Why hasn't AI eaten the world, yet? And how can we help it?

AI is one of the biggest technology shifts of our lifetime. Trillions are going into data centers, every serious software company is rebuilding around agents, and even my mom is using ChatGPT now.

So why hasn’t AI eaten the world yet? Why are so many companies investing in AI without seeing real, tangible benefits?

My current answer: adoption is blocked less by model intelligence, and more by three boring constraints: inference capacity, organizational inertia, and trust.

Solve one of these bottlenecks, and you potentially can get rewards proportional to the value being delivered. BitGN is my bet on the third bottleneck: making reliable AI architectures discoverable faster by turning business problems into fun engineering challenges.

The first bottleneck: inference capacity is limited. It is bound by GPUs, electricity, and data centers. Even major AI providers are struggling with that. Just take Anthropic: within a few weeks they had to push OpenClaw and other third-party harness usage out of subscription limits, then publish a Claude Code postmortem after product changes made the tool feel worse for developers.

Of course, if you have money, tokens are not a problem. Large companies can easily hand out six-digit LLM token budgets to their employees (I heard three separate stories about that in the past few weeks), but that doesn’t solve the problem for the business in general.

Fixing the problem would require a lot of CapEx, ranging from buying an Nvidia DGX cluster to building a data center. So let’s skip that for now.

The second bottleneck in AI adoption is organizational inertia. For instance, a company could follow a top-down adoption path, where directors just buy Copilot licenses or some other tool to sprinkle fairy AI dust over the processes. Pilots could look good on paper and even in PoCs, yet stall completely in reality.

This happens because workflows need to change, and people are not stupid.

Say employees get AI coding tools that make them twice as productive at the same quality. What happens?

Option A: they announce it. Management says, “we can’t give you a 2x raise, but we will call you an AI champion and expect 2x output from now on.”

Option B: they keep it quiet, deliver as before, and use the spare time for learning, fun, or personal projects.

Whenever I tell that story to audiences, I can easily see who is a manager and who is an employee: they either look annoyed or smile and nod vigorously.

Either way, dealing with that bottleneck takes patience and time for change management. Let’s put it aside for now.

The third bottleneck is more interesting. Most existing AI tools are not reliable and trustworthy. They can’t be easily controlled or audited in a way that complies not only with enterprise realities but also with common sense (and “log every single LLM request” is not a common-sense solution).

Fortunately, that bottleneck doesn’t require huge CapEx, OpEx, or time to solve. It requires a more liquid resource that is also easier to scale: knowledge, insights, and expertise.

If you know how to build AI tools that are reliable and trustworthy, and that can integrate well into enterprise environments with all their requirements, you can speed up adoption.

But nobody really knows how to build the best agents yet. Anthropic Mythos and OpenAI Codex may be the best evidence we have right now: one at the scary edge of security capability, the other turning code-driven agents into business infrastructure. And both will keep getting better in the next few years. That is the point. The frontier is moving, but we don’t know where we will end up.

So how can we do better than large companies, with fewer resources? Just think bigger and let the brightest AI engineers across companies work on problems together.

For example, take the problem of identifying the best architectures for personal and trustworthy AI agents, hand it out to a few hundred pairs of eyes at once in a fun and competitive way, and then see what best solution emerges.

So here is my take and bet on unlocking AI adoption around the world. Attack the bottleneck that can be solved with knowledge and insights by making it easy for the smartest AI engineers to collaborate on the same challenge. Then review results and share insights.

This is what BitGN is about. And you can even take this recipe and repeat it on your own!

On April 11th we ran the BitGN PAC1 challenge, which involved 860 engineers from ~80 cities around the world. The best architectures that emerged? Code-driven AI agents with durable sandboxes, controlled gates, and indirect tool use.

And that was even before OpenAI started pitching Codex to business teams :)

As a community, we already know a lot about these architectures: strong and weak points, performance, and even source code. And so do you; that is captured in the insights and shared publicly:

• Codex-on-Rails: Code-Mediated Execution - insights, source code.

• Operation Pangolin: Checklist-Driven REPL Agent - insights, source code.

Obviously, we are not stopping there. Once insights are shared, engineers need time to study and internalize them, then apply them to the next challenge. So at the end of May we are going to run another challenge about Agentic Ecommerce (ECOM1). And sometime in the fall we will run PAC2, with more challenging tasks based on the adoption of PAC1-based agents by companies and the edge cases they discover.

Engineers are already working on improving their agents. At the time of writing, we have 908 interested engineers in 91 cities and 20 offline hubs. 970k Agentic Trials have been run and 35M AI Tool calls have been recorded.

So, who should get involved?

If you are an engineer, check out the PAC1 benchmark and join the next challenge. The platform is free, and it is one of the fastest ways to learn the best practices hands-on while working on a concrete business problem with a fast feedback loop. People learn fast! And yes, people near the top of the leaderboard tend to be noticed.

If you want to support the local ecosystem, you can sponsor a hub by hosting engineers for the next event in your city on May 30th. What is needed: a venue, a big screen or projector for leaderboards, a good Wi-Fi connection, and some snacks. Your own engineers get to meet other like-minded builders, compete in a friendly challenge format, and then share insights.

And if your company benefits from this kind of work and wants BitGN to keep improving as a global platform for collaborative innovation, we are open to recurring partners and sponsors. Monthly support funds benchmark design, infrastructure, publishing, and the work of turning raw challenge activity into useful insights. Sponsors can get visibility on the website, tailored briefings, or both.

If any of this sounds useful, reply to this post or reach out to me directly.