01/What AI actually is
The AI everyone is talking about is a large language model. An LLM. It predicts the next token.
That sounds small. It is not.
In 2017 a team at Google published Attention Is All You Need and introduced the transformer (Vaswani et al., 2017). In 2018 Alec Radford at OpenAI built GPT-1 on top of it (Radford et al., 2018). Put the two together, scale it up, and a computer that predicts the next token can do pretty cool things. Like write code.
An LLM is the combined corpus of the internet's intelligence, curated by labs to produce outcomes. Give it a goal and the tools to reach it, and it works through the task on its own. That is the technical meaning of an agent.
Douglas Engelbart saw where this was going sixty years ago. That is the job. AI is one of the most important tools of this century because it enhances what a person can do.
02/The second gold rush
The first gold rush in tech was software as a service. The name of the game was automation. Companies paid millions to automate actions on the internet, and that meant paying millions for code.
Code was the gold mine of the 2000s.
Look at what a good engineer cost against what the company was worth. In 2005 the median software engineer made about $82,000 a year (Dr. Dobb's salary survey, 2005). Google went public in 2004 worth $23 billion with 3,021 people on payroll. That is about $7.6 million of company for every person in it. Roughly ninety times the salary.
So paying an engineer hundreds of thousands was cheap. Paying one millions was still cheap. Google paid Anthony Levandowski about $120 million while he built its self-driving car (NBC News). In 2025 Sam Altman said Meta was offering OpenAI researchers signing bonuses of up to $100 million (TechRepublic). When the company is worth trillions, $100 million is a rounding error.
That was the first gold rush. AI is the second.
The difference is who gets to dig. Before, you needed money to hire great engineers. Now you can hire fewer engineers, or hire people with good engineering principles, and let AI carry the rest.
An AI-powered person raises the baseline. Almost anyone can mine for gold now.
03/AI does not replace people. It cuts the bloat.
The market is wrong about one thing: that AI replaces humans.
It does not. It replaces the need to keep hiring.
Look at it the right way. The question is not which engineers AI can replace. The question is why you would scale your engineering team at all.
Smaller is better. It is more cost effective to move lean than to move big. Every person you add is another set of conversations to keep in sync. Five people have 10 lines between them. Fifty people have 1,225. That is why decisions take a while as a company grows, and why adding people to a late project makes it later.
The best companies already build this way. Bezos keeps teams no bigger than two pizzas can feed, because the smaller the team, the better the collaboration (AWS). A study of more than 65 million papers, patents and software projects found that small teams disrupt, while large teams develop what already exists (Wu, Wang and Evans, Nature, 2019). And look back at Fig. 2: Instagram was 13 people. WhatsApp was 55.
AI makes lean the default. More than a quarter of all new code at Google is generated by AI, then reviewed and accepted by engineers (Fortune, 2024). With the right setup, a small team can run a huge codebase.
And it lifts the people you keep. When 5,179 customer support agents got an AI assistant, they resolved 14% more issues per hour. The newest and least skilled went up 34% (Brynjolfsson, Li and Raymond, NBER). The study measured a lift, not a layoff.
Do not replace your engineers. Do not bloat around them. Keep the team small and make every person count.
Holistically, most engineers are bad programmers. Only a few are worth their skill, and the ones who are have coding principles. AI amplifies whatever you bring to it. Bring principles and it compounds them. Bring none and you get slop.
So here is the market going forward. People need capable hardware to put AI to work. Faheem Labs makes that translation easier: we sell the hardware that unlocks human potential with local AI, free from bottlenecks.
04/Find the bottleneck
Right now most of us rent these tools from Codex and Claude. They are great tools.
What happens when they go down? Your work stops.
What happens when you get banned? Everything you built on them is gone.
It is not hypothetical. On June 10, 2025, ChatGPT and the OpenAI API were down for over ten hours (Tom's Guide). StatusGator has logged 427 Anthropic outages since January 2025 (StatusGator).
The bottleneck is the AI distributors.
If you are good at business, the first rule of action is identifying the bottleneck. If this one thing goes down tomorrow, can you survive? Are you even aware of it?
In the age of AI, can your AI business survive the worst case? If your intelligence lives on someone else's server, the answer is not up to you.
05/Do you trust them?
The other factor is your data. AI companies use the information you send them. How much varies by company, by plan and by whatever the terms say this year. In 2025 Anthropic asked Claude's consumer users to choose whether their chats train its models, and kept the data of those who said yes for five years instead of thirty days (Anthropic, 2025).
Then look at their stances. The same labs selling the tools are some of the loudest voices saying the tools might end the world. The heads of OpenAI and Anthropic both signed the Center for AI Safety's extinction statement.
Do you trust Claude and ChatGPT with your code, your customers and your ideas?
To me, the answer is clearly no.
06/Independence
The goal is to be independent of them. With local AI, you can be.
Software is close to free. The best open-weight models trail the best closed ones by about four months (Epoch AI, 2026). Models small enough to run on a single consumer GPU catch the frontier within six to twelve months (Epoch AI, 2025). They cost nothing to download, and every workstation we ship has them preinstalled.
What you cannot download is the machine to run them. That is the moat, and that is what we build.
True intelligence is spreading the wisdom to others. Faheem means “one who understands.” But one is not the goal. Hoarding intelligence for the few is how we got here. The goal is to enable it for the many.
Not one who understands. Many who understand.
That is the goal for Faheem Labs.
You own the GPU.
You own the intelligence.