
The Great Unlocking: Start Building Now, and Remember Everything
The Great Unlocking: Start Building Now, and Remember Everything
以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。
Generative Ai Tools
AI
Agentic Ai
Agentic Ai Architecture
The Great Unlocking: Start Building Now, and Remember Everything
Dr. Majid Fekri
9 min read·
Just now
--
A forecast, not a fantasy
My career started in atmospheric science. The first thing that field teaches you is humility about the future.
You don’t forecast by wishing. You read the current state of the system, you measure the trends, and you follow the physics forward. Then you say what you see, plainly, even when it sounds bold.
So when I tell you we are entering the most abundant season for builders in human history, I want you to know I am not daydreaming. I am reading the instruments.
The instruments say a front is moving through. Not a storm. A clearing. And behind it is a wider sky than any generation of builders has ever looked up at.
Here is the first instrument. In March 2025, the research group METR measured how long a task AI agents can finish on their own, scored by how long the same task takes a skilled human. That length had been doubling roughly every seven months for six straight years. Since 2023, the pace has quickened to roughly every four months.
The curve did not flatten. By May 2026, frontier agents had climbed high enough on METR’s scale that the team added a note: their task suite can no longer reliably measure anything above 16 hours. The ruler is running out.
Here is the second instrument. Epoch AI found that the price of reaching a fixed level of AI performance fell between 9x and 900x per year across the tasks they studied. GPT-4-level answers on PhD-level science questions got about 40 times cheaper every year.
Put those two readings together. Capability doubles on a clock measured in months. The cost of yesterday’s capability collapses on a clock measured in quarters. When something gets much better and much cheaper at the same time, it stops being scarce. That is what abundance means.
A good forecaster also reports the uncertainty, so here it is. METR is candid that its tasks are clean and self-contained, closer to what a capable new hire could do with no prior context. And when METR asked real open-source maintainers to review agent-written pull requests that had passed the automated tests, roughly half would not have been merged.
I find that second number exciting, not discouraging. The gap between a benchmark and real work is mostly context: history, judgment, and the feedback loop. That gap is exactly where you come in. Hold on to that thought. We will come back to it.
Let me describe what I see coming in the next few years. None of it requires a miracle. It only requires the curves above to keep doing what they have done for years.
This will not arrive evenly, and it will not arrive without friction. Every big shift brings disruption, and we owe each other honesty and care through it. But the direction is clear. The cost of turning an idea into a working thing is falling toward zero, and when that happens, the number of things worth building explodes.
There is a myth that this future belongs to the big labs and the giant companies. It does not. They build the engines. The rest of us decide where to drive.
The most valuable thing in this new world is not access to a model. Everyone has that now. It is knowing a problem better than anyone else does.
The nurse who knows exactly which shift handoffs go wrong. The logistics manager who knows why the trucks are always late on Thursdays. The farmer who knows what the soil does after a dry August. Domain knowledge used to need a software team to become a product. Now it needs a weekend and an agent.
Starting early matters more than starting perfectly. Every project teaches you how agents reason, where they stumble, and how to guide them. That intuition compounds, and the person who starts today will have a year of it by the time someone else decides to begin.
I have seen this pattern before. I built a startup in an earlier wave of technology, and the lesson was the same then. The winners were rarely the ones with the most resources. They were the ones who started, shipped, learned and kept going.
So pick something. Start small. Start today.
Now the most important part of this article.
When you build with agents, you will move faster than you ever have. In a single week you might make a hundred decisions with them. Why you chose this database over that one. Which instruction finally made the agent behave. What broke at 2 a.m. and how you fixed it. What your first customer said that changed the whole design.
By default, agents wake up with no memory of yesterday. Every new session starts from zero. You re-explain the project, the agent repeats a mistake you already solved, and a new teammate, human or AI, has no idea why things are the way they are. You pay the same tuition twice, then three times.
In science, we keep research notebooks. Not because anyone loves paperwork, but because an experiment you cannot reproduce is an experiment you did not really do. The notebook is the difference between a lucky result and real knowledge.
Agentic work needs the same discipline, and it needs it more. Remember what METR found: the gap between a benchmark and real work is mostly context. Context is history, and history only exists if someone writes it down.
Here is the deeper reason it matters. Models are getting better and cheaper every quarter, and you will swap them many times. What you cannot swap, and what no one can sell you, is your accumulated understanding: your decisions, your failures, your customers, your way of doing things. The model is rented. The memory is yours. Over time, that memory becomes the most valuable thing you own.
There is one more reason: trust. As agents act on our behalf, we need to be able to ask what they did and why. A record of the process is what makes agent work accountable, easy to debug, and safe to hand more responsibility.
A context window is not memory. The context window is the agent’s desk: everything on it is visible right now, and it gets cleared when the session ends. Agentic memory is the filing cabinet behind the desk. It is a persistent store the agent writes to and reads from across sessions, across tools, and even across different models.
Good agent memory holds a few kinds of things:
A few principles hold no matter which tools you use:
Plugging it in is simpler than most people think:
A word of honesty: I build memory infrastructure for agents, so weigh my enthusiasm accordingly. But everything in this section works with a text file and a bit of discipline. The habit matters far more than the tool.
If you want a place to start, here is one.
Then do it again next week, a little bigger.
Forecasters have a quiet joy that few people talk about. It is the moment you watch a front pass on the radar and know that, a few hours from now, people will step outside into a clean, bright sky.
That is how I feel about this moment. The tools are here. The costs are falling. The opportunities are wider than any of us can see from where we stand today.
So start building. Choose a problem you care about and point an agent at it. And as you go, write it down: every decision, every failure, every small win. Give your agents a memory, and give yourself one too.
Ten years from now, people will look back on these years the way we look back on the early days of the web. Some will say they watched it happen. I hope you will be able to say you built something, and show exactly how you did it.
正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。