My reflection of education & training from the "DeepSeek Moment"

Today's "DeepSeek?moment" on Wall Street sparked an exciting new opportunity, profound impact for AI startups, especially for small and open source based companies in education and training industry or alike.

DeepSeek displaces ChatGPT as the App Store's top app since Sunday Jan 26th after their 2nd appl release since Dec. The public info says: DeepSeek researchers only used ~US$6m to train their open-source LLM whereas chatGPT burnt ~US$300m to train theirs, so DeepSeek is >50x more efficient. For the same query, DeepSeek's cost is ChatGPT's 1.12%. This creates lots of attention, in contrast to the $500b project "Stargate" announcement by Trump.

What is DeepSeek??And what is the implication on learning, education, and training??

// What is DeepSeek? //

DeepSeek is based in Hangzhou China (where Alibaba HQ is). Founder Liang Wenfeng is an electrical engineer and had success in his quantitative hedge fund (High-Flyer). He started to build DeepSeek in 2023 and published "DeepSeek-R1" papers in 2024 to explain its new RL LLMs.

DeepSeek's papers say it was using older mid-ranged H800 GPUs (due to US export restriction of Nvidia chips). It also uses data from Llama and other LLMs to generate "synthetic data" to train its model.

How amazing is DeepSeek? It's free to download to your own hardware, fine-tune it with your own codes and data-set (the power of open-source!). Besides its affordability and transparency, DeepSeek-R1 is as good as OpenAI's o1 models in math and coding (source). Per E! HuffPost, it has achieved a 97% accuracy rate in solving math problems and has outperformed 96% of humans in coding and natural language programming tests.

Worth noting, DeepSeek has yet developed its capability of producing images and videos which chatGPT is capable of.

// Implications on education and training //

3 week ago, I attended the world largest consumer electronic show - CES 2025 in Las Vegas (my 7th) and I also went to one of the largest education trade shows - Bett Global in London last week. Aggregating most of what I saw and learnt in the last six months, the success of DeepSeek-R1 is indeed trending in the same direction.

Synthetic Data Factory and "Skinny" LLMs: Heading into 2025, one big debate is whether AI models have started to plateau amid difficulty associated with sourcing high-quality, human-made training data. This lack of data problem exemplifies in highly privacy-protected environments, like in schools and hospitals. One of the many inspiring technologies demonstrated at CES is that Nvidia launched its Cosmos (digital twin technology and foundation model) by having it watch 20 million hours of video about nature, humans, and anything to do with the physical world. Based on those real scenarios, it can also create synthetic data to create even more scenarios.

You can call DeepSeek is "intelligence on a budget" but its success has demonstrated that startups can leverage open source foundation models to train their SLMs (Small Language Models) to generate synthetic data for specific use cases, while larger organizations can combine synthetic data with their proprietary datasets. I expect there will be more syntenic data factory and tools available to support Ai practitioners to build digital twins and simulations at scale.

Proliferation of simulations and future AI tutors: In a digital twin of education, synthetic data generated can be used?to create realistic virtual student populations, allowing for the simulation of various learning scenarios and the testing of educational interventions which can be limited or privacy-sensitive;?this enables educators to optimize teaching strategies, identify potential issues, and improve learning outcomes by analyzing the simulated data from diverse student profiles generated by the GAI system.?

Training and assessment is an area ripped for AI disruption. For example, in a war combat training, it used to take $2-3m to develop a single-player simulation. Now with AI and metaverse, you can build a multi-player simulation, mixing avatars and real human interactions with one-fifth of the costs.

In academic learning environment, I prefer live human tutoring experience to AI-generated avatars or chatbots. Especially for younger children, the current AI technologies have not be proved to help children develop discipline and empathy. However, creating an AI clone of a human super star teacher with consistent realistic characters has already been possible with AI tools, e.g. https://app.heygen.com/ (for deepfake), https://bolt.new/ and https://www.synflow.com/ (for building a full tutor app with no coding experience at all).

Embed Human Values and Ethics in AI

Last week from World Economic Forum Davos 2025, the most interesting takeaway I have is from a panel discussion (video link). Stanford professor Yejin Choi said something really sparked a lot of thought in me. She wishes the speed of AI development can be slowed down so that we can teach AI human values and ethical norms. She highlighted the limitations of today’s AI, which learns from the internet—often a reflection of harmful societal biases.

The current tech is not perfect but the quality of learning outcomes and affordability will be significantly improved, as DeepSeek has demonstrated to us. We also know that the successful adoption of any new tech is never a linear or hockey stick path. It usually relies on inclusion of existing stakeholders (like teachers and doctors), not to replace them but enabling them or transforming them to different change agents.