Alex Boris — AI 2025 → 2026 Live Show | Part 2

In one sentence

The AI landscape is shifting from a focus on "bigger is better" to a "capability-per-dollar" era where reasoning models and efficient small-scale models are redefining the economic value of human labor.

Why read this (and who it's for)

This summary decodes the geopolitical, regulatory, and technical shifts expected between 2025 and 2026. It is for AI researchers, policy-makers, and tech investors who need to understand why "older" hardware is still relevant, how "vibe-coding" and agents are changing the job market, and why the US-China chip war is becoming a matter of national security rather than just trade.

The big idea

The central thesis of the conversation is that we are entering a "paradigm shift" in AI development. While the public focuses on the sheer scale of frontier models, the real structural change is the collapse of cost. As models become more efficient, the "economic value" of a task (like software engineering or data analysis) becomes decoupled from the size of the chip required to perform it.

Furthermore, the conversation argues that we are moving toward an era of autonomous agents and robotics. The "chat" era is giving way to the "action" era, where the primary metric of success is no longer how well a model can speak, but how reliably it can execute complex, multi-step workflows (coding, research, physical tasks) without human intervention.

Key ideas

The "National Security" Model of Regulation

Alex Boris argues that AI regulation should not be viewed as "law enforcement" (punishing a crime after it happens) but as "national security" (preventing a catastrophe before it occurs).

• Why it matters: This shifts the burden of proof. Instead of proving a model is "bad," the focus is on identifying "extreme risks" (like bioweapons) that are so catastrophic they require proactive prevention.
• In the book: Boris highlights that the RAISE Act is about preventing "the bad thing from happening" rather than just dealing with it in a court process later.

The "Affordability x Capability" Multiplier

Peter Wilfofrd highlights a critical, often ignored metric: the massive drop in the cost of inference.

• Why it matters: If a model becomes 10x cheaper and 2x more capable, the total utility of the technology explodes. This allows for "on-device" AI (phones, laptops) to perform tasks that previously required massive data centers.
• In the book: Wilfoford notes that "old capabilities" can now be done by open-source models, which creates a massive opportunity for deployment in areas where privacy and cost are paramount.

The "Job-to-Job" Economic Reality

The conversation explores how the economic value of a task (e.g., a software engineer's output) remains constant even as the technology changes.

• Why it matters: If a 7-billion parameter model can do the work of a much larger model, the "job" is still being done, but the cost to the employer drops. This creates an "oversupply" of human labor for certain tasks, as the demand for human output may stay flat while the supply of AI-augmented output scales.
• In the book: The speakers note that a software engineer's job is still being done, but it is being done on "older, smaller chips," which fundamentally changes the labor market's equilibrium.

The "Agentic" Shift

Dean Ball and the others discuss the transition from "Chat" to "Agents."

• Why it matters: We are moving toward "coding agents" that can do "basically anything you ask them to." This isn't just a better chatbot; it's a system that can plan, execute, and verify its own work.
• In the book: The conversation notes that "Claude Code" and similar tools are reaching a point where they can act as autonomous agents, which has massive implications for policy and job displacement.

The China Chip "Renting" vs. "Selling" Dilemma

Peter Wilfofrd discusses the nuance of the US chip ban.

• Why it matters: The concern isn't just about "selling" chips; it's about whether China can "rent" the capability of those chips to power a military-grade AI stack.
• In the book: Wilfoford argues that because China is "weaponizing" AI for cyber-attacks and domestic human rights abuses, the US must ensure that the "power" of the chips doesn't just go to a competitor's economy, but specifically to their military/surveillance capabilities.

Frameworks & tools

The National Security Regulatory Framework

To evaluate if an AI policy is effective, use this three-part filter:

1. Prevention vs. Prosecution: Does the law stop a "bioweapon" level event (National Security) or just punish a "copyright" violation (Law Enforcement)?
2. Downstream Liability: Does the law change what is a crime? (It shouldn't—the goal is to prevent the capability from being misused).
3. Democratic Influence: Does the policy allow for a "democracy where we actually get to influence" how the technology develops, rather than letting 5 people decide?

The Capability-Cost Multiplier (Forecasting Tool)

When predicting the impact of a new AI model, do not just ask "How smart is it?" Ask:

1. Capability Score: How much better is it at the specific task?
2. Affordability Score: How much cheaper is it to run?
3. The Multiplier: Multiply these two. The winner is the model that offers the highest "capability-per-dollar" because that is what determines mass adoption.

Memorable examples & stories

The "Human Butler" vs. "Robot Butler"

Peter Wilfoford envisions a world where, by the late 2030s, the cost of hiring a "robot butler" to do chores becomes cheaper than hiring a human. Lesson: The ultimate test of AI's economic value will be its ability to replace physical and logistical labor, not just digital labor.

The "Political Genius" of the Presidency

Dean Ball discusses the "perplexing" moves of the current administration, specifically the tension between wanting to protect states' rights and the reality of selling H200s to China. Lesson: AI policy is often a messy, contradictory mix of national security, domestic politics, and economic reality.

The "Political Incorrectness" of AGI

Dean Ball mentions that some of the most important information is currently "politically incorrect," such as the claim that certain current models (like Claude Opus) are already "basically AGI." Lesson: The "whisper networks" of the industry often move faster than the public discourse.

Key quotes

"The point of the legislation... does not change downstream liability, does not change what is a crime... This is more on this national security side." — Alex Boris Why it matters: It clarifies that AI regulation is about preventing catastrophic capabilities, not just policing behavior.

"I think what goes a bit unnoticed is just like how much cheaper it's getting to do things as well." — Peter Wilfoford Why it matters: Cost is the primary driver of adoption, not just raw intelligence.

"A job is a job. That software engineer's job is still being done. It's being done on an older, smaller chip." — Speaker 2 Why it matters: This highlights the "commoditization" of human-level tasks.

How to apply it

This week

• Audit your "Agent" potential: Identify one repetitive, multi-step digital task you do weekly (e.g., "researching X and summarizing it into a table"). Ask: "Could a coding agent do this if I gave it the right tools?"
• Analyze "Capability vs. Cost": The next time you see an AI news headline, don't just look at the "bigger" model. Look for the "cheaper" model. Ask: "Who wins if this becomes 10x cheaper?"

Ongoing

• Monitor the "Whisper Networks": Follow the researchers and forecasters (like those mentioned: Tim Lee, Nathan Lambert, etc.) rather than just the mainstream "hype" news.
• Think in "Capabilities": When evaluating AI, stop thinking about "Chat" and start thinking about "Action." What can the AI do?

Common pitfalls

• The "Bigger is Better" Trap: Many people assume that the winner of the AI race will be whoever builds the largest model. The conversation suggests that the winner may be whoever finds the most efficient way to run a "good enough" model on cheap, old hardware.
• The "Law Enforcement" Fallacy: Don't confuse regulating AI usage (which is already covered by existing laws) with regulating AI capabilities (which is what the RAISE Act aims to do).

Connects to

• The Alignment Problem: How do we ensure these "Agents" don't take actions we didn't intend?
• The Moore's Law of AI: The idea that capability is increasing while cost is decreasing.

If you remember only three things

• The Multiplier: AI's impact is a product of Capability x Affordability.
• The Agentic Shift: We are moving from "Chat" (answering questions) to "Agents" (completing tasks).
• The Economic Decoupling: High-value jobs are being "decoupled" from expensive hardware; they can now be performed on smaller, cheaper, and older chips.

Internalize it

• Reflection questions

  • If an AI agent can perform your current job's core tasks at 1/100th of the cost, what is your unique "human" value add?
  • How would your industry change if the "cost of intelligence" dropped to near zero tomorrow?
  • Is the current AI regulation focused on the right things (preventing catastrophic risk) or just the wrong things (policing content)?

• Flashcards Q: What is the difference between "Law Enforcement" and "National Security" in AI regulation? A: Law enforcement deals with crimes that have occurred; National Security deals with preventing catastrophic risks (like bioweapons) before they happen.

Q: What is the "Capability-Cost Multiplier"? A: The idea that AI's impact is determined by the product of how much better it is (Capability) and how much cheaper it is to run (Affordability).

Q: What does it mean that "a job is a job" in the context of AI? A: It means that as models become more efficient, the same economic task (like software engineering) can be performed on older, smaller, and cheaper chips, potentially creating an oversupply of human labor.

Q: What is the "Agentic" shift in AI? A: The transition from "Chat" (models that talk) to "Agents" (models that can plan and execute multi-step tasks like coding or research).

Q: Why is the US-China chip ban a "threat model" issue? A: Because China can use high-end chips to power a military-grade AI stack for cyber-attacks and domestic surveillance, making it a national security concern.

Fidelity note

This summary is grounded in the provided transcript. I am confident in the distinctions made between "Law Enforcement" vs. "National Security" and the "Capability x Affordability" framework. The specific mentions of "Claude Opus" and "coding agents" are directly from the transcript. The "Robot Butler" anecdote is a primary example used by Peter Wilfoford to illustrate the future of robotics.