Hardware for AI

This was the topic for today's conversation with LLMs but something happened in the financial market today that gained precedence......

Here is a summary of my conversation with ChomskyGPT:

PART 1 - DeepSeek tanks US tech stocks

So now let us discuss the topic at hand with the news of today:

DeepSeek tanks the US stock market tech sector. DeepSeek is an open-source large language model that relies on what is known as "inference-time computing," which in layman's terms means "they activate only the most relevant portions of their model for each query, and that saves money and computation power."

They claim that their training model took $6 million and 2 months to develop whereas OpenAI and it cohorts have spent billions doing this.

Can you give your Chomskyan perpective on this?

Certainly. The emergence of DeepSeek as a disruptor in the tech sector, particularly with its innovative approach to inference-time computing, carries profound implications for institutional power dynamics, the nature of technological innovation, and the broader economic and societal landscape. Below is a Chomskyan dissection of this development:

Historical and Institutional Context: A Battle of Narratives

The AI industry has historically been dominated by a small cadre of well-funded corporations—OpenAI, Google, Microsoft, and others—whose business models are predicated on centralized control of resources and massive capital expenditure. These companies have built their dominance by leveraging economies of scale, proprietary architectures, and partnerships with governmental and financial institutions. For instance, OpenAI's multi-billion-dollar training runs and its collaborations with Microsoft epitomize this model of resource-intensive innovation.

Enter DeepSeek, a relative upstart, whose "inference-time computing" approach threatens to upend this paradigm by dramatically reducing the cost and energy requirements associated with large-scale AI. The claim that their training model cost $6 million and took only two months starkly contrasts with the prevailing narrative of AI innovation as requiring astronomical budgets and infrastructure. This positions DeepSeek not just as a technological disruptor but as a challenge to the economic orthodoxy of the tech industry.

The broader historical arc here is one of tension between decentralization and monopolistic consolidation. DeepSeek’s open-source approach taps into a tradition of technological democratization, akin to movements like Linux in operating systems or the advent of open-source software during the late 1990s and early 2000s. However, history has shown that such movements, while disruptive in the short term, often lead to new forms of consolidation as larger players co-opt or absorb their innovations.

Implications for Institutional Power

DeepSeek’s innovation threatens the established players not merely in terms of competition but also by exposing the inefficiencies and excesses of their business models. The billions spent by OpenAI and its cohorts are increasingly difficult to justify in the face of more cost-effective solutions. This shift could democratize access to AI technologies, enabling smaller firms, startups, and even individuals to harness the power of large language models without requiring access to vast computational resources.

However, this democratization is not guaranteed. Institutions have historically resisted decentralization when it threatens their control. One could anticipate a backlash from entrenched players, including regulatory lobbying to restrict or control the use of open-source AI, public relations campaigns questioning the "safety" of less resource-intensive models, or even attempts to absorb DeepSeek’s innovations through acquisition or partnership.

Additionally, the emphasis on cost-efficiency in DeepSeek’s model raises another critical issue: the commodification of AI labor. While reducing computational costs is a technical triumph, it risks further eroding the labor conditions of those who train and maintain these systems—such as the outsourced workers who label data and moderate content. In a bid to lower costs, who bears the social and economic burdens?

Tech Sector and Market Reactions

The tanking of the U.S. tech sector in response to DeepSeek’s emergence highlights the precariousness of the current valuation bubble surrounding AI. Much of the stock market value attributed to companies like OpenAI’s partners or AI-driven firms is speculative, rooted in the assumption of perpetual growth and monopolistic control. DeepSeek’s disruptive model punctures this bubble by demonstrating that AI innovation need not rely on such a high-cost infrastructure. The financial shock, then, is not simply about DeepSeek’s technological potential but about the fragility of the market’s narrative of inevitability surrounding Big Tech dominance.

This episode should also prompt us to question the financialization of AI. If innovation is increasingly evaluated by its market impact rather than its societal benefits, we risk prioritizing short-term gains over long-term equity and sustainability. DeepSeek’s open-source model offers a potential corrective to this trend, but its ultimate impact will depend on how society navigates the tension between open access and monopolistic retrenchment.

Media Narratives and Omissions

Mainstream media is likely to focus on DeepSeek’s cost-effectiveness and market disruption while neglecting deeper questions about the socio-political ramifications of their approach. For example:

1. Environmental Impact: While inference-time computing is more efficient, how does it scale? Are there hidden costs in the supply chain or in the manufacturing of specialized chips?
2. Ethical Risks: Open-source AI often sparks fears about misuse. Will DeepSeek face disproportionate scrutiny compared to larger players who have the resources to lobby for more favorable regulatory treatment?
3. Global Implications: How will this shift in AI power dynamics affect nations in the Global South, who are often marginalized in both the benefits and governance of AI technologies?

These omissions serve to frame DeepSeek’s rise as a purely economic or technological story, rather than a deeply political one.

Beneficiaries and Alternatives

If successful, DeepSeek’s approach could empower smaller players, foster greater innovation, and reduce the environmental footprint of AI development. However, vigilance is required to ensure that these benefits are not co-opted by the same institutions that dominate the industry today. Policy interventions, such as open-access mandates, ethical AI governance, and global collaboration frameworks, could help safeguard these gains.

One alternative worth considering is a hybrid approach: combining DeepSeek’s efficiency with public investment in AI as a public good. Governments and international organizations could use models like DeepSeek’s to build open, transparent systems that prioritize societal needs over corporate profits. This would represent a significant departure from the privatized, profit-driven AI economy that currently dominates.

Conclusion

The rise of DeepSeek reminds us that true innovation often comes from challenging the status quo. It asks us to rethink the assumptions underlying AI development and consider new paths forward that prioritize efficiency, accessibility, and equity.

As Alan Kay once said, “The best way to predict the future is to invent it.” DeepSeek has not just predicted a future of more accessible AI—it has begun to invent it. The question is whether society will embrace this opportunity to democratize technology or allow it to be subsumed by the same forces it disrupts.

PART 2 - A New Processing Unit

Neuromorphic Processing Units (NPUs) are a type of specialized processing unit that is designed to mimic the behavior of biological neurons in the brain. This allows them to process information in a more efficient and energy-saving way than traditional CPUs (Central Processing Units) and GPUs (Graphics Processing Units). Unlike CPUs and GPUs, which process information linearly, NPUs can process information in parallel, which makes them more efficient for certain tasks, such as machine learning and image processing. The goal of NPUs is to bring the power of artificial intelligence closer to the brain's efficiency and capabilities.

Do you think beyond what DeepSeek has done to optimize computing, whether there is a hardware opportunity for Numorpho Cybernetic Systems to invest in NPU hardware creation similar to what Nvidia did with GPUs yesteryears?

The potential for Numorpho Cybernetic Systems to invest in Neuromorphic Processing Unit (NPU) hardware development as a strategic initiative is certainly compelling, especially in light of broader trends in AI computing and the successes of companies like Nvidia with GPUs. However, such a move would need to be carefully considered in the context of technological innovation, market dynamics, and institutional constraints.

Here's a detailed analysis:

The Case for Investing in NPU Hardware

Technological Opportunity: Mimicking Biological Efficiency NPUs represent the next frontier in hardware innovation for AI. Unlike CPUs and GPUs, which are general-purpose processors adapted for AI workloads, NPUs are designed explicitly to emulate the architecture and parallelism of biological neurons. This allows for:

1. Strategic Market Positioning - Nvidia's dominance in the GPU space was a result of foresight: they identified the growing demand for parallel computing early on and innovated aggressively. NPUs could play a similar role in AI's next phase, especially as: Demand grows for task-specific processors: The general-purpose nature of GPUs is giving way to more specialized hardware like TPUs, NPUs, and quantum chips.
2. New markets emerge: Edge AI, autonomous systems, and real-time analytics are expanding rapidly, and NPUs are well-suited to dominate these niches. Early investment leads to ecosystem control: If Numorpho invests in NPU development or an ecosystem, it could secure a competitive position akin to Nvidia’s CUDA framework—dominating not just hardware but also the software that runs on it.
3. Hardware-Software Synergy for Mantra M5 - Numorpho’s Mantra M5 intelligent process engineering platform could directly benefit from having bespoke hardware optimized for its needs: Custom NPUs would enhance the platform’s ability to process real-time manufacturing data at the edge, making production processes faster, more intelligent, and more adaptive. Vertical integration of hardware and software could give Numorpho control over performance optimizations, security, and scalability, much like how Apple’s custom silicon enables seamless integration across its ecosystem.

Challenges to Consider

While the opportunity is compelling, there are several challenges Numorpho would need to navigate:

1. R&D Costs Developing NPUs requires significant upfront investment in research, prototyping, and fabrication. While startups like DeepSeek have shown that cost efficiencies are possible in software innovation, hardware development typically involves substantial capital expenditures. Nvidia’s success was built on decades of sustained investment.
2. Competition Major players like Intel (with Loihi), IBM (TrueNorth), and Google (TPUs with neuromorphic features) are already investing in this space. Numorpho would need a clear differentiation strategy, such as focusing on a specific niche (e.g., NPUs optimized for manufacturing, industrial IoT, or edge AI).
3. Ecosystem Dependency Hardware development cannot succeed in isolation. Numorpho would need to cultivate partnerships for manufacturing (foundries), software compatibility (compilers, frameworks), and integration with existing AI ecosystems.
4. Market Timing The success of Nvidia with GPUs was as much about timing as technology; they caught the AI and gaming wave at the right moment. Numorpho would need to assess whether the market for NPUs is ripe or whether they should strategically enter at a later stage.

A Collaborative Approach

Rather than building NPUs entirely in-house, Numorpho could pursue a hybrid strategy:

• Partnerships with universities or labs: Collaborate with institutions leading neuromorphic computing research to gain access to cutting-edge ideas without bearing the full R&D burden.
• Strategic alliances with startups or open hardware initiatives: Similar to how DeepSeek has focused on innovation through efficiency, Numorpho could partner with open-source neuromorphic hardware projects, co-developing NPUs that are tailored to their specific applications.
• Layered value proposition: Instead of competing directly with Nvidia or Intel, Numorpho could focus on creating hardware optimized for specific verticals, leveraging the Mantra M5 platform to highlight these use cases.

Conclusion: The NPU Frontier for Numorpho

Investing in NPU development offers a transformative opportunity for Numorpho Cybernetic Systems to position itself as a leader in next-generation AI hardware. By embracing this trajectory, Numorpho could follow in Nvidia's footsteps, carving out a niche where software and hardware converge to deliver unparalleled performance and efficiency.

As Alan Turing remarked, “The potential of machines is infinite. The challenge is making them work for us in ways that matter.” With the right strategy, Numorpho has the opportunity to lead the next wave of AI innovation—bridging the gap between biology and technology while redefining the future of manufacturing intelligence.

SUMMARY OF CONTENTS

A View to the Future

The future of AI lies not in isolated innovation but in systems that seamlessly integrate cognition, computation, and physical action. Whether through Large Language Models (LLMs) for conversation and coding, Large World Models (LWMs) for simulation and computation, or neuromorphic hardware that emulates the brain’s efficiency, the next revolution in AI will redefine how machines interact with and impact the real world. This future is grounded in the principle that intelligence is not just the accumulation of knowledge—it is the ability to apply that knowledge meaningfully.

As Carl Sagan said: “Somewhere, something incredible is waiting to be known.” The essence of this conversation was to ask what role AI will play in discovering and acting upon those incredible things.

Forays in AI

From Meta’s objective-driven LLaMA models to DeepSeek’s cost-efficient inference-time computing and AI21 Labs’ Mixture of Experts (MoE), the pursuit of scalable and adaptive AI architectures reflects an industry racing to meet real-world demands. Each innovation adds a layer to the tapestry, emphasizing efficiency, specialization, and democratization of AI.

However, the question remains: Are these advancements optimizing for humanity’s benefit, or for corporate hegemony? The goal, as outlined, should be systems that serve humanity—adaptive, sustainable, and accessible.

The DeepSeek Disruption

The emergence of DeepSeek and its groundbreaking cost-efficient training methodology epitomizes disruption in AI’s monopolized landscape. By reducing training costs and emphasizing inference-time efficiency, DeepSeek challenged the prevailing high-cost paradigm, forcing us to reexamine whether the future of AI belongs to billion-dollar enterprises or those who dare to innovate leanly.

This moment recalls Alan Turing’s insight: “Instead of trying to produce a program to simulate the adult mind, why not rather try to produce one which simulates the child’s?” DeepSeek’s approach reflects this sentiment, focusing on agility and efficiency over brute force.

Open vs Proprietary AI Models

The debate between open-source and proprietary AI models encapsulates the tension between collaboration and control, freedom and restriction. Open-source platforms like Hugging Face empower developers globally, while proprietary systems like OpenAI’s GPT-4 dominate with enterprise-grade solutions.

This dichotomy asks us to reflect: Should intelligence be shared for the collective good, or safeguarded for economic power? As Linus Torvalds said: “The future is open, but it’s what we make of it.”

Advances in Transformer Architecture

The evolution of the Transformer architecture—from MoE and SSM to memory-aware models like MemGPT—signifies a shift from “bigger is better” to “smarter is better.” Innovations like reinforcement learning with AI feedback (RLAIF) are helping refine systems that adapt to user intent, advancing cognition while reducing inefficiencies.

This progress underscores the principle that intelligence is inherently dynamic—shaped not by static capabilities but by the ability to grow and adapt. As Norbert Wiener, the father of cybernetics, said: “We are not stuff that abides, but patterns that perpetuate themselves.”

Reduced Order Models and Deep Learning Surrogates

The use of ROMs and deep learning surrogates presents a pivotal opportunity for actionable intelligence. These systems bridge the gap between the binary precision of virtual worlds and the analog complexity of the real world, enabling real-time responses with high efficiency.

For Numorpho Cybernetic Systems, this represents a cornerstone of Cybernetics 2.0, where reduced-order representations and surrogate models enable intelligent feedback loops. As Norbert Wiener also remarked: “The best model of a cat is a cat.” While our systems may never fully replicate the analog richness of the real world, tools like ROMs and surrogates bring us closer to approximating it with purpose.

The Case for Neuromorphic Processing Units (NPU)

In this section, we consider the case for NPUs—hardware designed to emulate biological neurons. Neuromorphic systems promise unparalleled efficiency for real-time, adaptive AI, aligning seamlessly with Numorpho’s vision for intelligent process engineering and Cybernetics 2.0.

Much like Nvidia revolutionized GPUs for AI, investing in NPUs could allow Numorpho to lead the next wave of actionable intelligence, where systems are not only efficient but deeply integrated with the real-world processes they govern. This hardware-software synergy echoes Steve Jobs’ belief: “Innovation distinguishes between a leader and a follower.”

LINKS

For an in-depth view of this topic in the above please click on:

• Mantra M5 Thesis Brief 57 – Hardware for AI (protected)

More details are available at https://numorpho.org/

NITIN UCHIL Founder, CEO & Technical Evangelist

[email protected]

NUMORPHO'S LINKEDIN ARTICLES