learning with rewards, without supervised fine tuning first.
Treating AI ethics as only a technical issue. Bias can be reduced technically, but accountability, copyright and jobs also need law, policy and governance.
Check yourself
What does Multi-head Latent Attention improve in DeepSeek-V3?
It compresses keys and values into a small latent vector. This shrinks the KV cache and makes inference cheaper.
Why did DeepSeek train R1 with cold-start data instead of RL alone?
R1-Zero, trained by RL alone, showed repetition, poor readability and language mixing. Cold-start data before RL fixed these issues.
Name two disputes AI raises in hiring or lending decisions.
Bias and fairness, because biased data can produce unfair outcomes, and explainability, because deep models are hard to audit.
Topic 2 of 2 in AI
Test yourself on AI
The full Huawei HCIA-AI H13-311 question bank has 600 questions, with timed practice exams and an explanation for every answer.
This study guide is independent and is not affiliated with, endorsed by, or sponsored by Huawei. Vendor names, certification names and exam codes are trademarks of their respective owners and are used for identification only.