Narzędziownik AI 2.0 Reloaded - sesja 1

Overview

Session 1 of “Narzędziownik AI 2.0 Reloaded” (Toolmaker AI series, Tomek Turba) is an introductory, practical training. It covers:

• Core AI concepts and popular models
• Prompt engineering techniques and demos
• Security, ethics, and risk management
• Course logistics (14 sessions total)

Emphasis is hands-on: choosing models/tools, crafting prompts, reducing hallucinations, deploying safe RAG/local setups, and integrating AI into business workflows.

Logistics / Product of the training series

• 14 scheduled sessions (recordings + slides emailed after each meeting)
• Certificate + test (Polish & English) with CPE points
• Minibook listing ~120 tools
• Community/communication: Sekurak.pl resources, Securitum company trainings, Discord channel (Sekurak), mailing list
• Future demos & labs: local model setups, building assistants, agents, automation, multimedia creation, Microsoft Copilot, SEO/AEO/GEO positioning

Key technological concepts explained

• Model families and architectures: Transformer, LLMs (large language models), LRMs (deep reasoning/long-context models)
• Core mechanisms: embeddings, attention, tokenization, context window/tokens
• RAG (Retrieval‑Augmented Generation) — local/document RAG for safer corporate answers
• Embedding databases / vector DBs (Qdrant mentioned) for document search/RAG
• Hardware & scaling: GPU (NVIDIA) vs CPU; clusters/supercomputers required to train very large models
• Local vs cloud models: tradeoffs between privacy/safety and knowledge/freshness/compute
• Model marketplaces and datasets: Hugging Face and model/dataset discovery

Models, platforms and notable products mentioned

Cloud / general models:

• OpenAI (ChatGPT / GPT‑5.2 mention)
• Google Gemini (Gemini Pro / deep reasoning mode)
• Microsoft Copilot / GitHub Copilot
• Anthropic Claude
• Meta LLaMA
• Mistral
• X (Grok)

Specialty / deep‑reasoning:

• Deepsig R1 (deep reasoning model) — local download recommended vs using a questionable frontend
• Perplexity (aggregator + deep reasoning/search), Perplexity Pro

Local / OSS / Polish:

• LLaMA variants
• Belik (Polish model)
• GGUF format for CPU runs

Tools & services demoed or referenced:

• Gemini image editor (image-to-image & avatar edits)
• Perplexity, Hugging Face, platform.openai.com (sandbox)
• Manus AI (agent/builder)
• Fireflies (transcription/meeting capture)
• Leonardo / “Nano Banana Pro” (image generation model label)
• tataf.com (tool directory)

Prompt engineering — structure and techniques

Four core components of a good prompt:

Instruction (verb), context, data input, answer format/constraints

Seventeen prompting techniques covered:

1. Zero‑shot prompting
2. One‑shot prompting
3. Few‑shot prompting
4. Instruction‑based prompting
5. Role/persona prompting
6. Chain‑of‑thought prompting
7. Step‑by‑step prompting (explicit)
8. ReAct (reason + action)
9. Tree‑of‑thoughts
10. Self‑consistency (generate multiple solutions then pick)
11. Prompt chaining (series of smaller prompts)
12. Metaprompting (prompt to create/optimize prompts)
13. Contextual prompting (provide rich document/code)
14. Structured/output format constraints (JSON/XML/table)
15. Negative prompting (specify what to avoid)
16. Prompt injection awareness / jailbreaking mitigation
17. Multi‑agent / planner‑executor prompting

Practical tips:

• Prefer structured outputs (JSON) for integrations
• Use delimiters for context and to reduce injection risk
• Combine techniques (e.g., few‑shot + role + format constraints)
• Prefer English for multimodal/generative prompts (most models trained predominantly on English)
• Use negative prompts to avoid artifacts
• Use metaprompts to iterate and improve prompt quality

Product demos and practical examples

• Photo editing/generation: uploaded Tomek’s photo to Gemini and generated full‑body knight armor + lasers — highlighted speed vs manual Photoshop
• Valuation & sales advice: uploaded photo of an Amiga 500 set; model recognized components and produced pricing, market tactics, and sales copy
• Travel planning: complex Peru trip prompt using deep reasoning mode for updated itineraries and safety tips
• Sandbox usage: platform.openai.com for testing system/developer messages, comparing models, and optimizing token costs

Security, ethics, regulation, and risk management

• Hallucinations: what they are and mitigations (deep reasoning, verification, delimiters, human‑in‑the‑loop)
• Prompt injection and jailbreaking: demonstrated as real risks; importance of AI cybersecurity measures
• Data leaks: avoid personal/free accounts for corporate confidential data — use business licenses or local models for sensitive data
• Explainability & bias: risks such as hiring discrimination; importance of Explainable AI and transparency
• EU AI Act / Code of Good Practice: model cards, transparency requirements, copyright/licensing implications, model risk classification (forbidden/high/medium/low), and regulatory challenges
• Copyright & authorship: AI‑generated works require significant human input to qualify as protected authorship; check model/tool licenses

Practical recommendations & best practices

• Use business accounts/licenses for company work
• For sensitive corporate content prefer local models + RAG with vetted embeddings
• Verify AI outputs with independent sources before acting (medical/legal/critical decisions)
• Build human‑in‑the‑loop controls; avoid full autonomy for critical transactions
• Optimize token usage for cost control; test models in sandboxes before production
• Combine multiple models/platforms where appropriate (aggregators like Perplexity can be cost‑effective)
• Develop AI competencies: data literacy, low‑code/no‑code skills, prompt engineering, critical thinking, basic model ops

Course / Tutorial — future sessions (high level)

1. Intro to AI, models, prompting (this session)
2. Multimedia AI: image, audio, video, 3D, local creative tools (create anthem)
3. Business tools: project management, BI with AI
4. Microsoft Copilot (Chat, 365, Studio) + risks
5. AI for parents & children (education)
6. Cybersecurity, leaks, deepfakes, attacks
7. Building personal assistants, persona/prompt management
8. Designing AI-enabled applications, low/no‑code builders
9. Safe company implementation (local models, configuration)
10. AI agents & automation (multi‑agent setups)
11. Sales & social media automation with AI
12. Research & education use cases
13. Deep technical anatomy of models (algorithms, layers)
14. SEO / AEO / GEO positioning in the AI era (bonus / date tbd)

Additional resources: OSINT toolkit course, minibook of tools, Sekurak AI columns, possible workshops on model tuning and local deployments.

Mentions of reviews / comparisons / recommendations

• Perplexity Pro: recommended as a cost‑effective aggregator
• Gemini: rated subjectively strong for images and multimodal tasks
• Deepsig R1: flagged as a major LRM with good deep reasoning; security concerns about data routing
• Claude (Anthropic): noted for large context windows and suitability for programming tasks (Cloud Code)
• Grok (X): described as a “dark horse” with fewer safety constraints but ethical considerations
• Advice: evaluate models per task, test in sandbox, prefer local installs where confidentiality matters

Main speakers / sources

• Main presenter: Tomek (Tomasz) Turba — cybersecurity & AI trainer, Securitum / Sekurak.pl
• Organizations / platforms referenced: Securitum, Sekurak.pl, Hugging Face, OpenAI, Google, Anthropic, Deepsig (DeepSeek R1), Perplexity, Meta (LLaMA), Manus AI, Qdrant, tataf.com
• Contributors referenced: Łukasz Łopuszański (colleague, AI columns concept), guest contributors planned (e.g., Michał Sajdak for deepfake session)

Extras (available outputs)

• Short checklist for secure corporate AI adoption (extractable)
• Concise prompt engineering cheat‑sheet (templates for zero-/few‑shot, role prompts, JSON output, negative prompts)