2025

Part I

Geopolitics

April 2nd got a name: "Liberation Day." Trump announced tariffs on basically everyone, and by everyone I mean 145% on China, 25% on Canada and Mexico, and a 10% baseline on the rest of the world, which meant that companies who had spent the last decade optimizing their supply chains around cheap Chinese manufacturing and NAFTA suddenly found themselves scrambling to figure out what their cost structure would look like in six months.

Markets did what markets do. The Dow dropped 2,200 points in two days and supply chain people started making calls and everyone waited to see if this was a negotiating tactic or actual policy, which of course turned out to be both, because the 90-day pause announced in May became less of a pause and more of a new permanent state of uncertainty where companies stopped making long-term plans and started making contingency plans instead.

Ukraine entered year four with no resolution in sight. Peace talks happened but produced nothing concrete and the front lines barely moved and Europe kept asking what security looks like if America pulls back, a question to which nobody had a particularly good answer.

Sudan became the crisis nobody covered, which is a strange thing to say about what is probably the world's worst humanitarian catastrophe right now. Civil war between the army and RSF displaced millions of people, famine conditions spread across multiple regions, and the international response remained fragmented and underfunded in a way that would be unthinkable if this were happening somewhere the Western media cared more about.

India and Pakistan flared up again with cross-border incidents and diplomatic expulsions, which is the usual dance between two nuclear powers that share a border and a grudge and occasionally remind everyone that South Asia remains one of the most dangerous places on Earth for reasons that have nothing to do with AI or tariffs.

The bigger story though was fragmentation, and not any single dramatic event but rather the cumulative weight of a hundred small realignments. BRICS expanded. De-dollarization efforts accelerated. Belt and Road kept building. The post-WWII order continued its slow unwind into something that looks less like a new world order and more like several competing orders that haven't quite figured out how to coexist yet.

Part II

Artificial Intelligence

January brought DeepSeek R1, a reasoning model from a Chinese startup that matched GPT-4 performance at a training cost of around $6 million, which is notable because frontier models from OpenAI and Anthropic have been costing hundreds of millions of dollars and suddenly here was a model that achieved comparable results for roughly 1/50th the price, built by a team that most people in Silicon Valley had never heard of.

Nvidia lost $600 billion in market cap in a single day on January 27th, which tells you something about how much of the AI thesis had been built on the assumption that you needed massive compute budgets to compete, and how quickly that assumption could unravel when someone demonstrated that algorithmic efficiency might matter just as much as raw GPU hours.

The training paradigm shifted in ways that felt both gradual and sudden at the same time. RLVR became the acronym of the year, which stands for Reinforcement Learning from Verifiable Rewards and basically means models that think step by step before answering rather than just pattern-matching their way to a response. OpenAI shipped o1 and o3, Anthropic upgraded Claude, Google pushed Gemini, and the benchmark became whether a model could reason through a problem rather than simply respond to it.

Agents escaped the lab and became products you could actually buy. Systems that browse the web and write code and execute multi-step tasks and interact with your computer the way a human would, which sounds like science fiction except it's not anymore, it's just a feature you can turn on in Claude or GPT and watch as the AI figures out how to navigate your desktop to accomplish whatever you asked it to do.

Andrej Karpathy coined "vibe coding," which is exactly what it sounds like: you describe the vibes of what you want, the AI writes the code, and programming starts to change shape in ways that make some people excited and other people nervous. GitHub Copilot and Cursor and Claude Code became standard tools, junior developer hiring slowed down because senior engineers could now do the work of entire teams, and the conversation shifted from whether AI would replace programmers to what programming even means when most of the syntax is written by machines.

Sam said the quiet part loud and Dario at Anthropic predicted "powerful AI" by 2026-2027 and whether any of this is genuine confidence or fundraising narrative or something in between became the parlor game that everyone in tech played at dinner parties. The definition of AGI stayed contested but the confidence of the people building it grew more pronounced with each passing quarter, which either means they know something we don't or they're very good at convincing themselves and everyone else that they do.

Part III

Crypto

Bitcoin crossed $100K and then pushed above $126K by October, and the difference from previous cycles is that this time it was institutions buying rather than retail mania, which changes the character of the whole thing in ways that are hard to fully appreciate until you realize that BlackRock's ETF became one of the most successful fund launches in history and corporate treasuries started adding BTC to their balance sheets like it was just another line item in their asset allocation strategy.

Congress actually did something, which is remarkable enough to note on its own. The GENIUS Act created a federal framework for stablecoins and the CLARITY Act distinguished securities from commodities and after years of regulation by enforcement, where the SEC would just sue people and let the courts figure out what the rules were, the industry finally got actual legislation that told them what they could and couldn't do.

Larry Fink completed his arc from skeptic to evangelist in a way that probably deserves its own case study in how institutional finance adapts to things it initially dismisses. The BlackRock CEO now talks about tokenization the way he used to talk about index funds, arguing that every asset will eventually be tokenized and settlement will become instant and ownership will be fractionalized, and he launched tokenized money market funds to prove that this isn't just talk.

Tokenized real-world assets grew from $5.5 billion to nearly $20 billion over the course of the year (and over $250 billion if you include stablecoins), which includes treasuries and real estate and private credit and all the boring financial instruments that nobody associates with crypto but which actually represent the future of the industry more than any memecoin ever will. Chainlink connected traditional finance to blockchain infrastructure and the boring use case of efficient settlement started to overtake the exciting use case of speculation in terms of actual dollars moving through the system.

The administration floated a Strategic Bitcoin Reserve, which is the idea of the government holding BTC as a national asset the way it holds gold or foreign currency reserves. Whether this was serious policy or political positioning or something in between remained unclear, but the fact that it was being discussed at all marked a shift in how crypto gets talked about at the highest levels of government, from fringe curiosity to legitimate policy consideration.