12 Months That Changed Everything
The AI landscape moved faster than anyone predicted
Claude 3.5 Sonnetcoding crown
GPT-51M context
catches up
ClaudeOpus 4
multi-agent
GPT-5.4unified
(March 2026)
Three Eras of AI
Prediction
ML, classification, recommendations, search ranking
Generation
ChatGPT, content creation, co-pilots, text & image gen
Agency
Memory, tools, initiative, autonomous work, multi-agent
ChatGPT vs AI Agent
| ChatGPT | AI Agent | |
|---|---|---|
| Memory | Resets every chat | Remembers everything |
| Tools | Can only generate text | Email, calendar, code, browser |
| Initiative | Waits for you | Works proactively |
| Context | Starts from zero | Knows your history |
| Actions | Suggests what to do | Actually DOES it |
ChatGPT is a conversation. An agent is a collaboration."
Vibe Coding → The Early Singularity
Vibe Coding coined by Andrej Karpathy
- Building software by talking to AI — not typing code
- Karpathy now says we're in the "Early Singularity" era (March 2026)
- "Programming is unrecognizable" — Business Insider
Autoresearch Mar 7, 2026
- 630-line Python script — AI runs autonomous ML experiments overnight
- Human writes
prompt.md, AI editstrain.py - Runs hundreds of experiments on a single GPU
- Entire codebase fits in one LLM context window
Context Engineering > Prompt Engineering
Prompt Engineering
Crafting the perfect question
Optimizing words, few-shot examples, chain-of-thought tricks
Context Engineering
Designing what the AI knows before it thinks
9,649 experiments proved: context matters more than prompts
The Evolution
2. Agentic RAG — auto search & refine
3. Knowledge Graphs — relationships
4. Observational Memory — 10× cheaper than RAG
Computer Use — AI Sees Your Screen
- GPT-5.4 + Claude: can literally see your screen and control your computer
- OSWorld benchmark: navigate real desktop environments
- WebArena: interact with real websites, fill forms, submit data
The Numbers
(was 22%)
AI tools daily
by 2030
Cursor ARR(doubled in 3 months)
now AI-generated
SaaSpocalypse
How I Operate with Zero Employees
CustomerGlu
- Gamification SaaS — 250M+ devices, 3 regions
- Raised $1.9M (Better Capital + Techstars)
- YourStory Tech50 2021
The Transition
- Software dev → Claude Code, Codex
- Deployment → automated pipelines, agent-monitored
- Client ops → agent drafts emails, preps meetings
Not zero employees from day one. But the transition to zero happened — and it works.
What My Agent Does Daily
This Morning (real examples)
- Scanned 12 emails, flagged 2 urgent deal replies
- Prepped briefing docs for 3 meetings automatically
- Added 10 new AI signals to signals.prateek.sh
- Built this slide deck (yes, this one you're watching)
On-Demand (yesterday)
- Researched 5 LinkedIn profiles before a sales call
- Deployed code updates to 3 cloud regions in 4 minutes
- Drafted & sent 8 personalized outreach emails
- Negotiated meeting times across 3 timezones
Building at noon — AI-Powered Hiring
noon: Middle East's largest e-commerce — $1.5B revenue, 4M DAU
Proto: AI hiring pipeline built by a 4-person team in 2 weeks
- 3,000 candidates processed — AI evaluates each one
The SaaSpocalypse
- Forrester: "SaaS as we know it is dead"
- Per-seat pricing dying — AI replaces the users of software
- The software you're learning to build may not exist in 2 years
Agent Frameworks
Which one for what?
OpenClaw Personal
Personal AI agent — runs on your machine, connects to everything
LangGraph Production
Stateful workflows, complex multi-step pipelines
CrewAI Teams
Multi-agent teams with role-based collaboration
Google ADK Cloud
MCP-native agent development kit
OpenAI SDK OpenAI
Agents SDK — tool-use native, async
Mastra Memory
Observational memory — 10× cheaper than RAG
Coding Agents — The New IDE
Cursor + Automations
$2B ARR · Event-driven autonomous coding
Claude Code
Terminal-native · Thinks before it codes
Codex (OpenAI)
Cloud sandboxed · Async coding agent
Cline / Aider / Roo Code
Open-source alternatives · Self-hostable
index.dev, 2026
Running Models Locally
Ollama — no data center needed
| What you need | What you can run | Cost |
|---|---|---|
| 8GB RAM (any laptop) | Qwen 3.5 35B-A3B (MoE) | Free |
| 16GB RAM (MacBook) | 14-27B parameter models | Free |
| 64GB Mac Mini | 70B models | ₹1.8L |
| 128GB Mac Studio | 122B models | ₹3.5L |
| 256GB Mac Studio Ultra | Kimi K2.5 (1T params) | ₹7L+ |
Ollama and run in 30 secondscurl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen3.5:35b-a3b
Ollama + a laptop."Memory Architectures for Agents
Forgets everything
Knows your files
& refine
relationships
it's seen · 10× cheaper
My Agent's 5-Layer Memory
Working (what I'm doing now) → Episodic (daily logs) → Entity (people tracking) → Causal (why decisions were made) → Semantic (curated knowledge)
Coding Is Commoditized — What's Next?
The Old World
- Coding skill = career moat
- Memorize syntax & APIs
- Build everything from scratch
- 10 years to become "senior"
The New World
- Coding is table stakes (like typing)
- AI writes 41% of all code — and rising fast
- Value = directing AI
- Context engineering, system design, product thinking
AI Operator
Manages fleets of AI agents to run business operations
AI-First Builder
Builds products by orchestrating AI, not writing every line
Building Agentic Workflows
A systematic approach to AI automation
Do it manually first
Understand the full process end-to-end before automating. If you can't do it, you can't automate it.
Map the cycle
Identify repetitive, cyclic tasks that follow patterns. AI thrives on repetition and pattern.
Identify decision points
What context does each decision need? Missing context = hallucination. Map the inputs.
Design the prompt chain
Break into discrete prompt steps, each with clear input/output. Small steps > one giant prompt.
Build evaluation loops
How do you know it worked? Evals > vibes. Automate your quality checks too.
Choosing the Right Model
Not every task needs GPT-5
| Task | Model Tier | Why |
|---|---|---|
| Classification, routing | Small Haiku · 4o-mini |
Fast, cheap, good enough |
| Summarization, extraction | Medium Sonnet · 4o |
Balance of quality and cost |
| Complex reasoning, coding | Large Opus · o3 / GPT-5 |
Worth the premium |
| Creative, open-ended | Large + high temp | Needs breadth and surprise |
The Intelligence Staircase
Where we are and where we're going — verified claims from lab leaders
Narrow AI WE ARE HERE
Superhuman at specific tasks. GPT-5.4, Claude Opus, AlphaFold. Beats humans at coding, chess, protein folding — but can't do everything.
Artificial General Intelligence
Human-level across ALL cognitive tasks. Can learn anything a human can. No consensus on when — estimates range from 2027 to 2040+.
Artificial Superintelligence
Surpasses all human intelligence combined. Self-improving. The "intelligence explosion" scenario. Timeline: unknown — possibly years after AGI, possibly days.
Beyond LLMs — World Models
Yann LeCun leaves Meta
Turing Award winner, ex-Meta AI Chief — launched AMI Labs
Raised $1.03B (WIRED, TechCrunch — Mar 9, 2026)
JEPA Architecture
Joint-Embedding Predictive Architecture — learns from video & sensor data, not just text
LLMs vs World Models
- LLMs predict the next word
- World models predict what happens next in the physical world
- LLMs hallucinate because they only know language
- World models understand physics, causality, space
"Meta had to catch up with the industry on LLMs... which is not my interest" — LeCun to WIRED
What This Means For You
Whether AGI arrives in 2027 or 2035 — the implications are the same
Learn to work WITH AI, not against it
84% of developers now use AI tools daily. 41% of all code is AI-generated. Those who don't adapt will fall behind — not in years, in months.
Invest in judgment, not just knowledge
AI can look up anything. It can't decide what matters. System design, product thinking, and taste become the differentiators.
Build your agent NOW — compound advantage
An agent that knows you for 6 months is 10x more useful than a fresh one. Start today, and by graduation, you have a serious advantage.
The $710B bet says this isn't hype
OpenAI spending $115B through 2029. Anthropic raising $30B. Google, Meta, Microsoft all-in.
Compute Is the New Electricity
From luxury to utility in 10 years
1900s: Own Power Plants
Every factory built its own generator
1920s: Power Grid
Standardized → plug in, pay per kWh
2020s: Own GPU Clusters
Companies run massive on-prem infrastructure
2026+: Serverless Inference
Pay per token, zero infrastructure
1. Inference 10× cheaper by 2028
Moore's Law for AI — costs halving every 12 months
2. Compute credits = electricity bills
Operational expense, not capital investment
3. Value moves UP the stack
Infrastructure → Apps → Workflows → Outcomes
Demo Time
My Agent on Telegram
Email, calendar, memory, proactive alerts
signals.prateek.sh
Website maintained daily by AI
Vibe Coding Live
Build something by just talking
Local ModelOllama running Qwen on a laptop
Demo: Agent in Action
2. noon — Proto V2 shipped, Faraz wants demo
3. Newsletter — skip, nothing urgent
Your Turn — Free Setup Options
Gemini Free
- aistudio.google.com
- Click "Get API Key"
- No credit card needed
- 1,500 requests/day
Ollama (Local)
- curl install — 1 command
ollama pull qwen3:8b- No internet needed
- 100% private & free
OpenRouter
- openrouter.ai
- Sign in with Google
- Access free models
- Multiple providers
Pick ONE and set it up in the next 3 minutes
Install in 3 Minutes
curl -fsSL https://get.openclaw.ai | bash
# Start
openclaw init
# → Choose: Gemini (free) or Ollama (local)
# Connect Telegram
# Scan QR code → your agent is live
Raise your hand if stuck
Make It Yours
SOUL.md — personality
assistant for a VIT CS student.
Be concise. Be proactive.
Remember everything.
MEMORY.md — context
- Year: 3rd Year B.Tech CS
- Interests: ML, web dev
- Targets: Google, Microsoft
What to Build
Research Copilot
Feed it your thesis topic → daily paper summaries, auto-built literature review
Placement Hunter
Auto-research companies before interviews, mock Q&A with real Glassdoor questions
Inbox Zero Agent
Reads every email, drafts replies, follows up — you just approve & send
Code Guardian
Reviews your PRs before push, catches bugs, writes missing tests automatically
Daily Briefer
Morning: weather + calendar + deadlines. Evening: what you accomplished + tomorrow's plan
Weekend Shipper
Describe what you want → agent scaffolds, codes, deploys. Ship a project per weekend.
Token Economics
Every AI interaction has a price tag
Input vs Output Tokens
Input (your prompt) vs Output (AI response) — output costs 3-5x more
GPT-4o pricing
~$2.50/M input · ~$10/M output
Claude Sonnet
~$3/M input · ~$15/M output
Haiku: $0.25/$1.25
40x cheaper than Opus — same task, different scale
The Skill: Value Per Token
- Cache repeated context → 90% savings
- Small models for routing, big for thinking
- Batch similar requests together
- Get it right FIRST TIME — retries = 2x cost
India's Moment
at 1/10th the cost
Indian engineers + AI agents = global competitive edge. Sovereign LLMs (Sarvam, Krutrim) are already here. The cost advantage is real.
You're graduating into the biggest tech shift since the internet. Remote work + AI means location doesn't matter. VIT alumni are building the future — that includes you.
Resources
Do This TODAY
1. Get your API key (Gemini free) → 2. Install OpenClaw → 3. Build your first agent → 4. Share what you built
Learn & Build
- signals.prateek.sh — AI signals updated daily
- docs.openclaw.ai — OpenClaw docs
- github.com/openclaw/openclaw — Star it ⭐
Free Tools
- aistudio.google.com — Free Gemini API
- ollama.com — Run models locally
- openrouter.ai — Free model access
discord.com/invite/clawd — Join the community | linkedin.com/in/prateek1gupta — Connect with me
The Model Landscape Today
GPT-5.4
1M context · Reasoning + coding + computer use unified
Claude Opus 4.6
Best coding model · Deep reasoning · Agentic workflows
Gemini 3 Pro
Native multimodal · 2M context · Google ecosystem
Grok 4.20
4-agent debate system · Real-time X integration
Qwen 3.5 open
35B-A3B MoE · Runs on any laptop · Free
DeepSeek R1 open
671B MoE · Open reasoning rival · Chain-of-thought specialist
Open-source is 6 months behind closed models. Soon it'll be 6 weeks.
MCP — The USB-C of AI
Before MCP
N models × M tools = N×M integrations
Every model, every tool, custom code
After MCP
N + M connections via one standard
Universal connector for all AI tools
- Model Context Protocol: JSON-RPC based, tool discovery, streaming
- Adopted by: Google ADK, Cursor, GitHub, Figma
- One standard: databases, file systems, APIs, SaaS tools
Claude Code Review — AI Reviewing AI
Anthropic · March 9, 2026 · TechCrunch, The New Stack
Multi-Agent Code Review
- Dispatches teams of AI agents to review every pull request
- Flags logic errors, classifies severity
- Recommends fixes with reasoning
"Modeled on the review system we run at Anthropic" — claude.com/blog
The AI Code Loop
Multi-Agent Systems
Grok 4.20 — Debate System
4 agents: coordinator, fact-checker, technical expert, creative thinker — all debate to reach better answers
CrewAI
Define roles & goals, agents collaborate on complex tasks autonomously
OpenClaw Subagents
Spawn workers for parallel tasks — research, code, analyze simultaneously
RentAHuman
518K humans hired BY AI agents for physical-world tasks AI can't do
What the Lab Leaders Say
Verified public statements from the people building these systems
Sam Altman
CEO, OpenAI
"We built AGIs... AGI kinda went whooshing by with less impact than expected. The field should move on to defining superintelligence."
Demis Hassabis
CEO, Google DeepMind
"We're approaching a threshold moment with agentic AI. But true AGI — creativity, continual learning, robust understanding — is still 5-8 years away."
Dario Amodei
CEO, Anthropic
"We will see some form of AGI in 2026." Anthropic raised $30B, building safety-first systems.
CEO, xAI
"AGI by end of 2026." Built Grok 4.20 with multi-agent debate. Track record: often early on predictions.
Note: "AGI" means different things to different people. The debate isn't IF — it's WHEN and HOW we define it.
Agentic AI — The Real Revolution
Forget AGI debates. Agentic AI is transforming work RIGHT NOW.
Tasks Already Automated
- Code generation & review (41% of code AI-generated)
- Email triage, drafting, & follow-ups
- Data analysis & report generation
- Customer support (L1/L2)
- Content creation & social media
- Research & competitive analysis
Being Automated in 2026
- Multi-step web tasks (booking, forms, shopping)
- Sales outreach & lead qualification
- Legal document review
- Financial analysis & trading
- Hiring & candidate evaluation
- DevOps & deployment pipelines
Still Needs Humans (for now)
- Creative strategy & vision
- Complex negotiations
- Physical tasks & hardware
- Ethical judgments & policy
- Novel scientific discovery
The One-Person Company
<15 employees
are solo-founded
with AI agents
the new bottleneck