KBKAST Episode 355 Deep Dive: Sam Cummings | Will We See Current LLM Technology Reach Its Limits in 2026?

Date: February 18, 2026
Host: KBI Media
Guest: Sam Cummings, Director of Education at Gen AI Works

Episode Overview

This episode tackles the evolving landscape of Large Language Models (LLMs), their functional and economic limits, and what the future holds for AI innovation—especially in context of cybersecurity. Sam Cummings draws on his background in customer success and AI to offer a strategic perspective, arguing that while current LLM technology may be reaching its architectural and cost efficiency ceiling, an innovation boom in areas like advanced reasoning, "world models," and memory architectures is imminent. The conversation covers impacts on security, business consolidation, startups, the job market, and the changing nature of digital influence.

Key Discussion Points & Insights

LLMs: Functional Nature and Limitations

• LLMs = Talking Models:
  LLMs function primarily as stateless, conversational agents—excellent at pattern matching and text generation, but fundamentally limited in “understanding” and reasoning complexity.

  “Large language models are talking models...it doesn't have an understanding. So...you get these hyperbolic performances where it scales. That’s the big thing that really, from a physics perspective, is underwriting this problem.” — Sam Cummings [00:00]

• Memory and Reasoning Bottlenecks:
  Current LLM architectures struggle with context management. As context windows enlarge, hallucination and cost rise disproportionately.

  "More context doesn’t mean better performance...as the models get larger...they hallucinate more, they get confused..." — Sam Cummings [07:08]

• Cost of Reasoning:
  Running LLMs—particularly for reasoning-heavy, always-on tasks like security monitoring—is currently too expensive for widespread deployment.

  “If I wanted to make more monitoring security services, the amount of data I have to pass through a large language model makes most use cases not viable, cost wise.” — Sam Cummings [04:58]

The Coming Transition: From LLMs to "World Models"

• Stateless vs. Stateful Models:
  The future lies in models that separate reasoning and memory from the underlying architecture—akin to the difference between the CPU and a GPU in computers.

  “When you separate out reasoning and memory from the underlying model, that’s where you’re able to really move the needle.” — Sam Cummings [13:51]

• Alternating Current Analogy:
  LLMs today operate as “direct current,” reasoning in one direction (step by step). The next generation (“stateful engines”) will be to AI what alternating current was to electricity: enabling persistent, long-term, efficient reasoning.

  “What changed our universe was alternating current...I can now reason for long periods of time with lower cost. That’s going to explode the security industry.” — Sam Cummings [16:07]

• Rise of World Models:
  Beyond LLMs, “world models” are on the horizon. These will enable factual grounding and visualization, revolutionizing domains like robotics and manufacturing.

  “Nvidia just announced at CES Cosmos...These world models will be the substrate...that powers smart cars, robots, and mechanized manufacturing.” — Sam Cummings [23:22]

Industry Structure: Monopolization and Startup Dynamics

• Big Tech Will Continue to Dominate:
  Expect consolidation; most AI innovation will be absorbed into the portfolios of hyperscalers (Microsoft, Google, Amazon).

  “We are in the middle of a monopolization era in civilization...across all industries.” — Sam Cummings [23:22]

• Local vs Cloud Token Consumption:
  While big providers want all LLM computation routed through their clouds, Sam predicts a coming shift toward more local, specialized, and cost-effective reasoning models.

  “As we get the innovations we’re talking about, models can run cheaper, they can run locally...the global amount of token consumption is going to be more local...” — Sam Cummings [23:22]

• Startup and Acquisition Conveyor Belt:
  Most startups will aim for acquisition by larger players rather than building independent empires, echoing patterns seen over decades in tech.

  “It’s almost going to be like an ice cream shop where you can get 150 flavors, but it’s the same base.” — Sam Cummings [34:11]

Security Industry: Explosive Opportunity and Risks

• Double-edged Arms Race:
  AI has accelerated both cyber-attack techniques (deep fakes, automated exploits) and defensive capabilities (real-time threat detection, codebase inspection).

  “The amount of arms race increase that language models has done is already known...Order of magnitude increase of deep fakes, just unprecedented capability.” — Sam Cummings [35:51]

• Exponential Growth in Security Startups:
  The cyber sector is facing a surge of startups, many destined for acquisition, potentially resulting in considerable personal financial opportunity for cybersecurity professionals.

  “You could work at six or seven companies over the next five years and retire with huge stock portfolios.” — Sam Cummings [35:51]

Broader Societal Trends: Job Market, Knowledge Economy, and Influence

• Job Market Turbulence:
  The rise of AI lowers the cost of “thinking” work (text, analysis, reasoning). Success will be less about degrees and more about networks, community, and distribution.

  “Your moat is not what you can do, it's who you know and your ability to leverage that to an audience...Culture and community and distribution are where I'd put my focus.” — Sam Cummings [39:51]

• Bots and Information Trust:
  With bots comprising the majority of web traffic, influence and branding will center on authenticity, micro-communities, and engagement rather than mere reach or SEO.

  “We're now entering a world where we're optimizing for what the language model sees.” — Sam Cummings [39:51]

• Wave Cycles in Media:
  Rejection of the mainstream in favor of micro-influencers and authentic pockets of community will drive the next cycle of media and influence.

  “We are currently in a place where the micro influencer is going to be more valuable...People are going to be seeing so much AI generated stuff...they’re going to want something that seems like, oh, I can trust this.” — Sam Cummings [45:22]

Notable Quotes & Memorable Moments

• On Limits of LLMs:

  "The solution of talking your way to understand has its functional limits. What you need instead is a world building model."
  — Sam Cummings [11:00]

• On Tech Monopolies:

  “We are in the middle of a monopolization era in civilization that’s been unmatched in hundreds of years. And that’s important to understand across all industries...”
  — Sam Cummings [23:22]

• On Opportunities for Cybersecurity Pros:

  “Can I say something to your audience directly? Y'all my people, cybersecurity folks, y'all gonna be doing really well. Matter of fact, if y'all...one of you on this call is gonna be a millionaire billionaire coming soon.”
  — Sam Cummings [35:51]

• Advice to Next-Gen Candidates:

  “Your ability to have a malleable mind that’s not anchored on the way we’ve done things is a superpower. Leverage that. Explore these technologies...”
  — Sam Cummings [50:04]

• On Media Trends:

  “We are an industry of waves...the micro influencer is going to be more valuable...because people will want to trust.”
  — Sam Cummings [45:22]

Timestamps for Important Segments

• LLM Functional Limits Explained – [00:00–01:59], [07:08–11:00]
• Cost/Scaling of Reasoning – [04:58], [13:51–15:56]
• Direct vs Alternating Current Analogy – [16:07]
• Consolidation & Monopolization – [23:22–28:19]
• Startups and Acquisition Model – [31:29–34:11]
• Cybersecurity arms race & Opportunity – [35:51–38:52]
• Job Market & Knowledge Economy – [39:51–43:57]
• Future of Media & Micro-influencers – [45:22–49:15]
• Final Strategic Takeaways – [50:04–52:12]

Final Insights / Advice

• Exploit the Opportunity:
  Security professionals and tech entrepreneurs are entering a period of immense opportunity, both financially and in terms of impact—but expect the big firms to remain the ultimate owners.
• Bet on Reasoning & Memory Innovation:
  Advances in affordable, deep, and long-term reasoning models and architectures will drive a new wave of use cases and security breakthroughs.
• Adaptability Is Key:
  For individuals, being able to pivot, network, and adapt is more important than ever. Leverage community, distribution, and new tech early.
• Stay Authentic in an Automated Age:
  In both business and media, authenticity and community engagement will increasingly trump scale and polish.

“The ability for you to work at more core companies that are working in cybersecurity is a privilege...[and] the AI boom is going to put a pressure on electricity and water. If you invest in electricity or energy companies and energy portfolios, you’re most likely to be okay...”
— Sam Cummings [50:04]

Summary prepared for those seeking sharp, strategic insights—whether you're a cybersecurity professional, founder, or just AI-curious.