IQ is a commodity. Models are cheap. Proprietary data and workflows are the new moat.

The fight moved to the physical world. Power, grids, and nuclear energy are back in focus.

Agents are turning AI into labor. We are moving from selling software "seats" to selling outcomes.

DeepSeek demonstrated a 46x reduction in training costs and 90% lower inference costs. This creates massive deflationary pressure on "intelligence," transforming it into a commodity that will collapse software's cost structure.

Nvidia is winning right now, but 90% gross margins never last forever. Major customers (Amazon, Google, etc.) are building their own chips. At the same time, more efficient models like DeepSeek mean you need fewer GPUs to get the same work done. This challenges the "infinite demand" thesis

We are entering a phase defined by physical realities rather than by algorithmic discovery alone. The main constraint in the AI race is now securing gigawatts of power and the physical infrastructure to deploy it.

Low token cost is a weapon. When it drops far enough, it lets companies squeeze everyone else. As new chips like Nvidia’s Blackwell push costs down, the advantage will be longer thinking. Models can afford to reason for more steps, run more checks, and handle messy, multi-stage work like sales.

Early adopters of general-purpose AI in physical industries (robotics, logistics) are establishing 15-20% annual cost reductions. These advantages compound, creating "power law businesses" with cost structures that late-coming competitors cannot match, even with equal capital.

AI agents sell labor. When an agent does the work, charging “per user” no longer makes sense. Companies will have to price around outcomes, not logins. A lot of SaaS math will get rewritten here.

As models become easier to copy, data becomes the chokepoint. The most valuable companies will keep high-quality data within their walls while still squeezing value from models. Leverage moves from model vendors to data owners.

Startups can move fast, but Google is a behemoth. They can wire Gemini into Docs, Drive, Search, and everything else people already use. Google’s edge is having the model knowing your files, your emails, and your workflow. Context beats cleverness.

Governments are realizing they cannot rely on foreign AI infrastructure. This is driving a trend toward "Sovereign AI," where nations build domestic compute clusters and foundation models to ensure national security and cultural alignment.

As strong models become easily accessible, value moves away from the model itself and toward whoever owns the workflow. Customer relationships, proprietary data, and deep integration are what truly matter.

The current AI product cycle is the fastest in tech history because models keep showing new behaviors no one explicitly designed for. It feels less like software and more like biology. Capabilities show up, builders notice, and suddenly a dozen people are testing weird new uses at the same time.

AI products really take off when they add to how a company already makes money, not when they just cut costs. And when you capture outcome data end-to-end, the product gets better over time in ways competitors can’t easily copy.

Instead of betting only on big tech, a more durable exposure can be through power, cooling, datacenter services, networking, and software layers that monetize every incremental unit of AI demand. These layers don’t move in lockstep, which matters if you care about risk and correlation.

Public markets often underprice companies with persistent high growth (>30%) because it is unnatural to model such long, multi-year trajectories. This can make high-growth AI companies less risky than low-growth buyouts at reasonable multiples.

Future AGI may not be a "finished mind" but rather a deployed learner that improves on the job. Superintelligence will look less like a breakthrough moment and more like steady compounding in production.

Similarly, Dwarkesh believes that further leaps in AI capability will be driven by continual learning rather than a software singularity. Achieving human-level on-the-job learning may take another 5 to 10 years of incremental progress.

Coding is the ideal testbed for AGI because it provides a verifiable, closed-loop environment where an AI can write, execute, and correct its own output. Solving the reasoning required for complex software engineering is a proxy for solving general reasoning.

AI struggles with "folklore knowledge". That’s expertise known to practitioners but never written down in the literature. This implicit knowledge remains a major barrier for AI in specialized scientific fields.

The recent jumps in models like GPT-5 aren’t only about more data. Training now rewards completing multi-step tasks, not single replies. That pushes models to plan, execute, and recover when things go wrong. The result feels less like a chatbot and more like a junior employee.

Research into simple games like Othello proves that LLMs build internal representations (models) of the board state and rules, suggesting they truly "understand" the world they are simulating.

Synthetic data helps coverage, but it also sneaks in model bias. If you can’t measure when it drifts from reality, it quietly poisons your training set. The real advantage is knowing exactly when that data helps and when it hurts.

The most immediate risk from AI is Orwellian surveillance and censorship rather than an AGI takeover. Protecting privacy and preventing the state-level abuse of AI are more urgent concerns than sentience.

AI is a deflationary force so powerful it makes traditional economic metrics (like CPI/GDP) obsolete. In an economy where intelligence cost trends toward zero, the only inflation will be in assets that AI cannot reproduce (energy, land, trust).

Salesforce pulled off something most companies only dream about. By rolling out internal AI tools, they increased developer productivity by 30% in a year and moved 9,000 support engineers into new roles, rather than hiring more people.

By 2026, the role of a software engineer will transition from writing syntax in an IDE to managing a fleet of AI agents. The primary skill will shift from coding to architecture and agent orchestration.

Advanced agents now operate like self-improving recipes. You seed them with initial instructions, and they try variations, score the results, and rewrite their own prompts to "hill climb" toward better outcomes without human intervention.

Once agents can run code and move money, they become serious attack surfaces. Guardrails based on simple permissions stop working. Security now has to reason about what an agent is trying to do, not only what it’s allowed to touch.

Companies like Cognition (Devin) proved a new go-to-market strategy: generate massive viral attention with a "translucent overlay" UX that feels like magic, acquire users rapidly ($1M ARR in 10 weeks), and use that data moat to iterate faster than incumbents.

Harvey’s (Legal AI) GTM lesson is that enterprise adoption moves fastest when you can make a credible promise that customer data will not be used for model training. It changes procurement timelines and lowers the odds that the rollout dies in security review.

AI research tools can unlock decades of "institutional memory" trapped in old digital documents. One biotech customer avoided eight months of redundant mouse experiments after an AI tool found relevant prior internal studies.

Most AI benchmarks no longer work. Models hit the ceiling so fast that static tests stop telling us who’s better. What we need are moving targets: stitched-together tests and statistical views, like Epoch AI’s S-curves, that show progress over time instead of one-off wins. Otherwise, we’re reading noise as signal.

To hit the same capability level, we now need 2–3x less compute every year. That gain is coming from algorithms, not chips. Stack that on top of hardware improvements, and the pace of progress effectively doubles. This is why forecasts keep missing on the low side.

It is shockingly easy to "steal" a proprietary model by querying its API and training a smaller model on the outputs. This effectively commoditizes proprietary weights. "Moats" based solely on model performance are weaker than investors think.

The provisional "prompt box" interface will soon give way to proactive AI teammates that observe work and propose actions. Users will spend more time approving smart suggestions than crafting prompts themselves.

We are running into a thermodynamic wall. “Beff Jezos” proposes a shift to "thermodynamic computing". Instead of fighting noise and entropy, this approach uses them. Hardware is designed to compute probabilistically, not deterministically, which fits how modern AI actually works.

The next phase is Physical AI. We are moving from software that thinks to systems that act. The constraint is no longer reasoning capability. It is grid capacity, process knowledge, and the friction of the real world. Compute is abundant. Atoms are not. The winners in 2026 will be the companies that can wire AI into energy systems, factories, robots, and biology.

AI lets games write themselves as you play. Instead of fixed story branches, quests, and dialogue shift based on your choices in real time. That makes worlds feel alive and replayable without studios hand-authoring every path.

AI models like Microsoft’s MatterGen are doing for materials what AlphaFold did for biology. We are entering an era of "generative matter," where AI discovers new materials for batteries and solar panels tailored to specific properties.

Tools like Magic School are seeing explosive growth (5M users) not through top-down district sales, but via bottom-up adoption by teachers. When AI delivers such immediate ROI from administrative drudgery, users will adopt it without permission.

The energy demands of AI data centers are single-handedly reviving the nuclear industry. Tech giants are now looking at Small Modular Reactors (SMRs) as the only viable solution to provide carbon-free, baseload power for training clusters.

Networks like Bittensor are creating "free markets for intelligence," where different AI models compete to provide the best answers and are rewarded in tokens. This incentivizes a diverse ecosystem of specialized models rather than a single monolithic winner.

As AI generates infinite content, proving human origin becomes critical. Zero-Knowledge (ZK) proofs are becoming fast enough to work in real time, enabling content to verify its source without revealing identity. That creates a trust layer that the internet has been missing.