Generic selectors
Exact matches only
Search in title
Search in content
Post Type Selectors
About the Author
About the Author
About the Author

Accelerated Innovation-AI in AI – DeepSeek: A Modern David Slingshots AI Goliaths, Redefining the Art of the (Im) Possible

Being Human for a Better Tomorrow in the Age of AI
  • January 24, 2025
  • 32nd edition
Accelerated Innovation AI in AI DeepSeek A Modern David Slingshots AI Goliaths Redefining the Art of the Im Possible

Context:

2024 has seen a early majority of enterprises begin the adoption journey for AI in the enterprise . The statistics are telling when you compare enterprise AI adoption in 2023 vs 2024:

  • In 2023, 55% of companies had integrated AI into at least one business function. By 2024, this number jumped to 72%, representing a substantial leap in adoption (Source: Statista)
  • In 2023, only 18% of organizations had deployed AI in production, with 24% engaged in pilot projects. By 2024, nearly a third of businesses were running AI in production, and the share of organizations with mature AI initiatives doubled from 4% to 8% (Source: Enterprise Strategy Group)
  • 92% of respondents reported an increase in AI use in their organization over the previous 12 months (Source: Enterprise Strategy Group)
  • Organizations are pursuing multiple AI experiments, with most running 20 or fewer experiments or POC and over two-thirds said that 30% or fewer of their experiments will be fully scaled in the next three to six months. (Source: Deloitte)

What’s Amiss

In the 14th edition of the Mindvista newsletter we spoke about Bi-Modal Exploration (AI Enhanced and AI Driven) where
  • AI Enhanced: Permeate and enhance key existing processes with AI with a diffusion by infusion model to eventually cover all key processes.
  • AI Driven: Start now for long-term fence and form model to create new AI-driven experiences

Yet most of the conversations and POCs and pilots are not about new disruptive innovation but about saving costs by bringing labor and process efficiencies (AI Enhanced) . Statements from CEOs freezing new hires due to AI efficiencies have added fuel to the debate.

While no one will argue against improved efficiency, focusing solely on incremental betterment misses the opportunity to think anew, use AI for innovation, and drive grounded disruption. For example to make the AI Driven new experiences the 19th edition of the Mindvista newsletter illustrated new ideas for Apple (Digital Assistant like Ask Jeeves); Marriott (Virtual Bonvoy) and LinkedIn (Delphi for hiring).

While these were ideas, in future editions, we will explore both conceptual and real-world disruptive innovations using AI across industries. And what better way to start this series than by addressing disruption within the field of AI itself.

Disrupting the disruptors: DeepSeek slingshots AI frontier models

We all know how much funding, business and policymaking attention is on AI. Billions and hundreds of billions of dollars are being invested by venture capitalists, corporate funding with full support from policy makers.

Yet quietly on January 20, 2025 , DeepSeek a Chinese company released a model for public use and the AI world is stunned by its elegant design, performant engineering and stunning hardware and inference efficiency as seen below

DeepSeek what did they do differently

Wiithout going into the technicalities and to keep it simple for readers this is what DeepSeek did.

In an industry dominated by the idea that “more is better,” Deepseek broke free of the collective mindset by questioning assumptions, rethinking design, and engineering for efficiency.

  • 1. Rethinking the Design: Instead of relying on brute force—throwing more data and compute at the problem—they asked, “What’s a better way to achieve superior outcomes?” This bold question led them to reimagine how reasoning and learning interact.
  • 2. A New Approach to Training: While training methods like supervised fine-tuning and reinforcement learning remain industry standards, DeepSeek introduced a combinatorial innovation. By processing reasoning outputs first and applying fine-tuning afterward, they optimized both accuracy and resource usage.
  • 3. Elegant Engineering: DeepSeek’s model demonstrated exceptional efficiency by: Capturing long-range dependencies, which are crucial for complex reasoning tasks. Streamlining memory usage and optimizing compute cycles, dramatically reducing training costs.
  • Capturing long-range dependencies, which are crucial for complex reasoning tasks.
  • Streamlining memory usage and optimizing compute cycles, dramatically reducing training costs.
  • Competitive language and coding benchmarks, outpacing Meta’s Llama and rivaling OpenAI’s GPT-4.
  • 4. Efficient training: trained a frontier AI model at one-tenth the cost and time compared to industry giants. The entire training cost under $6 million—a fraction of what others spend—and delivered results that rival the most advanced models in the field.
  • 5. Results: While the entire community is abuzz and testing the results (it has just been 5 days), the early indicators and comparison of DeepSeek to frontier models is very impressive (Source Deeplearning)
  • a. Dominated 5 out of 7 coding benchmarks.
  • b. Shows superior performance in math & reasoning: (AIME, MATH benchmarks widely regarded as the gold standard for reasoning and problem-solving). It did a perfect run in 45 seconds where humans spend 90 minutes on a paper to get to top maths courses at elite universities such as Oxford and Cambridge.
  • c. Competitive language and coding benchmarks, outpacing Meta’s Llama and rivaling OpenAI’s GPT-4.
  • 6. A commercial implementation models of DeepSeek suggest 27x cheaper cost for input and output token costs.

DeepSeek questioned the status quo, breaking free from the ‘bigger is better’ mindset. Their approach reimagined how reasoning and learning interact, resulting in a model that’s not only performant but also cost-effective. By focusing on reasoning-first training, elegant engineering, and efficient memory use, they achieved what seemed impossible: a state-of-the-art AI model trained at a fraction of the usual cost.

There is no doubt we are in a rabbit hole moment with the DeepSeek model questioning the design, engineering and price /performance that is a seismic break from industry group think.

On hindsight it may seem obvious but that is so until some one sees it. As Jean Piaget said “The challenge is not seeing different things but seeing things differently on what everyone sees”.

DeepSeek- how did they do this differently

Nothing DeepSeek did was conventional that you would see in any team or resources thrown in a large enterprise.
1. Unconventional Team:
  • DeepSeek’s founders weren’t from the typical AI elite. With a background in quantitative finance (not computer science), they brought fresh perspectives to AI design. Their team was entirely local, comprising self-taught engineers driven by passion, not pedigrees. This approach prioritized ingenuity and resourcefulness over credentials.
2. Purpose-Driven Openness:
  • Unlike traditional players, DeepSeek openly shared its model ((671 billion, with 37 billion activated per forward pass) releasing it as a MIT license open-source for the global community and published the technical approach (arxiv) . This open-source approach democratizes access to cutting-edge AI tools.
3. First Principles Thinking:
  • They approached AI design with the simplicity of first principles, questioning every assumption and focusing on solving real-world problems with measurable outcomes.This pragmatism allowed them to design a model that excelled in efficiency without compromising performance.
4. Innovation Under Constraints:
  • Working with limited resources—a fraction of the funding and compute power of AI giants—they turned constraints into a competitive advantage.
5. Focused Execution:
  • Their approach was results-oriented, targeting tangible, measurable outcomes. They avoided feature bloat and speculative experiments, focusing on delivering real-world impact.

What DeepSeek has done is truly inspiring. They have defied all checklist criteria for success and if it plays out as promises we have a new breakthrough way for AI not only to be better but far more efficient and sustainable.

A commercial implementation at the data center or at the edge in devices can give absolutely new art of the possible use cases for AI driven innovation.

We will keep a close watch out on how this shapes and congratulations to Deepseek team for the courage and enterprise!

Takeaways

We live in a world where the art of possible is getting democratized. Doing incremental is one and being disruptive is another. Both are paths to pursue.

Any one, any where can do any thing if they can show the courage , intelligence and tenacity.

The innovation zone lies directly in what you know and what skills you have.

AI is a critical foundational technology for humanity and access is critical.

Policy makers must rethink their priorities and mindset even in ensuring what is built is accessible and relevant to the daily lives of citizens.

As Dr APJ Abdul Kalam, former President of India, said- “Small ambition is a crime” and there is no room for small ambition on technology.

I hope these narratives inspire and trigger a question- What problem can we solve?

What a promising start to 2025—the first year of the next quarter-century.

Explore and stay tuned for more!

Cheers

“AI is a language. Treat it like one: practice, iterate, and mind your grammar prompts, assumptions, and verification.”

II. Questioning / Asking

Good conversations flow from well‑sequenced questions—topical, simple, coherent, cohesive.
LLM Conversation Example 1
Q: What are empirical judgments?

A: Empirical judgments are based on observation, experience, or experimentation.
Q: What are moral judgments?

A: Moral judgments are based on ethical principles and values.
Share: