AIceberg · Deep dive through several layers
0 meters · depth
Surface
Descent 00%
Welcome to your descent

AIceberg

Deep dive through several layers.

Scroll to begin your descent
How this works

A learning journey, not a course.

You'll explore the fundamentals of AI at the surface and progressively dive deeper into more advanced concepts as you descend through the iceberg. Each layer has clickable tiles — open one to learn the topic any way you like.

Watch
a short focused video
Listen
to the conversational deep-dive
📄
Read
the Pax8-branded whitepaper
Try it
against real partner work

Or upload the whitepaper into ChatGPT, Claude, or Microsoft NotebookLM and ask it to summarise, quiz you, or generate a podcast — learning is flexible, and personalised to you.

You only need to complete one format per topic — then mark the section complete and continue the descent.

00 — The Surface · 0 m

What everyone sees.

The foundation that works whether this is your first week or your fifth year — the AI strategy, tools, data discipline and judgement habits that sit at the centre of how we work and how we help partners become MIPs.

Why AI matters here Pax8 EMEA · Strategy

Pax8 helps partners evolve from Managed Service Providers into Managed Intelligence Providers — at its heart an AI story. This is why we work on Copilot every day and why MIP language shows up in every QBR.

The 6 plays at a glance Pax8 EMEA · Strategy

The official EMEA framework for the partner journey to MIP. Discover · Sell · Buy · Build · Implement · Manage. You don't need to master it — just know the language well enough to follow a conversation and sound credible.

M365 Copilot (Pax8 tenant) Office Skills with Amy

Tier 1 enterprise tool — approved for D0–D3 data including partner-confidential content. QBR drafts (cycle drops from 4–8h to 1–2h), partner business plan synthesis, marketplace listing guidance. The single tool worth touching every day.

Copilot Cowork M365 creator

Tier 1 in-tenant tool for the heavier work — multi-step research, building docs and decks across M365 files, pulling a partner picture across Outlook, Teams, OneDrive and SharePoint. The deep-dive partner.

Which tool for which data Pax8 · InfoSec

Before anything goes into AI, check the official Tool & Data Classification document. D0–D4 tiers, the never-list (no partner personal data, no customer data, nothing pre-GA) regardless of how good the tool is.

The data mandate Pax8 EMEA · Strategy & Ops

Every AI summary, insight and recommendation is built on data we record. Recording partner conversations where permitted, capturing outcomes, and keeping Salesforce current are the fuel that makes AI useful. AI helps you complete the record faster; it doesn't excuse leaving it blank.

AI in the field Pax8 EMEA · Enablement

Events, posture, follow-up. Before (research who's attending, draft intros), at (capture quickly with a voice note), after (clean recap and next step within 24h). AI handles the admin so you can be more human in the room.

Where AI ends and your judgement begins Advait Sarkar · TED

AI is a draft, not the truth — verify names, numbers and facts before anything leaves the building. Six scenarios where AI is the wrong answer. The one question that catches them all: "if this were subtly wrong and I didn't notice, what would happen?"

01 — Shallow · 120 m

Familiar waters.

The vocabulary anyone who's read a few articles already knows. Important ground rules — they keep you from saying embarrassing things at dinner.

AI for content creation Wes McDowell

How to use AI to draft, edit and polish written content — and what to keep human. The blank page problem, solved.

ChatGPT for emails & research Jeff Su

Turn ChatGPT into the second brain that handles the email, summaries and research you keep putting off. Two five-minute habits that pay back forever.

AI social media schedulers Buffer (official)

Buffer, Hootsuite, Publer and the AI features that turn “I should post more” into “I have a quarter of content already queued”.

Prompt templates Prompt Engineering

The starter pack of templates that works for the work you actually do. Saved prompts, ready to paste.

Resume building with AI Resume Genius

AI rewrites are genuinely transformative for CVs and LinkedIn — and the few traps to avoid. The use case where AI shines hardest.

AI meme generators ChatGPT Tutorials

02 — Twilight Zone · 400 m

How it actually works.

The engineering layer. Most engineers shipping "AI features" only have a hand-wavy grip on this. It's where the magic starts to look like math.

Prompt Engineering basics Tina Huang

Beyond templates — the patterns the best prompters use, and how to apply them to the actual work you do at Pax8.

ElevenLabs, D-ID, Pictory Skill Leap AI

Realistic AI voiceovers, talking-head avatars and auto-generated explainer videos — and where each one actually fits.

No-code AI workflow builders Nate Herk

n8n, Flowise, Make and Zapier — what each one is best for and how to ship your first automation this week.

AI automations (Zapier, Make) Kevin Stratvert

The boring-but-massive lever most people skip past. Zapier and Make with AI nodes — every repeated task becomes background work.

AI avatars & video creation Curious Refuge

HeyGen and Synthesia for scaled video — what they

Speech-to-text with Whisper sentdex

OpenAI API / Anthropic API Patrick Loeber

The plumbing under every AI app: requests, responses, tokens, models, costs. Even if you don\

Fine-tuning LLMs IBM Technology

When fine-tuning makes sense, when it doesn\

Vector databases Fireship

Where AI stores what it knows about you. Embeddings, similarity search and the data layer powering every smart search and chatbot.

LangChain & LangGraph IBM Technology

The orchestration layer for LLM apps. Frameworks that let you compose LLM calls, tools and decisions into real applications.

Embeddings IBM Technology

Tokenization Matt Pocock

03 — Midnight Zone · 1,500 m

The research frontier.

Below recreational diving depth. The vocabulary of arXiv papers and lab slack channels. Most of this was unknown five years ago and is unstable even now.

Autonomous AI agents IBM Technology (Maya Murad)

Tool use via ReAct Sam Witteveen

Reason + Act — the single pattern behind almost every agent that works: thinking out loud, then taking one step.

Multi-agent collaboration IBM Technology

Agent Workflow orchestration Sequoia Capital

Memory management Underfitted

Short-term context, long-term memory, retrieval — the three layers behind every agent that "remembers you".

Function calling Independent educator

The protocol that lets language models drive APIs, query databases and operate other software — the bridge from talking to doing.

04 — Abyssal · 3,500 m

Alignment, interpretability, safety.

The field stops being engineering and becomes something stranger: an attempt to inspect, predict, and corral systems we made but did not design. Cold, dark, very serious people.

LLM Monitoring & Evaluation IBM Technology

The discipline that separates production-grade AI apps from impressive demos. If you can\

Self-healing AI pipelines LangChain

Retries, graders, fallbacks and reflection — the patterns that turn fragile pipelines into reliable ones.

Custom embeddings & reranking James Briggs

Retrieval-Augmented Generation IBM Technology (Marina Danilevsky)

Probably the most important AI app pattern of the decade. The pattern behind almost every "AI that knows your stuff" product.

Synthetic data generation IBM Technology

Using LLMs to manufacture training data, test data and edge cases — and the trap to avoid.

Model quantization Olivio Sarikas

The technique that lets a 70B-parameter model run on your laptop — what it is, how it works, what you lose.

05 — Hadal · 6,500 m

The esoteric end.

Memes, metaphors, and decision-theoretic spelunking from rationalist forums and alignment Twitter. Half-folklore, half-philosophy, occasionally prescient. Approach with curiosity, not credulity.

Chain-of-thought reasoning Serrano.Academy

The 2022 paper that quietly transformed how we prompt LLMs — and what it tells us about how they actually reason.

Emergent behaviors in LLMs AI Coffee Break with Letitia

Why bigger models suddenly do things smaller ones can\

Instruct tuning vs RLHF Ari Seff

The two-step process that turns a text-completion engine into ChatGPT, Claude or Copilot.

AI Alignment & safety Robert Miles AI Safety

Why making AI helpful is harder than making it powerful, and the field that has emerged to make that tractable.

Reward modeling Robert Miles AI Safety

The clever trick that turns subjective human preferences into a number a model can optimise.

Causal inference in ML CodeEmporium

Why machine learning is great at prediction and bad at "what if" — and what\

06 — The Void · ∞

What we do not know — yet — outweighs everything above.

AGI design principles Computerphile

The honest landscape of artificial general intelligence — what we mean, what we don't, where serious researchers actually disagree.

Self-improving AI Computerphile

Recursive self-improvement, AI improving AI, and the intelligence-explosion thesis — without the doom-mongering.

Autonomous economic agents IBM Technology

When AI agents transact, hold accounts, hire other agents — the economic-actor frontier.

Neural-symbolic systems IBM Research

Combining neural networks (great at pattern) with symbolic reasoning (great at logic) — the long-running quest to get both.

Artificial consciousness TED

Could AI be conscious? What would we look for? Where serious thinkers actually disagree.

AI × Neuroscience fusion TED

What AI teaches us about brains, and what brains keep teaching us about AI. A genuinely virtuous circle.

And below even these — the questions nobody knows how to answer.

Is there something it's like to be GPT?

We have no agreed-upon test for consciousness, and no reason to assume it can't arise in silicon. We also have no reason to assume it can.

Will we recognise AGI when we build it?

The goalposts keep moving. Chess was AGI, then it wasn't. Conversation was AGI, then it wasn't. What's left?

How quickly does this go?

Estimates from serious researchers span "decades" to "next presidential term." Both groups have looked at the same data.

What is the model doing, really?

We can read every weight and still not know what a frontier model is thinking. Interpretability is in its infancy.

"We are building minds
we cannot yet read."
— the entire field, more or less
↓ Keep descending