AI.Cruz - Learn AI Through Interactive Visualizations

RESEARCH

AI Tools for Researchers

LLMs accelerate literature search, summarization, and writing — but they hallucinate citations and cannot verify claims. Use them as drafting assistants, not authorities.

1AI-Powered Literature Tools

A new generation of AI-native tools sits between raw search engines and general-purpose LLMs. They are trained specifically on academic literature and can extract claims, synthesise evidence, and map citation networks in ways that keyword search cannot. These tools do not replace reading papers — they help you decide which papers to read.

Elicit
AI research assistant. Finds papers matching a research question, extracts key claims, populations, methods, and outcomes. Excellent for systematic literature reviews.

Consensus
Synthesises evidence across papers for a yes/no question. Useful for fact-checking claims like "does batch normalisation improve training stability?" with citations.

Connected Papers
Visualises a citation graph for any paper. Enter one seed paper, get a force-directed graph of related work. Maps a sub-field visually in under 5 minutes.

ResearchRabbit
Recommends papers based on your reading history. Tracks new publications from authors you follow. Known as "Spotify for papers".

Semantic Scholar
Free AI-powered index by Allen AI. Citation graph, influence scores, author profiles, and free paper alerts. No paywall.

Perplexity
General web search with inline citations. Useful for quick background checks. Always click through to verify — summaries can be wrong even when the citations are real.

2AI Writing Assistants

LLMs like Claude, GPT-4, and Gemini are powerful writing partners for researchers. They excel at improving prose clarity, restructuring arguments, generating first drafts of boilerplate sections, and explaining math in plain language. They fail at original claims, accurate citations, and anything requiring knowledge of work published after their training cutoff.

✓ Use AI for

✓ Improving sentence clarity and flow
✓ Drafting the related work outline
✓ Explaining your method to a non-expert
✓ Generating alternative phrasings for an abstract
✓ Summarising a paper you just read

✗ Never trust AI for

✗ Citation generation (it invents plausible-sounding but non-existent papers)
✗ Claims about recent work (post training-cutoff)
✗ Mathematical derivations (errors are confident and subtle)
✗ Experimental results (numbers are fabricated)

3Code Generation for Research

AI coding assistants — GitHub Copilot, Claude, GPT-4 — accelerate the implementation phase of research significantly. They are excellent for standard algorithms, boilerplate, test suites, and debugging error messages. They are unreliable for novel architectures, cutting-edge library APIs, and correctness-critical numerical code.

The best workflow: write pseudocode or a precise spec first, then use AI to draft the implementation, then review every line yourself. Never copy AI-generated numerical code (loss functions, gradient computations, custom CUDA kernels) without a unit test. AI-generated code is confident and wrong in subtle ways.

Data loading boilerplate
Dataset classes, DataLoader configs, augmentation pipelines — AI handles these reliably since the patterns are repetitive.

Experiment scaffolding
Argument parsers, logging setup, checkpoint saving, evaluation loops — generated quickly and reliably. Focus your manual effort on the core contribution.

Unit tests
AI is excellent at writing tests for pure functions. Pair AI-generated tests with your own edge-case tests for numerical code.

4The Critical Caveat: Always Verify

The single most dangerous AI research failure mode is citation hallucination: an LLM generates a paper title, authors, and venue that sound completely plausible but do not exist. Researchers have submitted papers with AI-generated references and been caught at review. The fix is simple: look up every citation you did not personally retrieve.

Verification checklist for any AI-generated reference: (1) does this paper exist on Google Scholar, arXiv, or Semantic Scholar? (2) are the authors and year correct? (3) does the paper actually say what the AI claims it says? All three must pass. Skipping step 3 is how reviewers catch "the citation exists but doesn't support the claim."

AI as accelerator

✓ Use AI to draft, then verify every claim
✓ Let AI surface candidates, then retrieve originals yourself
✓ Use AI for structure and clarity, own the content

Red flags to catch

✗ "Author et al., 20XX" — verify the year exists
✗ Too-perfect paper titles — check they are real
✗ Claims that match your argument too perfectly

Use Claude API to summarise a paper — with hallucination guards

# Requires: pip install anthropic
# Set ANTHROPIC_API_KEY environment variable

import os

SYSTEM = """You are a research assistant helping summarise academic papers.

STRICT RULES:
1. Only summarise what is in the provided text. Never add external facts.
2. If you are unsure about a claim, say so explicitly.
3. Do NOT generate citations or reference other papers.
4. Mark any inference with [inferred].
"""

def summarise_paper(paper_text: str, model: str = "claude-haiku-4-5-20251001") -> dict:
    """Summarise a paper text using Claude with hallucination guards."""
    try:
        import anthropic
        client = anthropic.Anthropic(api_key=os.environ.get("ANTHROPIC_API_KEY"))
        prompt = f"""Summarise the following paper excerpt. Provide:
1. Central claim (1 sentence)
2. Method summary (2-3 sentences)
3. Key result (1 sentence)
4. Main limitation stated by authors (1 sentence)

Paper text:
{paper_text}"""
        message = client.messages.create(
            model=model,
            max_tokens=512,
            system=SYSTEM,
            messages=[{"role": "user", "content": prompt}],
        )
        return {"success": True, "summary": message.content[0].text}
    except ImportError:
        return {"success": False, "error": "anthropic package not installed. Run: pip install anthropic"}
    except Exception as e:
        return {"success": False, "error": str(e)}

# Example paper excerpt (from the Attention Is All You Need abstract)
paper_excerpt = """
The dominant sequence transduction models are based on complex recurrent or convolutional neural
networks that include an encoder and a decoder. The best performing models also connect the encoder
and decoder through an attention mechanism. We propose a new simple network architecture, the
Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions
entirely. Experiments on two machine translation tasks show these models to be superior in quality
while being more parallelizable and requiring significantly less time to train.
"""

result = summarise_paper(paper_excerpt)
if result["success"]:
    print("=== AI Summary ===")
    print(result["summary"])
    print()
    print("IMPORTANT: Always verify — read the original abstract yourself to confirm accuracy.")
else:
    print("Demo mode (no API key):")
    print("Central claim: The Transformer architecture replaces recurrence with attention.")
    print("Method: Encoder-decoder with multi-head self-attention and position-wise FFN.")
    print("Key result: Superior translation quality with faster training.")
    print("Limitation: [not stated in excerpt]")
    print()
    print(f"API error: {result['error']}")