Two years of coding in two weeks
But are we trading safety for velocity?
So I’m done. In two weeks, I ‘coded’ an application that would have taken two years by hand. Or, perhaps more realistically, a team of 5 six months to build. I used o3-mini-high (a lot) and Claude. But I never once looked at docs, never once looked at Stack Overflow, never once used Google.
I built scaffolding to help me go faster. Tools to automate code mgmt and diffing. I used Claude and o3 to review the code. I reviewed the key parts of the code myself, sure, but used Claude to generate control flow diagrams. To explore specific parts of the logic.
It’s amazing. It’s exciting. It’s exhausting. I keep thinking of Dr Seuss - and the places I can now go that I couldn’t before. I never want to go back to the old ways.
This breakthrough's timing is no accident - it's a direct result of the accelerating AI arms race, with OpenAI rushing o3-mini out in response to DeepSeek's R1. Which is where things get interesting...
Back to DeepSeek
In the three weeks since R1 was released it’s become clear much of the initial narrative was flawed. The $6M ‘cost’ that spooked the US was incorrect; DeepSeek have spent billions on hardware.
Dario Amodei, CEO of Anthropic, recently noted DeepSeek were able to exploit “an interesting crossover point, where it is temporarily the case that several companies can produce good reasoning models”. This is because we’re early in the evolution of thinking (test-time-compute) models. But as time passes “this will rapidly cease to be true.” And the US, with their superior hardware will be able to draw away.
He’s almost certainly right. DeepSeek are smart, but they haven’t done anything that couldn’t be replicated by a US lab. And DeepSeek published their research - sharing it with the world. It’s easy for the US labs to copy.
AI remains a compute intensive game. The more compute you have, the better models you can build. The more efficiently you use your compute, the better models you can build. DeepSeek found a clever way to improve efficiency and shared their secrets. Now the US labs will apply those learnings to their models on their superior hardware. DeepSeek will help the US go faster.
Making it worse for China, DeepSeek appears to be an outlier. Most of the Chinese AI companies are run as traditional software dev houses, driven by KPIs and OKRs. DeepSeek are different. They have a structure which encourages innovation. Which enables people to invest significant time, money and resources in promising hunches. But will they manage to retain that after their recent success? Can any of the other Chinese labs find a way to duplicate the DeepSeek model?
The reality is China is behind. For now it is able to do amazing things due to innovation and talent. Arguably some of the clever efficiency gains they found were forced on them by US export controls that prevent the export of the best AI hardware to China. But innovation and talent can only take China so far. The US has plenty of innovation and talent. And a lot more hardware. The US is comfortably ahead.
Why should we care?
DeepSeek changed the mood music in the US. There’s an increasing belief the US is (a) in a race with China to AGI/ASI and that (b) the US needs to win at any cost. But what happens when we build AGI/ASI?
By definition AGI/ASI means smarter than human intelligence. Building machines smarter than us is something to take seriously. There isn’t a good track record of less intelligent species dominating more intelligent ones. Many smart folk have worried about this for many years, but as yet there are no answers. No strategy.
It’s easy to see the appeal of AGI/ASI. It has many upsides - medical advances, new materials, deeper insights. But it almost inevitably comes with a whole set of downsides too. And we don’t know what those are. However, we may not have long to wait before we find out. Two leading labs in the US (OpenAI and Anthropic) are openly talking of delivering AGI/ASI by 2026/2027. It’s a believable timeline; progress has significantly accelerated over the past six months.
Given the potential risk, you’d imagine safety would be top of mind. But, amazingly the key players are all backing away from the safety discussion.
Take Anthropic. Until now they’ve been a voice of responsibility. They’ve focused on a do-no-harm ethic. Their mission statement is: “Anthropic is dedicated to building systems that people can rely on and generating research about the opportunities and risks of AI.”
But Amodei no longer mentions alignment or safety. Instead he’s focused on export controls and preventing China getting chips. He talks about bipolar and unipolar worlds. Ones where either China and the US dominate or just the US dominates. The rest of the world is forgotten. It’s a significant shift.
Then we had the AI Summit in Paris this week. JD Vance opened with “I'm not here this morning to talk about AI safety… I'm here to talk about AI opportunity.” And followed up with: “The AI future will not be won by hand-wringing about safety. It will be won by building.”
The head of Softbank (a major OpenAI investor) told the world on Monday: “I would say if their (AGI) source of energy was protein, then it's dangerous. However, their source of energy is not protein, so they don't have to eat us.” It’s an, err, unusual viewpoint from someone so senior.
There’s the simmering fight between Elon Musk and Sam Altman (CEO of OpenAI) over control of OpenAI. The two fell out around 2018 over how to develop AI safely and responsibly. Musk’s bid of $97.4 billion this week complicates Altman’s attempts to convert OpenAI from a non-profit to a for-profit. And Musk could yet use his influence over Trump to try to wrest control of OpenAI.
Finally there’s the declaration from Paris. The US demanded that the final statement excluded any mention of the environmental cost of AI, existential risk or the UN. The UK refused to sign. And, despite all their demands, the US didn’t sign either - due to DEI objections.
What happens next?
The US is racing towards an uncertain future at full speed, pulling the whole world along with it. Maybe JD Vance is right - perhaps AI will simply supplement rather than replace human work, making us all more productive. After all, I've seen first-hand the massive productivity gains that supplementing can bring.
But there’s inevitably going to be change. 2025 is shaping up to be the most disruptive year ever in software engineering. Jobs as translators are disappearing. AI can compress months of dissertation work into 4 days. This transformation is just the beginning - we're witnessing the early tremors of a seismic shift that will reshape not just how we work, but what kinds of work remain relevant for humans.
The shift in mood music is concerning. Abandoning safety discussions in favour of beating China may turn out to be a mistake. The US may win the race to AGI, but with no clear plan for controlling AGI/ASI we may unleash a force we are ill prepared to manage. We may not have to wait long to find out how this story unfolds.