https://www.youtube.com/watch?v=rnMDF3kraiI&t=161s
无作者
This is RollerCoaster Tycoon. It was released in 1999, and it was written almost entirely in x86, Assembly. Because of that, Chris Sawyer managed to simulate complex physics along with 10,000, individual AI guests on a 90 megahertz Pentium computer. Hardware so limited,it would struggle to load a basic modern website today.For decades, the industry moved away from this level of efficiency. We’ve been living in the, computational golden age, the so-called “free lunch” era. We did this because of Moore’s Law.For decades, microchips got twice as fast and twice as cheap every two years. Since hardware, was practically free, the machine’s performance didn’t really matter. Only developer time did.Essentially, high-level languages won. Why worry about language efficiency when Moore’s Law bailed, us out every single time?
bail sb ↔ out, to leave a large sum of money with a court so that someone can be let out of prison while waiting for their trial
〔交保释金〕把某人保释出来:
Clarke’s family paid ￡500 to bail him out.
克拉克的家人交了 500 英镑保释他出来。
And this, the era of “Move fast and break things” was born.But Moore’s Law has just hit a physical brick wall. Microchips have become so dense that.
we are reaching the literal limits of physics. Since 2004, the free hardware, lunch officially ended. Software companies are now racking up record levels of server costs,
rack up: (especially NAmE) to collect sth, such as profits or losses in a business, or points in a competition
The company racked up $200 million in losses in two years.
公司两年内损失累计达 2 亿美元。
In ten years of boxing he racked up a record 176 wins.
在十年的拳击生涯中，他累计获胜 176 次，创下纪录。
spending upwards of 400 million dollars per year just to run basic apps on the server,while user interfaces have slowed down to a crawl. I’m looking at you, Jira.But Google dropped a bombshell last month. They announced that 75% of all new code at Google, is now AI-generated and approved by engineers, up from 50% last fall. That’s the Google we’re, talking about. Yeah, they’re using an internal LLM called Goose trained on their own codebase, but, that’s a crazy rate of progress. At this rate, AI will be writing 100% of the new code in no time.But if that’s the case, then why are we continuing to optimize for the human with, high-level languages when we should be optimizing for the machine with low-level languages and, reaping the performance benefits? And by the end of the video, you’ll have an understanding.
of what kind of engineers will be obsolete and what kind will still have jobs—hopefully.The next logical step for the software industry is an AI-driven migration back to the metal,using AI to write high-performance compiled code to wipe out those multimillion-dollar server bills, to a fraction of what they are today.
wiped out
to destroy or remove sb/sth completely
Link to original
This migration brings huge cost savings because, high-level languages pay a massive abstraction tax to stay human-readable. They rely on heavy, runtime interpreters to translate instructions on the fly. They use dynamic typing that forces, the CPU to constantly guess data types, and automatic garbage collection algorithms that, steal processing power to clean up the data. When a company scales to a 400-million-dollar, server bill, they are paying a massive premium just to keep that software middleman alive.But we can’t do this yet for two reasons. First, there isn’t enough low-level code on the internet.
for the AI to learn from. Second, the stakes are way too high. If an AI messes up in Python,the app just crashes safely. But if it messes up at the low level, it creates, a hidden memory leak and massive security backdoors instead of a clean error message.That being said, frontier AI labs are bypassing this wall using compiler-in-the-loop reinforcement, learning. Instead of just training the AI to mimic text, they hook the training loop up to a live, compiler and test environment using algorithms like GRPO (Group Relative Policy Optimization).The AI gets an immediate, automated reward signal: did it compile and pass? This, environment effectively beats syntax and formatting errors out of the AI models.And to fix the data scarcity for ultra low-level targets like Rust,Zig, or hardware-specific assembly, labs are deploying automated synthetic, curriculum generation. Advanced AI models are being paired with strict compilers to.
mass-produce their own hyper-optimized, error-free low-level training data,scaling the data reservoir synthetically without relying on the public internet.But once the AI starts pumping out millions of lines of low-level code, we hit a lethal wall:the watchmen problem. High-level languages like Python were explicitly designed to, be readable by human eyes. Low-level languages are optimized for the machine,largely ignoring human comfort. Because of this, traditional human code review becomes more taxing.The codebase becomes an unreadable black box where no human engineer alive can realistically, audit or catch subtle logic flaws across millions of lines of low-level instructions.Therefore, the old way of thinking about programming is dead. Since humans can no, longer read or audit the underlying low-level code anyway, the human workflow shifts entirely, to spec-driven development, like with Speccit. The industry doesn’t need people to type out.
code syntax line by line anymore. Instead, human product architects write airtight,unambiguous logical specifications, defining the exact inputs, outputs,and business rules, and the AI treats that specification as the new high-level language.And because manual line-by-line debugging becomes completely impractical at this scale,we must pivot our quality assurance to agentic testing. Instead of forcing tired engineers, to read endless files to find a needle in a haystack, we build an adversarial red team of AI, agents whose only job is to relentlessly bombard, probe, and try to break the system. If a fleet of, AI testers attacks a piece of software for 24 hours and can’t find a single vulnerability,that code achieves a level of operational safety that manual human review simply cannot match.But for mission-critical systems like aerospace, defense, or medical software,even aggressive testing cannot rule out rare cosmic events or rogue hardware.
anomalies where failure means loss of human life. Therefore, the highest stakes code, will rely on formal mathematical verification using advanced frameworks like Kani. It looks, and runs exactly like a traditional unit testing framework, but instead of executing sample inputs,it uses mathematical solvers to exhaustively prove the code is completely safe across, every single possible input before it ever touches a physical microchip.But this doesn’t mean human programmers are going away entirely. It just means the day-to-day job is, shifting. Instead of spending your day writing out syntax or hunting down bugs line by line,your role moves to a higher plane. Humans become the system architects,the data seeders, and prompt designers who guide the core logic,managing the AI agents that do the heavy lifting under the hood.This doesn’t mean high-level languages like Python or Javascript are completely dead. They aren’t.
going anywhere. But we are moving into a brand new era where, whenever efficiency and server costs, actually matter, companies will heavily prefer to let AI compile code straight down to the metal.The new high-level language isn’t code at all; it is pure human intent written in specifications,allowing us to build software at a scale Chris Sawyer could have only dreamed of.
--- 由 vCaptions 生成 ---