The Future of AI: Where It Came From, Where It Is, and Where It's Going

Five years ago, AI struggled to form coherent sentences. Today it's reaching beyond human level on a widening range of tasks. That gap, five years, is the whole story of this article.

I gave a talk on this recently to a room of engineers, and I want to write down the version of it that lives outside a slide deck. Not as a hype piece, and not as a doom piece either. Just an honest walkthrough of what's happening, why it's happening this fast, and what it means for people who build technology for a living.

Why now

The length of coding task AI can complete has been doubling roughly every 4 months since 2023, according to METR's time horizon research. Extrapolate that curve and AI is expected to handle week-long software engineering tasks by September 2026. A new frontier model launch pushes the state of the art forward nearly every month.

This isn't a discussion about the distant future. AI is already changing how software is built, secured, and operated. If you work in technology, your decisions this year, what you adopt, what you automate, what you secure, and what you govern, will shape how competitive and resilient your organization is in an AI-driven industry.

What AI can already do

In just five years, AI has gone from struggling with simple text to generating images, video, software, and human-like speech. Progress that once took decades is now happening in months.

A quick tour of what's already routine:

• Image: Photorealistic images, illustrations, and product mockups from a single text prompt. The first text-to-image model, back in 2015, produced 32x32 pixel outputs. Today's models produce gallery-quality art.
• Video: Four years ago, AI video generation didn't exist in any usable form. Now realistic video from a prompt is a normal creative tool.
• Code: Describe an application in plain English and get a working prototype in minutes. Claude, GPT-5, Gemini, GLM, Kimi, and Composer all ship production-grade code today.
• Design: Presentations, UI mockups, marketing materials, and illustrations on demand. What used to take a design team days now takes minutes.
• Audio: Natural voice generation, translation, transcription, and realistic voice cloning.
• Reasoning: Analyzing documents, solving complex problems, and assisting with decisions that used to require a room full of specialists.
• Agents: Completing multi-step tasks across tools with minimal supervision. This is the newest and, in my view, the most consequential capability.

The shift from tools to agents

Until recently, most AI systems behaved like tools. You asked a question, you got a single response. Predictable, scoped, passive. Think of a calculator, or ChatGPT answering a coding question. The human did all the orchestration.

Today we're seeing the rise of AI agents. You give them a goal, and they plan, take multiple steps, use other software, and report back with a result. It's the difference between a calculator and a junior teammate who is autonomous over a defined task horizon.

Real agents shipping right now:

• Claude Code writes and refactors software autonomously.
• ChatGPT Agent researches, browses, writes reports, and executes workflows.
• Copilot assists across email, documents, meetings, and business apps.

The key difference is autonomy. Tools answer questions. Agents pursue goals. That distinction has profound implications for productivity, software engineering, cybersecurity, and governance, because agents don't just generate content anymore. They perform work.

The evolution: ANI, AGI, ASI

AI capability is usually described in three stages.

Stage 1: ANI, Artificial Narrow Intelligence. This is where we are today. AI designed to perform specific tasks exceptionally well, but unable to generalize the way humans do. Spam filters, face ID, chess engines, credit scoring models, Siri, Alexa, Google Translate: all narrow. Nearly all production AI in the world right now is ANI, no matter how impressive it looks.

Stage 2: AGI, Artificial General Intelligence. Plausibly within this decade. An AI system capable of learning and performing almost any intellectual task a human can, not just one task, but many. Think of a single system that can do your job and mine, autonomous scientific research at human level, or self-improving codebases running end to end operations.

Stage 3: ASI, Artificial Superintelligence. Still speculative. AI that vastly surpasses human intelligence across every field, including scientific creativity, general wisdom, and social skill. Capabilities beyond human comprehension, scientific breakthroughs at machine speed, and the ability to solve problems humans can't even formulate yet.

How AGI could become ASI: the intelligence explosion

The idea, going back to I.J. Good in 1965 and revisited by Nick Bostrom in 2014, goes like this. Once AI becomes as capable as top human AI researchers, it can start contributing to AI research itself. Instead of only using AI, researchers use it to discover better algorithms, write code, and run experiments. Each new generation helps design an even more capable successor. If that feedback loop holds, capability could rapidly surpass human intelligence across many domains.

Why it could be fast: designing better AI is largely an intellectual task, and once AI performs that task as well as human researchers, it can help accelerate its own improvement.

Why it's uncertain: no one knows whether the feedback loop continues indefinitely. Progress could slow because of limits in data, compute, algorithms, or something we haven't hit yet. Experts genuinely disagree on both the likelihood and the timeline, and I think that honesty matters more than a confident prediction either way.

The three engines behind all of it

AI capability doesn't improve by magic. Every major breakthrough comes from advancing one or more of three things:

1. Data, the fuel. More high-quality experience means the model learns more patterns about the world. Think of it as giving a new employee more experience.
2. Compute, the engine. More computing power trains larger models faster. Think of it as giving that employee a faster computer.
3. Algorithm, the recipe. Better learning methods extract more intelligence from the same data and hardware. Think of it as teaching that employee a better way to solve problems.

These three reinforce each other. More data needs more compute to process. Better algorithms extract more from the same data and compute. The result is capability growth that's faster than any single trend would predict on its own.

What's at stake

The upside

Dario Amodei, CEO of Anthropic, put it directly in his 2024 essay "Machines of Loving Grace": powerful AI could compress a century of progress into a single decade, across disease, poverty, and governance.

A few concrete threads of that promise:

• Health: faster drug discovery and medical research, including treatments for diseases that have been historically underfunded.
• Prosperity: affordable AI expertise could give small businesses access to legal, financial, and technical advice that used to be out of reach.
• Government and policy: AI could help test policies before they're implemented, improving planning before real people are affected.
• Education: personalized tutoring for every child, in their own language, at their own pace, regardless of where they live.

The promise of AGI isn't just smarter technology. It's making expertise, education, healthcare, and innovation accessible to people who have never had access to them before.

The risks

None of this is free of danger, and treating it that way would be dishonest. There are three broad families of risk worth taking seriously:

1. Terrorism, misuse by non-state actors. AI could lower the expertise required to develop cyber, biological, or chemical threats. A capable model in the wrong hands is a public safety problem on a new scale.
2. Authoritarianism, concentration of power. If only a handful of companies or governments control the most powerful AI systems, they could gain outsized influence over information, surveillance, and decision-making.
3. Loss of human control. As AI systems become more capable, reliably directing, predicting, and supervising them becomes harder, not easier.

These are risks, not predictions. The goal isn't fear. It's preparation, understanding the risks well enough to actually manage them while still capturing the benefits.

What you can actually do about it

You don't have to be an AI researcher to shape where this goes. Three paths, and you can start any one of them this month.

1. Learn continuously. Stay informed about what's changing, on both the capability side and the risk side. Experiment with AI tools responsibly. Knowledge is the foundation for making good technology decisions, and it decays fast in this field if you're not deliberately refreshing it.

2. Use AI responsibly in your own work. Think critically about where it adds real value, where it introduces risk, and what controls need to exist before you deploy it. Start by identifying one repetitive task AI could take off your plate, and think through the security, privacy, and governance implications before you automate it.

3. Help build a culture of responsible adoption around you. Share what you learn. Ask good questions instead of either blind enthusiasm or blind dismissal. Support thoughtful experimentation instead of either banning AI outright or adopting it without any guardrails.

Five things to remember

If you take nothing else from this:

1. AI is moving faster than most people think. The length of task AI can handle doubles roughly every 4 months, and traditional benchmarks are becoming saturated, which means we need new ways to even measure progress.
2. Today's AI is narrow, but AGI may be closer than expected. It's plausibly within this decade, and the jump from AGI to ASI could be much faster still, through the intelligence explosion feedback loop.
3. Three engines drive everything: data, compute, and algorithms, reinforcing each other and accelerating progress beyond what any single trend would predict.
4. AI presents extraordinary opportunity and serious risk, simultaneously. A century of progress could compress into a decade. The goal is maximizing the benefit while actually managing the risk, not picking one and ignoring the other.
5. You can shape this starting now. The future of AI won't be decided only by the labs building frontier models. It will be shaped by the decisions the rest of us make about how we build, adopt, and govern it, starting today.

---

If this resonated, I'd genuinely like to hear what stage of the ANI to AGI to ASI progression you think will matter most in your own field over the next five years, and why.