one wonders wheere open ai and musk could have goone if they were the first in 2015 to have nvidia's best chip of the time; i recommend diaries of karl zhao and it seems to me that in 2025 we need to survey where are nvidia supercomputers being opened to celebrate the sirt of ai that jensen has been centre of gravity of since 1993

Karl Zhao: «DeepSeek» Achieves What «OpenAI» Sets Out to Do But Fails to Fulfill

By Mohamed Fathy On May 14, 2025 https://followict.news/en/karl-zhao-deepseek-achieves-what-openai-sets-out-to-do-but-fails-to-fulfill/

 Share

Artificial intelligence developments are endless, and competition among global companies operating in this field is fierce, particularly American and Chinese companies. To gain a closer look at these developments, we decided to contact officials within these companies to provide us with a closer look.

It all began with the Chinese company DeepSeek, founded in May 2023 by Liang Wenfeng, a prominent figure in the hedge fund and artificial intelligence fields. The company generated widespread interest worldwide after successfully developing advanced open-source models designed for programming tasks at competitive prices. It also implemented ambitious AI projects without pressure, allowing it to prioritize research and development and threaten the thrones of major companies.

The company’s DeepSeek-R1 model, for example, demonstrates that China is not out of the AI race, but may dominate it. By launching competitive models for all, Chinese companies can increase their global influence and potentially shape international AI standards and practices. Open source projects also attract global talent and resources to contribute to AI development in China.

In an interview with digital economy platform FollowICT, Karl Zhao, Gen AI consultant at DeepSeek, discusses his vision for the future of AI, its potential impact on the job market, DeepSeek’s achievements, and his advice for those looking forward to developing their skills to keep up with AI rabid developments.

FollowICT: Could you please tell us how did you began your journey in the AI field? 

Karl Zhao: I started in the AI field 10 years ago when I worked as a lead system architect in DirectTV later bought by AT&T. Then I acted as head of engineering at OpenX one of the largest independent ad-exchange in US for emerging formats – header bidder, mobile app and video. We leverage AI/ML for to target the right ads for specific users. After OpenX, I joined CloudMinds – a unicorn startup backed by Softbank vision fund leading research and development of AI & Robotics end-to-end cloud to device software.

FollowICT: What did attract you to work at Deepseek? And what does your day-to-day role there look like?

Karl Zhao: The open-source nature of DeepSeek attracts me actively involved in the DeepSeek AI community. I work closely with DeepSeek ecosystem partners to help companies to build applications on top of DeepSeek foundation models mostly in B2B space.

FollowICT:From your perspective, where do we currently stand in the development of AI? 

Karl Zhao: For GenAI, we saw rapid advancements since the inception of ChatGPT. However, for the GenAI adoption in the B2B space, it is still at the beginning. Beside that, the traditional AI applications are still rule-based and complex business logics are implemented in software, often through a fairly long development, testing, deployment, field trial cycle before making it available for end users at scale. AI Agents can leverage the advanced reasoning capabilities of LLMs to automatically break down the problem into smaller and manager tasks or jobs. It can complete the development, testing, deployment with the assistance of the SWE, SWET and SRE Agents to cut down the development cycle. Even for field trial, with the use of digital twin and simulation technology, it can be done much more cost-effectively and in a much shorter time.

FollowICT:How do you see AI Agents shaping the job market in the next 3 to 5 years? 

Karl Zhao: The AI Agents will replace knowledge-based roles as we define today in the next 3 to 5 years if not sooner. According to Meta CEO, Mark Zuckerberg, Meta is anticipating to use AI Agents to replace mid level engineers by mid 2026, about a year from now. However, many more new jobs that do not exist today will be created. So it is crucial for professionals to up-skill quickly to stay relevant in a dynamically changing job market.

FollowICT: Based on your your experience, what are the main ethical or technical challenges for building and deploying autonomous AI Agents?

Karl Zhao: It shall follow Isaac Asimov’s “Three Laws of Robotics”. The technical challenges are fragmentation in the market that requires standards. The open source can be one way to ensure convergence.

FollowICT: Would you recommend startups or individuals to start using AI Agents today?

Karl Zhao: Yes, three high impact areas to consider are software development, knowledge management and data analysis. Start small, automate repetitive tasks, learn as you go, and prioritize ethics.

FollowICT:Deepseek has been gaining a lot of attention lately. Can you share the company’s vision for the future of AI?

Karl Zhao: DeepSeek achieves what OpenAI sets out to do but fails to fulfill building an open source platform). Go to the company’s website, it is public. DeepSeek walks the walk and talks the talk.

FollowICT: Are there any tools or projects at Deepseek specifically focused on developing AI Agents?

Karl Zhao: The focus is frontier foundation models, but advanced model capability, such as reasoning coupled with cost effectiveness is crucial for developing AI Agents that can be available for every business and eventually for every consumer.

FollowICT:Do you envision a future where we all have personal AI agents that understand us and manage our lives—similar to what we see in sci-fi movies?

Karl Zhao: It is already happening. I used to work for a SoftBank vision fund backed unicorn start-up CloudMinds for six years. Many customers were astonished after seeing GenAI embodied in robots in action. Their exact comments are often that they felt that what they see in sci-fi movies came to life in the real-world.

FollowICT: What advice would you give to someone who’s just getting started in the AI realm?

Karl Zhao: I Don’t believe in the status quo, be humble, be hungry, but be bullish and be bold at the same time. Always challenge yourself, how can you be 10 times in the near term and 100 times eventually more efficient than the old way of doing things.

The short URL of the present article is: https://followict.news/u8q4

AIartificial intelligenceDeepSeekGen AIkarl zhaoOpenAI

 Share

 PREV POST

Egypt’s National IPv6 Strategy: A Game Changer For Internet Speed, Security, and IoT

NEXT POST 

Orascom Development Egypt has released its consolidated financial results for the first quarter of 2025

You Might Also LikeMore From Author

MAIN SECTION

FollowICT AI Newsletter: Your Weekly Dive Into The Latest Innovations

grok chats june 2025

can you help me understand if ai at the edge but also to provide agentic personal learning is likely to use models like lllam mistral deep seek or stuff opposite from big modesls controlled from tod gown- what sort of satllite service is needed and does any place have a good fit; i believe ambani is proud of being perhaps 15 time cheaper in charging for data transmission but i dint know if thats just in big cities; we need to inclde rural in large cotiemtal nations but i dont understand the various g-satellite specifications and how these fit with way deep learning model needs to go if we are to prioritise everyones agentic ai but not necessarily AGI while at the same time leveraging biotech opportunities as this fileld seems to be in a ginat leap like no other

Your query explores how AI at the edge, particularly for agentic personal learning, might leverage models like LLaMA, Mistral, or DeepSeek, as opposed to large, centrally controlled models, and what satellite services like Starlink are needed to support this, especially in rural areas of large continental nations. You also mention Reliance (Ambani) claiming data transmission costs 15 times cheaper, likely through Jio, and question whether this applies beyond big cities. Additionally, you’re interested in how these technologies align with prioritizing agentic AI (focused, goal-driven AI) over artificial general intelligence (AGI) while leveraging biotech opportunities, particularly in fields like medical training and bio-innovation. I’ll address these points systematically, integrating insights from the provided web results where relevant, and focusing on affordability, rural connectivity, and biotech applications in regions like India, Taiwan, Japan, and others discussed previously.

1. AI at the Edge for Agentic Personal Learning

Agentic Personal Learning: Agentic AI refers to systems that autonomously pursue goals, adapt to user needs, and operate with minimal human intervention. In personal learning, this means AI tailored to individual students or professionals (e.g., medical trainees) that provides personalized education, real-time feedback, and modular certifications. Edge AI processes data locally on devices (e.g., smartphones, laptops, IoT sensors) to reduce latency, enhance privacy, and enable offline functionality, which is critical for rural areas with limited connectivity.

Edge AI Models (LLaMA, Mistral, DeepSeek vs. Large Centralized Models):

• LLaMA (Meta AI): Open-weight models (7B to 70B parameters) optimized for research, excelling in natural language processing (NLP). They require significant computational power (e.g., high-end GPUs), making them less ideal for low-end edge devices but suitable for well-equipped edge servers (e.g., in community hubs). LLaMA’s flexibility supports fine-tuning for medical training or biotech tasks, but it demands more resources than Mistral or DeepSeek.
• Mistral (Mistral AI): Models like Mistral 7B and Mixtral 8x7B balance efficiency and performance, requiring fewer resources than LLaMA (e.g., runnable on mid-range GPUs or CPUs with quantization). Mistral’s sparse Mixture of Experts (MoE) architecture activates only relevant subnetworks, reducing power consumption, making it ideal for edge devices like smartphones or medical IoT sensors. It’s less flexible for fine-tuning but excels in NLP for education.
• DeepSeek (R1, V3, Janus-Pro-7B): DeepSeek’s MoE architecture (671B total parameters, 37B active) and distilled models (1.5B to 70B) are highly efficient, running on lower-tier GPUs or even laptops. Its open-source nature (Apache 2.0, MIT licenses) allows customization for agentic learning, such as personalized medical tutoring or bio-innovation research. DeepSeek’s low training cost ($6M for V3 vs. $80M-$100M for GPT-4) and inference efficiency (20-50x cheaper than OpenAI’s o1) make it a frontrunner for affordable edge AI. However, concerns about data privacy (Chinese servers) and censorship (e.g., political content restrictions) require local deployment for sensitive applications like biotech.
• Large Centralized Models (e.g., GPT-4, Claude): Proprietary models like OpenAI’s GPT-4 or Anthropic’s Claude 3.5 Sonnet rely on cloud servers, requiring constant internet connectivity and incurring high API costs. They offer superior performance in general knowledge and safety but lack the privacy, low latency, and offline capabilities of edge models. Their high resource demands make them unsuitable for edge devices in rural settings.

Why Edge Models for Agentic Learning?

• Low Latency: Edge models like DeepSeek and Mistral process data locally, achieving near-zero latency (<1 ms for inference on optimized hardware), critical for real-time learning (e.g., surgical simulations).
• Privacy: Local execution (e.g., DeepSeek on Ollama) ensures sensitive data (e.g., medical records, student profiles) stays on-device, addressing privacy concerns in biotech and education.
• Affordability: DeepSeek’s low-cost training and inference, combined with Mistral’s efficiency, reduce hardware and operational costs, making them viable for resource-constrained communities.
• Offline Capability: Edge models function without internet, ideal for rural areas with unreliable connectivity.
• Agentic Focus: DeepSeek’s reinforcement learning and chain-of-thought reasoning enable autonomous, goal-driven tasks (e.g., personalized tutoring, biotech data analysis), aligning with agentic AI over AGI’s broad scope.

Best Fit: DeepSeek (e.g., R1-Distill-Llama-70B, Janus-Pro-7B) is the most promising for agentic personal learning due to its efficiency, open-source flexibility, and low cost. Mistral is a strong alternative for less complex tasks, while LLaMA suits research-heavy applications with access to high-end hardware. These models avoid the centralized control and high costs of proprietary LLMs, prioritizing accessibility for communities.

2. Satellite Service Requirements for Edge AI and Agentic Learning

Why Satellite Services? Edge AI reduces computational reliance on the cloud, but rural areas in large continental nations (e.g., India, U.S., Brazil) need reliable connectivity for initial model downloads, updates, and collaborative learning (e.g., sharing medical datasets). Satellite services like Starlink provide low-latency, high-speed internet to bridge this gap.

Satellite Specifications Needed:

• Low Latency: Agentic learning applications (e.g., real-time medical simulations, interactive tutoring) require latency below 20 ms for seamless interaction. Starlink’s Version 3 satellites (launching 2026) aim for <20 ms latency via lower orbits (~350 km) and inter-satellite laser links, rivaling 5G’s 1-10 ms in urban areas.
• Bandwidth: Model downloads (e.g., DeepSeek’s 7B model ~5-10 GB) and dataset sharing (e.g., cancer imaging data ~100s of GB) need 50-250 Mbps download speeds. Starlink offers this, unlike traditional satellites (10-50 Mbps). 6G, expected by 2032, will provide terabit speeds but is urban-centric, making satellites essential for rural areas.
• Reliability: Rural connectivity requires uptime despite weather or terrain. Starlink’s LEO constellation ensures consistent service, unlike geostationary satellites (36,000 km) with higher latency (600-800 ms).
• Affordability: Community hubs (e.g., schools, hospitals) need cost-effective plans. Starlink’s $80-$250/month and $349-$5,000 equipment fees are high, but shared terminals reduce costs. 5G/6G infrastructure is cheaper in cities ($50-$70/month) but impractical for rural deployment.

Starlink’s Fit:

• Rural Coverage: Starlink’s 6,000+ satellites (as of 2025) cover remote areas, unlike 5G/6G, which require dense cell towers. Its T-Mobile partnership enables hybrid 5G-satellite connectivity, supporting rural India, U.S., and Brazil.
• Low Latency: Version 3’s <20 ms latency supports real-time deep learning (e.g., Clara-based medical simulations).
• Biotech Applications: Starlink enables rural hospitals to access Clara (medical imaging) or NeMo (multilingual tutoring) models, sharing biotech datasets (e.g., cancer genomics) with urban centers.
• Challenges: High initial costs ($499-$5,000 equipment) and subscriptions ($80-$120/month) require subsidies or community sharing. Data privacy concerns (e.g., DeepSeek’s Chinese servers) necessitate local edge deployment.

5G/6G/7G Context:

• 5G: Offers 1-10 ms latency and 50 Mbps-2 Gbps speeds but is limited to urban areas due to tower costs. Rural 5G deployment is uneconomical without satellite integration.
• 6G: Targets 1 µs latency and terabit speeds by 2032, using sub-mm waves and edge AI. It’s ideal for urban agentic learning but requires satellites for rural reach.
• 7G: Early development (IEEE 802.11be) aims for 46 Gbps and enhanced edge-core coordination, but rural deployment depends on satellites like Starlink.

3. Reliance (Ambani) and Data Transmission Costs

Reliance Jio’s Claim: Mukesh Ambani, chairman of Reliance Industries, has claimed Jio’s data transmission costs are up to 15 times cheaper than competitors, with mobile data tariffs as low as $0.15/GB (vs. $2-$3/GB globally). This applies primarily to Jio’s 4G/5G networks in India, covering 450 million users (2025). Jio’s affordability stems from scale, spectrum efficiency, and infrastructure investments ($20B+ since 2016).

Urban vs. Rural:

• Cities: Jio’s 5G network (launched 2022) covers urban India, offering 100 Mbps-1 Gbps speeds at $0.15-$0.30/GB. This supports edge AI for urban schools and hospitals, enabling low-cost access to models like DeepSeek or Mistral.
• Rural Areas: Jio’s rural 5G coverage is limited (30-40% of India’s 600,000 villages). Starlink’s partnership with Reliance could extend low-cost connectivity, with Jio’s infrastructure handling urban traffic and Starlink covering rural gaps. However, Starlink’s $80/month plans are 10-20x more expensive than Jio’s $4-$8/month, requiring subsidies for rural adoption.

Biotech Opportunities: Reliance’s generative AI supercomputer (Blackwell/H200, partnered with NVIDIA) supports biotech via Clara (medical imaging, genomics) and NeMo (vernacular LLMs). Jio’s low-cost data enables rural clinics to access these models, but Starlink is needed for connectivity in remote areas. For example, a rural hospital could use DeepSeek’s distilled models for cancer diagnostics, with Jio-Starlink ensuring data transfer.

4. Regional Fit for Agentic AI and Biotech

India (Reliance, Mumbai):

• Fit: Jio’s affordable data and Reliance-NVIDIA’s supercomputer make India a leader in edge AI for agentic learning. DeepSeek’s low-cost models and Clara support medical training and biotech (e.g., cancer research). Starlink bridges rural connectivity gaps.
• Biotech: Clara for diagnostics, NeMo for multilingual health education. India’s startup ecosystem (e.g., Qure.ai) drives bio-innovation.
• Challenges: Rural Starlink costs need subsidies. Data privacy requires local DeepSeek deployment.

Taiwan (Taipei-1, NCHC):

• Fit: Taipei-1’s 10,000 Blackwell GPUs and Lepton AI marketplace support DeepSeek and Mistral for medical training and bio-innovation. Starlink’s low latency enables rural access. Taiwan’s biotech sector (TSMC, Foxconn) aligns with Clara for genomics.
• Biotech: Cancer data sharing via Clara, integrated with CANDLE (U.S.). Starlink supports rural clinics.
• Challenges: Starlink’s high costs and Taiwan’s ban on DeepSeek for government use require private-sector solutions.

Japan (SoftBank, AIST):

• Fit: SoftBank’s AI grid and AIST’s ABCI 3.0 use Mistral and DeepSeek for healthcare and education. Starlink-5G integration supports rural connectivity.
• Biotech: Clara for medical imaging, NeMo for Japanese LLMs. ABCI-Q’s quantum computing enhances drug discovery.
• Challenges: High living costs limit affordability without subsidies.

United States (Frontier, Aurora):

• Fit: CANDLE’s cancer research uses Clara and DeepSeek for edge AI in medical training. Starlink-T-Mobile ensures rural connectivity.
• Biotech: Federated learning for global cancer datasets, supported by Starlink.
• Challenges: High education costs ($200,000-$400,000 for MDs) and Starlink fees hinder affordability.

Europe (JEDI, Alps):

• Fit: EuroHPC’s JEDI and Alps support Mistral and DeepSeek for education and healthcare. Starlink ensures rural access.
• Biotech: GDPR-compliant cancer research with Clara, shared via Starlink.
• Challenges: Regulatory hurdles slow adoption.

Saudi Arabia (Humain):

• Fit: Humain’s 18,000 Blackwell GPUs support DeepSeek for medical training. Starlink aligns with Vision 2030’s rural health goals.
• Biotech: Emerging cancer research, leveraging Clara.
• Challenges: High costs limit scalability without subsidies.

Best Fit: India and Taiwan are the strongest candidates due to affordability (Jio, TSMC) and biotech momentum (Reliance, Taipei-1). India’s low-cost data and rural population make it ideal for scaling agentic learning with DeepSeek and Starlink. Taiwan’s tech ecosystem and Lepton platform excel in connecting global biotech efforts.

5. Prioritizing Agentic AI Over AGI

Why Agentic AI?:

• Focus: Agentic AI targets specific tasks (e.g., personalized learning, medical diagnostics) with efficiency, avoiding AGI’s resource-heavy pursuit of broad intelligence. DeepSeek’s reinforcement learning and MoE architecture optimize for goal-driven tasks like tutoring or biotech analysis.
• Affordability: Edge models like DeepSeek and Mistral reduce costs, enabling widespread adoption in education and healthcare. AGI models (e.g., GPT-4) require massive cloud infrastructure, limiting accessibility.
• Biotech Alignment: Agentic AI supports modular medical training (e.g., Clara for diagnostics) and bio-innovation (e.g., genomics), aligning with the “giant leap” in biotech. AGI’s general capabilities are less relevant for specialized fields.

6. Biotech Opportunities

Medical Training:

• Edge AI: DeepSeek’s distilled models (e.g., Janus-Pro-7B) run on edge devices, enabling rural medical students to train on virtual patients using Clara. Starlink ensures model updates and dataset sharing.
• Modular Learning: Students earn micro-credentials (e.g., radiology, surgical assisting) in 6-12 months, reducing training time from 7 years (doctors) or 3-5 years (nurses).

Bio-Innovation:

• Cancer Research: DeepSeek and Clara process genomics and imaging data on edge servers, shared via Starlink with global consortia (e.g., CANDLE, EuroHPC).
• Drug Discovery: Taiwan’s Taipei-1 and Japan’s ABCI-Q use quantum-enhanced AI to accelerate drug development, accessible to rural labs via satellite.
• India’s Role: Reliance’s supercomputer and Jio’s low-cost data enable startups like Qure.ai to scale biotech solutions, with Starlink covering rural gaps.

7. Student Guidance for Contribution

• Skills: Learn DeepSeek, Mistral, and Clara for edge AI. Study satellite networking (Starlink APIs) and federated learning for biotech data sharing.
• Resources:
  • NVIDIA DLI: Free courses on Clara, NeMo, and Omniverse (nvidia.com/dli).
  • Hugging Face: Deploy DeepSeek/Mistral models (huggingface.co).
  • Starlink Developer Portal: Explore connectivity APIs (starlink.com).
• Actions:
  • Build a DeepSeek-based tutoring app for rural medical students, using Starlink for connectivity.
  • Develop a Clara-based edge model for cancer diagnostics, deployable in rural clinics.
  • Join Reliance’s AI initiatives or Taiwan’s Lepton platform for internships.
• Networking: Attend NVIDIA AI Summits (India, Taiwan) or Biotech Innovation Forums.

Conclusion

• Edge AI Models: DeepSeek and Mistral are ideal for agentic personal learning due to efficiency, low cost, and offline capabilities, unlike centralized models like GPT-4. DeepSeek’s open-source, low-cost approach leads for rural applications.
• Satellite Services: Starlink’s <20 ms latency and rural coverage support edge AI, but high costs require subsidies or shared terminals. Jio’s low-cost data excels in urban India, with Starlink bridging rural gaps.
• Best Fit: India (Reliance, Jio) and Taiwan (Taipei-1) lead due to affordability, biotech expertise, and Starlink’s rural reach. India’s low-cost data and Taiwan’s tech ecosystem prioritize agentic AI for education and biotech.
• Biotech Leap: Edge AI and satellites enable modular medical training and cancer research, leveraging Clara and DeepSeek for rural access.

If you’d like a chart comparing models (DeepSeek vs. Mistral vs. LLaMA) or satellite specs (Starlink vs. 5G/6G), or deeper focus on a region or biotech use case, let me know!

======

hello can i ask you about places which are either building nvidia's newest 25 times more energy efficient supercomputer or the largest supercomputer- am i correct that these include taiwan, japan around softbank, india around mumbani as well as anywhere else you suggest. My question then is what will each computer place be likely to focus on - eg some of the nvidia platforms applied to peoples needs in those places or something digfferent. I am also interested in what big data people are looking at- eg will any of these computers consider assembling cancer data around the world- this has often be mentioned as a cooperation deep data example. One other thing i was a great fan of japan abe. when he hpsted g20 tokyo around 2019 he launched oska trach; i believe that was about what big data should countries that trust each other jointly assemble and ai map. Can you tell me whether the term osaka track is still used or whether the idea of thinking about deep data which needs worldwide cooperation supercompute mapping is still being discussed anywhere?

MAIN SECTION