"OpenAI announced GPT-4 Turbo at DevDay with a 128K context window — then fired its CEO Sam Altman eleven days later, before reinstating him five days after that." — The Verge, November 2023
A year ago this month, OpenAI released ChatGPT to the public. It hit one million users in five days. One hundred million by January. Nothing in consumer technology had ever moved that fast.
Now, twelve months later, with the dust somewhat settled, I want to write an honest retrospective. What actually changed for businesses? What turned out to be hype? What surprised me? What should you be thinking about heading into 2024?
What Genuinely Changed
The talent market shifted. A year ago, "prompt engineering" wasn't a job title. AI engineering was a specialist field with a thin talent pool. Today, there are thousands of developers who've shipped AI-powered features, hundreds of consultancies specialising in LLM implementation, and a generation of business people who understand at least the basics of what's possible. The knowledge has diffused. That's good. The API ecosystem matured. OpenAI's API a year ago was reliable but limited. Today, we have function calling, improved context windows, fine-tuning for GPT-3.5, embeddings that are fast and cheap, and a growing ecosystem of tools built on top. The developer experience is substantially better. The use case library is real. Twelve months ago, we were largely figuring out what LLMs could actually do in production. Now there's a body of evidence. Document processing, customer support, code generation, internal knowledge search, content drafting, data extraction — these aren't theoretical. Companies are running them in production, measuring results, and expanding them. AI literacy in business increased. Not uniformly, and not always accurately — a lot of executives still have misconceptions. But the average business owner has a much more concrete idea of what AI can do than they did a year ago. Conversations are less exploratory and more tactical.What Didn't Change (Or Changed Less Than Expected)
Most businesses haven't implemented anything. Despite the hype, the majority of SMBs haven't deployed anything beyond individual employees using ChatGPT ad-hoc. The gap between "experimenting with AI" and "AI in production workflows" remains large. Implementation is harder than demos suggest, and most organisations haven't found the right entry point. Data problems are still data problems. The AI can only be as good as the data you give it. And most businesses' data — fragmented, inconsistent, inaccessible — hasn't changed in the last year. The AI hype hasn't accelerated data infrastructure investment proportionally. This is still the primary bottleneck. The hallucination problem hasn't been solved. Models have gotten better, but they still confidently make things up. Every production system needs guardrails, validation, and human oversight. Anyone building AI without accounting for this is building fragile systems. "AI strategy" still mostly doesn't exist. Most businesses have AI tactics — specific tools or implementations — but not a coherent strategy for how AI capabilities will change their business model over time. The organisations that will be most advantaged in three years are the ones building that strategic layer now.What Surprised Me
The speed of open-source development. I expected proprietary models to stay dominant for longer. The quality of Llama 2 and derivatives, released just this past summer, was faster than I anticipated. The open-source ecosystem is genuinely competitive for many use cases, which changes the cost and privacy calculus significantly. How much prompt engineering matters. I underestimated, a year ago, how much the quality of your instructions to the model affects output quality. Two teams using the same model can get dramatically different results based on prompt design. This is a real skill — not rocket science, but consequential and learnable. The multimodal acceleration. GPT-4V (vision capabilities) shipped in October. We can now feed images, charts, and documents directly to the model. The implications for business — processing invoices, analysing product photos, reading visual reports — are significant and we're just starting to explore them. The Japan market moving faster than expected. I've been based in Tokyo, and a year ago the conventional wisdom was that Japanese enterprises would be slow adopters. In practice, large Japanese corporations are moving aggressively on AI implementation — both out of genuine interest and out of concerns about falling behind international competitors. The government has also been supportive rather than restrictive.What Heading Into 2024 Means
The businesses that are six to twelve months into thoughtful AI implementation have a growing advantage over those that haven't started. Not because they have better technology — everyone has access to the same APIs. Because they have experience: they've learned what works in their domain, they've built the organisational muscle, they know which processes are good candidates and which aren't.
That experience compounds. It doesn't get easier to catch up.
The window for "getting started" isn't closed — it's not even close to closed. But the cost of waiting is incrementally higher each month, because the early movers are getting better while the late starters are still figuring out fundamentals.
2024 is going to be the year that separates the businesses with AI strategies from the businesses with AI experiments. Start planning which side of that line you want to be on.
Want to implement AI in your business? Book a 30-minute architecture call at blok-ai.com.