Meet the AI Platforms11 min read

Claude

An AI built from day one to be helpful, harmless, and honest — meet Anthropic's Claude
scope:Foundationaldifficulty:Beginner

Born from a Safety Mission

In 2021, a group of researchers left OpenAI because they believed the AI industry wasn't taking safety seriously enough. They founded Anthropic, and their first major product was an AI assistant called Claude.

The name isn't random. It's a nod to Claude Shannon, the father of information theory — the mathematician who figured out how to measure and transmit information, laying the groundwork for everything from the internet to, well, LLMs.

From day one, Anthropic's goal wasn't just to make the smartest AI. It was to make an AI that was helpful, harmless, and honest — the "3 H's" that guide everything Claude does.

Who Founded Anthropic?

Dario Amodei (CEO, former VP of Research at OpenAI) and Daniela Amodei (President, former VP of Safety at OpenAI) lead the company. They brought along several top researchers who shared their concerns about building increasingly powerful AI without adequate safety measures.

Anthropic isn't anti-AI. They believe powerful AI is coming regardless, and they'd rather be at the frontier building it safely than watching from the sidelines. They call this approach "responsible scaling" — push the boundaries, but with guardrails.

Constitutional AI: Teaching an AI Values

Most AI companies use RLHF (Reinforcement Learning from Human Feedback) — hire humans to rate AI responses, then train the model to produce responses humans rate highly. This works, but it has problems:

  • Humans are expensive and slow
  • Different humans have different values
  • It's hard to be consistent across millions of ratings

Anthropic invented something different: Constitutional AI (CAI). Here's how it works:

  1. Write a constitution — A set of principles like: "Choose the response that is most helpful while being harmless," "Don't help with illegal activities," "If unsure, say so honestly."
  2. Self-critique — The AI generates a response, then critiques its own response against the constitution. "Does this response violate any principles? How could it be improved?"
  3. Self-revise — The AI rewrites its response based on its own critique.
  4. Train on the improved responses — The model learns to produce the revised (better) version directly.

The beauty of this approach: the AI's values are explicit and readable. You can look at the constitution and see exactly what principles the AI was trained to follow. It's more transparent than a black box of human ratings.

Using Claude via the API

import anthropic


client = anthropic.Anthropic()  # uses ANTHROPIC_API_KEY


# Claude uses a messages API similar to ChatGPT
message = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=200,
    system="You are a helpful CS tutor. Be concise.",
    messages=[
        {"role": "user", "content": "What makes Claude different from ChatGPT?"}
    ]
)


print(message.content[0].text)


# Claude's strengths:
# - 200K token context window (read entire books)
# - Constitutional AI for safety
# - Strong at long document analysis
# - Careful about what it doesn't know
# - Less likely to hallucinate confidently
Output
Three key differences:
1. Safety approach: I'm trained with Constitutional AI, which
   means I have explicit principles I follow, rather than just
   learning from human ratings.
2. Long context: I can process up to 200K tokens — about a
   full novel — in a single conversation.
3. Honesty: I'm designed to say "I don't know" rather than
   confidently making something up.
Note: The Model Lineup: Claude comes in three sizes, each named for its role: Haiku (small and fast — great for quick tasks and chat), Sonnet (the balanced workhorse — great for most tasks), and Opus (the most powerful — best for complex reasoning, analysis, and nuanced writing). Think of it like cars: Haiku is a zippy scooter, Sonnet is a reliable sedan, and Opus is a luxury sports car.

What Claude Is Great At

  • Long documents — With a 200K token context window, Claude can read and analyze entire books, codebases, legal contracts, or research paper collections in a single conversation. Ask it to find contradictions in a 100-page contract? No problem.
  • Coding — Claude excels at writing, reviewing, and debugging code. It's particularly good at understanding large codebases and explaining complex code.
  • Careful reasoning — Claude is trained to think step by step and show its work. It's less likely to jump to conclusions and more likely to consider edge cases.
  • Saying "I don't know" — Unlike some models that confidently make things up, Claude is designed to express uncertainty when it's unsure. It will say "I'm not confident about this" rather than inventing a plausible-sounding falsehood.
  • Following nuanced instructions — Claude is strong at understanding complex, multi-part instructions and following them precisely.

Claude vs. ChatGPT: How Do They Compare?

Neither is strictly "better" — they have different strengths:

  • Context window — Claude: 200K tokens. ChatGPT (GPT-4 Turbo): 128K tokens. Claude wins for long documents.
  • Ecosystem — ChatGPT has plugins, DALL-E image generation, code interpreter, web browsing, and a huge app store of custom GPTs. Claude is more focused on the core chat experience.
  • Safety approach — Claude uses Constitutional AI (explicit principles). ChatGPT uses RLHF (human ratings). Different philosophies, both effective.
  • Hallucination — Claude tends to be more cautious, sometimes declining to answer rather than risk being wrong. ChatGPT is more willing to attempt an answer but may hallucinate more.
  • Writing style — Claude tends to be more measured and nuanced. ChatGPT tends to be more enthusiastic and verbose. Personal preference varies.

Anthropic's Bigger Vision

Anthropic isn't just building a chatbot. They're working on some of the hardest problems in AI:

  • Interpretability — Understanding what's happening inside neural networks. Most AI models are "black boxes" — we know what goes in and what comes out, but not why. Anthropic is leading research into opening that black box.
  • Responsible Scaling Policy — A framework for deciding when it's safe to train more powerful models and when to pause. As models get more capable, the stakes get higher.
  • Alignment research — Making sure AI systems do what humans actually want, not just what they literally asked for. The difference between "clean my room" (put everything away neatly) and a misaligned interpretation (throw everything in the trash).

Whether Claude or ChatGPT or Gemini ends up being the "best" AI assistant, the safety research Anthropic is doing benefits the entire field. If they figure out how to make AI provably safe, everyone wins.

The Three H's in Practice

Every Claude response is guided by three principles:

  • Helpful — Actually answer the question, provide useful information, and complete the task. Don't be so cautious that you're useless.
  • Harmless — Don't help with dangerous or illegal activities. Don't produce harmful content. Don't manipulate or deceive.
  • Honest — Tell the truth. Express uncertainty when uncertain. Don't pretend to be human. Don't make up facts.

These three principles can sometimes conflict. What if being maximally helpful means answering a dangerous question? Claude is trained to balance all three — being as helpful as possible while staying within the bounds of harmlessness and honesty.

Challenge

Quick check

What is Constitutional AI, and how does it differ from RLHF?

Continue reading