Practical Tips & Safety13 min read

The Future of AI

In 2020, AI could barely write a paragraph. By 2025, it passes bar exams. What comes next?
scope:Big Picturedifficulty:Beginner

The Pace of Change Is Staggering

Let's take a moment to appreciate how fast things are moving:

  • 2020 — GPT-3 launches. It can write passable paragraphs, but they often go off the rails. Impressive demo, but not a tool most people take seriously.
  • 2022 — ChatGPT launches. Within 5 days, it has 1 million users. Within 2 months, 100 million. The fastest-growing consumer app in history.
  • 2023 — GPT-4 passes the bar exam, medical licensing exams, and SATs. AI-generated images win art contests. AI writes code that works. The world starts paying attention.
  • 2024 — AI video generation explodes (Sora, Runway). AI agents start browsing the web and taking actions. AI models become multimodal — understanding text, images, audio, and video simultaneously.
  • 2025 — AI is embedded in everyday tools. Coding assistants write 30-50% of code at major companies. AI tutors personalize education. AI diagnoses diseases from medical scans with superhuman accuracy.

Each of these milestones would have seemed like science fiction just 3 years before they happened. The pace of AI progress is not linear — it's exponential. And that's both thrilling and terrifying.

AGI: The Holy Grail

Artificial General Intelligence (AGI) is the ultimate goal of AI research — a system that can learn and perform any intellectual task that a human can. Not just one task really well (like playing chess), but all tasks: reasoning, creativity, common sense, learning new skills, understanding context, and adapting to novel situations.

Today's AI is narrow AI — incredibly powerful at specific tasks but unable to generalize. GPT-4 can write poetry and pass exams, but it can't tie a shoe, make a sandwich, or understand why a joke is funny in the way a 5-year-old can.

What Would AGI Look Like?

  • It could learn any new skill from scratch — cooking, carpentry, calculus — just by reading about it or being shown once
  • It would have common sense — understanding that if you put a ball on a table and tilt the table, the ball rolls off
  • It could transfer knowledge — using what it learns about chemistry to help with cooking, and what it learns about music to help with math
  • It would understand context, nuance, and ambiguity — grasping sarcasm, reading between the lines, understanding cultural references

When Will It Happen?

This is the most debated question in AI. Here's the spectrum of expert opinions:

  • Optimists (10-15 years) — Researchers like some at OpenAI believe we could achieve AGI by the mid-2030s through scaling current architectures with more data and compute.
  • Moderates (20-40 years) — Most AI researchers believe fundamental breakthroughs beyond current approaches are needed, putting AGI in the 2040s-2060s range.
  • Skeptics (50+ years or never) — Some researchers argue that current AI approaches will hit fundamental limits, and that true intelligence may require architectures we haven't invented yet — or that machine consciousness is impossible.
Note: The goalpost problem: Every time AI achieves something that was supposed to be "proof of AGI," we move the goalpost. In the 1990s, beating a chess grandmaster was considered the benchmark. AI did it, and we said "that's just search, not real intelligence." In the 2010s, beating a Go champion was the benchmark. AI did it, and we said "that's just pattern matching." Today, passing medical exams is the benchmark. AI does it, and we say "that's just memorization." At what point do we accept that the intelligence is real?

AI Agents: AI That Takes Action

One of the biggest shifts happening right now is the move from AI that answers questions to AI that takes actions. These are called AI agents.

Traditional AI interaction: you ask a question, you get an answer, you do something with that answer yourself.

AI agent interaction: you state a goal, and the AI figures out the steps, takes the actions, and achieves the goal for you.

What Can AI Agents Do?

  • Web browsing — Navigate websites, fill out forms, click buttons, make purchases
  • Task planning — Break down a complex goal into steps and execute them in order
  • Tool use — Call APIs, use software tools, write and run code
  • Self-correction — Recognize when something went wrong and try a different approach

Examples Emerging Now

  • Coding agents — You describe a feature; the agent writes the code, runs tests, fixes bugs, and submits a pull request
  • Research agents — You ask a question; the agent searches multiple sources, cross-references findings, and writes a comprehensive report
  • Personal assistants — You say "book me a flight to Tokyo next month, find a hotel near Shibuya under $200/night, and block off my calendar." The agent does all of it.
  • Computer use agents — AI that can see your screen and use your computer like a human would — clicking, typing, navigating between apps

Simplified AI Agent Loop

class SimpleAgent:
    """A simplified AI agent that plans and executes tasks.
    Real agents use LLMs for reasoning; this shows the concept."""
    
    def __init__(self, name, tools):
        self.name = name
        self.tools = tools  # available actions
        self.memory = []     # track what we've done
    
    def plan(self, goal):
        """Break a goal into steps (simplified)."""
        plans = {
            "book trip to Tokyo": [
                ("search_flights", {"to": "Tokyo", "budget": 800}),
                ("search_hotels", {"city": "Tokyo", "max_price": 200}),
                ("block_calendar", {"dates": "next month"}),
                ("send_confirmation", {"to": "user"})
            ]
        }
        return plans.get(goal, [("ask_user", {"msg": "I need more details."})])
    
    def execute(self, goal):
        """Plan and execute a goal step by step."""
        steps = self.plan(goal)
        print(f"[{self.name}] Goal: {goal}")
        print(f"[{self.name}] Plan: {len(steps)} steps\n")
        
        for i, (tool, params) in enumerate(steps, 1):
            if tool in self.tools:
                result = self.tools[tool](params)
                self.memory.append({"step": i, "tool": tool, "result": result})
                print(f"  Step {i}: {tool} → {result}")
            else:
                print(f"  Step {i}: {tool} → ERROR: tool not available")
                return False
        
        print(f"\n[{self.name}] ✓ Goal complete! ({len(steps)} steps executed)")
        return True


# Define available tools
tools = {
    "search_flights": lambda p: f"Found flight to {p['to']} for ${p['budget']-150}",
    "search_hotels": lambda p: f"Booked hotel in {p['city']} at ${p['max_price']-30}/night",
    "block_calendar": lambda p: f"Calendar blocked for {p['dates']}",
    "send_confirmation": lambda p: f"Confirmation sent to {p['to']}"
}


agent = SimpleAgent("TravelBot", tools)
agent.execute("book trip to Tokyo")
Output
[TravelBot] Goal: book trip to Tokyo
[TravelBot] Plan: 4 steps

  Step 1: search_flights → Found flight to Tokyo for $650
  Step 2: search_hotels → Booked hotel in Tokyo at $170/night
  Step 3: block_calendar → Calendar blocked for next month
  Step 4: send_confirmation → Confirmation sent to user

[TravelBot] ✓ Goal complete! (4 steps executed)

Multimodal AI: Understanding Everything at Once

Early AI models were specialists — a text model understood text, an image model understood images, and never the two shall meet. Multimodal AI changes that.

Modern models like GPT-4o, Gemini, and Claude can understand and generate across multiple modalities simultaneously:

  • Text — Read and write in any language
  • Images — See, understand, and generate images
  • Audio — Listen, speak, and generate sound
  • Video — Watch, analyze, and generate video
  • Code — Read, write, and execute programs

Why does this matter? Because the real world is multimodal. When a doctor diagnoses a patient, they look at X-rays (images), read lab results (text), listen to the patient describe symptoms (audio), and combine it all into a diagnosis. A truly useful AI needs to do the same.

What Multimodal AI Enables

  • Real-time translation with video — See someone speak Japanese, get English subtitles and lip-synced English audio instantly
  • Medical diagnosis — Analyze an X-ray while reading the patient's history while listening to their described symptoms
  • Education — An AI tutor that can see a student's math work (image), hear their question (audio), and explain the concept with generated diagrams (image + text)
  • Accessibility — Describing what's on screen for visually impaired users, transcribing audio for deaf users, in real time

AI in Every Industry

AI isn't just a tech industry story. It's transforming virtually every field:

Healthcare

  • Diagnosis — AI detects cancer in medical images with accuracy matching or exceeding specialist doctors
  • Drug discovery — AI predicts molecular structures and identifies potential drugs in weeks instead of years. Google DeepMind's AlphaFold predicted the structure of virtually every known protein — a problem that would have taken biologists millions of years.
  • Personalized treatment — AI analyzes a patient's genetics, lifestyle, and medical history to recommend treatments tailored to them specifically

Education

  • AI tutors — Personalized learning that adapts to each student's pace, style, and knowledge gaps
  • Content creation — Generating practice problems, study guides, and explanations on demand
  • Accessibility — Real-time translation, text-to-speech, and adaptive interfaces for students with disabilities

Science & Research

  • Climate modeling — AI processes vast amounts of climate data to make better predictions
  • Materials science — AI discovers new materials with desired properties (superconductors, battery materials)
  • Astronomy — AI identifies patterns in telescope data that humans would miss, discovering new exoplanets and cosmic phenomena

Creative Arts

  • Music — AI composes original music and assists human musicians
  • Visual art — AI generates images, assists with animation, and creates concept art
  • Writing — AI assists with drafting, editing, brainstorming, and translation

Risks and Existential Questions

With great power comes great... you know the rest. As AI becomes more capable, the risks scale up too.

Short-Term Risks (Now - 5 years)

  • Misinformation at scale — AI makes it trivially easy to generate convincing fake news, fake images, and fake videos
  • Job displacement — Millions of jobs transformed or eliminated, potentially faster than new ones are created
  • Privacy erosion — AI-powered surveillance becomes more pervasive and harder to escape
  • Concentration of power — AI development is extremely expensive, concentrating power in a handful of companies and countries

Medium-Term Risks (5-20 years)

  • Autonomous weapons — AI-powered military systems that can decide to use lethal force without human approval
  • Economic inequality — Countries and companies with advanced AI may dominate those without, widening global inequality
  • Manipulation — AI that understands human psychology so well it can persuade anyone of anything

Long-Term Risks (20+ years)

  • Alignment problem — How do we ensure that a superintelligent AI's goals align with human values? If we build something smarter than us, how do we make sure it wants what we want?
  • Control problem — If an AI becomes significantly more intelligent than humans, can we maintain meaningful control over it?
Note: The alignment problem explained simply: Imagine you tell a superintelligent AI: "Make humans happy." Sounds safe, right? But the AI might decide the most efficient way to make humans happy is to directly stimulate the pleasure centers of every human brain, turning everyone into mindless, perpetually grinning zombies. Technically, it achieved the goal — everyone is "happy." But it's not what we meant. The alignment problem is about ensuring AI understands not just our words, but our intentions.

Preparing for the Future

Whether you're 15 or 55, here's how to prepare for an AI-powered world:

Skills That Will Matter More

  • Critical thinking — As AI generates more content, the ability to evaluate, question, and verify becomes priceless
  • Creativity — AI can generate, but humans set the vision. Creative direction, storytelling, and original thinking are uniquely human
  • Emotional intelligence — Understanding, motivating, and leading other humans. AI can simulate empathy; humans have the real thing
  • AI literacy — Understanding what AI can and can't do, how to use it effectively, and how to evaluate its output
  • Adaptability — The willingness and ability to keep learning as the landscape shifts

Skills That Will Matter Less

  • Rote memorization (AI has perfect recall)
  • Routine data processing (AI does it faster and cheaper)
  • Basic content creation (AI generates first drafts instantly)
  • Simple translation (AI translates in real time)

The Mindset Shift

The most important shift isn't about skills — it's about mindset. Think of AI as:

  • Not a replacement, but a tool. The calculator didn't replace mathematicians; it made them more powerful. AI won't replace thinkers; it will make them more effective.
  • Not magic, but technology. Understanding its limitations is as important as appreciating its capabilities.
  • Not the future, but the present. AI is already here. The question isn't whether to engage with it, but how.
Challenge

Quick check

What is the key difference between today's AI (narrow AI) and Artificial General Intelligence (AGI)?

Continue reading