How LLMs actually work β the engineering version, in plain language.
35 MIN
The RTFC framework. Five principles. Three live labs.
70 MIN
Applied β the lab the whole course is built around.
20 MIN
The vocabulary is sloppy β just know the shape.
Your brain: ~86 billion neurons, connected by synapses. Learning = strengthening or weakening connections through experience.
Nobody programs the rules. You don't write "if email contains URGENT then mark spam" β you show 10,000 labelled emails and the network figures out the rules itself.
if email contains "URGENT"
and sender not in contacts
then mark as spam
Can't handle anything the programmer didn't anticipate.
1. Show 10,000 labelled emails
2. Model finds its own patterns
3. Handles examples it's never seen
Same idea, scaled to internet-sized data, makes an LLM.
The skill is no longer how to build models.
It's how to use them well.
β¦at a scale and sophistication where it starts doing things that look like reasoning.
Wikipedia, books, GitHub, papers β essentially the public internet, compressed as numbers.
Multi-step reasoning, translation, code β none explicitly programmed in. Above a certain size, they just appear.
Frontier LLMs are estimated to have hundreds of billions to trillions of parameters β exact counts are not publicly disclosed. More parameters = more capacity to learn subtle patterns. Scale changes what's possible.
"Understanding AI is a superpower in 2026"
~3β4 characters per token Β· "ChatGPT" = 3 tokens Β· 1 page English β 500 tokens
"How many r's in strawberry" β the model sees tokens, not letters.
"1M-token context" β 750,000 words β 6β8 novels.
Maximum tokens per conversation. Coming up in two slides.
probability of next token at each step
Ask the same question twice β slightly different answers. By design. The creativity dial (technically: temperature) controls how often the model picks the top choice vs a less likely one.
Not deliberation. Pattern-matching at a scale where the patterns look like reasoning. Brilliant and confidently wrong β sometimes in the same breath.
The architecture built on attention is called a transformer. It now powers basically everything: text, image generators (Midjourney), video (Sora), audio, and AlphaFold for protein structures. Same engine, different fuel.
Snapshot Β· May 2026
Default in ChatGPT. Strong agentic coding. GPT-5.5 Pro for harder reasoning. ChatGPT crossed 900M weekly users in Feb 2026.
Frontier. Strong on long docs + following complex instructions. Sonnet 4.6 is the workhorse. Apps: Claude Code, Cowork, Excel, PowerPoint, Chrome.
Native multimodal: text, image, audio, video. Deeply integrated with Google Workspace (Docs, Gmail, Meet) and Chrome.
Free to download and run. Scout has the largest context window of any open model. Powers many third-party apps via providers like Groq.
Rule of thumb: the best model is the one your team actually uses. Switching costs are essentially zero.
The model generates something that sounds plausible but is factually wrong β and presents it with full confidence. Made-up citations. Wrong dates. Invented quotes. Fictitious features.
The model's job is plausible next tokens, not true statements.
There is no internal fact-checker.
Confidence in tone tells you
nothing about accuracy of content.
Hallucinations are getting less frequent with reasoning models β but they have not been eliminated. The most plausible-sounding ones are the most dangerous.
May 2026
max tokens per conversation
Unlocks: paste entire books, multi-doc repos, full meeting transcripts β the model reads them all.
Still remember: each new chat is fresh. The model does not remember last week unless you paste it in.
GPT-5.5
Claude Opus 4.7
Gemini 3.1 Pro
Most chat tools now search the web in real time. Use that for recent news, prices, regulations, breaking events. For everything else, the cutoff doesn't matter.
Think of an LLM as a brilliant colleague who just got back from a 6β12 month sabbatical. Sharp, well-read, great judgement β but hasn't seen the news since they left.
10-minute break. When you come back: the part that changes your work.
Open the companion β "one thing I want to remember"
BACK AT THE TOP OF THE HOUR Β· QUESTIONS AFTER THE BREAK
Pick something you already do, 10β30 minutes, with a draft-then-polish shape.
The boring task is where you have a baseline.
Some of you brought a real task from your work, as I asked in the pre-work. Good. We'll come back to it in Lab 3 in about 50 minutes.
The most banal task is the most valuable one. Boring is the point.
ChatGPT Β· Claude.ai Β· Gemini. Open browser, type, get a response. No setup. Start here.
Perplexity. ChatGPT and Claude also have built-in web search. Best for research and current events.
GitHub Copilot Β· Microsoft Copilot Β· Notion AI Β· Claude for Excel Β· PowerPoint Β· Chrome. The AI comes to where you already work.
Claude Code Β· Cowork Β· Cursor Β· Devin.
AI that takes action, not just talks.
The acronym: MCP.
Practical advice: don't use all of these. Pick one chat tool and go deep. Switching costs are zero.
The model doesn't remember last week. If you want context, you paste it.
"Make it shorter." "More formal." "Try again, in Romanian." Each turn builds on the last.
PDFs, images, spreadsheets β upload and the model reads them as context. We'll use that in Lab 3.
Imagine you're briefing a brilliant freelancer who has never met you, doesn't know your company, doesn't know what "done" looks like, and only sees the brief. The quality of their work is the quality of your brief.
NEXT 40 MIN Β· FIVE PRINCIPLES FOR WRITING BETTER BRIEFS
Vague in, vague out.
Help me with my email.
β Generic response. Doesn't know who, what, why, or how long.
Rewrite this email to sound more direct. Remove apologetic language. Max five sentences.
β Four things to optimise for: task, audience, format, constraints.
Those four words are actually the framework. Let me draw it out.
Who do you want the model to be? "Act as a senior PM." "Act as a corporate lawyer." Primes a tone and a body of knowledge.
What specifically should it do? Not "help with my deck" β "Write a one-page brief: problem, solution, metrics, risks."
Shape of the output. Bullets? Table? Email with subject line? 200-word exec summary? Say it.
Boundaries. Word count, tone, audience, things to avoid. "Under 400 words. No jargon. Don't use 'synergy.'"
[Role] Act as a senior product manager at a B2B SaaS startup. [Task] Write a one-page brief outlining the problem, solution, key metrics, and risks for a new onboarding flow. [Format] Markdown with four numbered sections. [Constraints] Under 400 words. Audience is our non-technical CEO. No jargon.
Write something about marketing.
Show, don't just tell. (few-shot prompting)
Write a meeting title for our Q3 planning session.
β "Q3 Planning Session."
Generic. Doesn't sound like your team.
Write a meeting title for our Q3 planning session.
Examples of how we name meetings:
Β· "Shipping or Sinking: H1 Retrospective"
Β· "The Money Slide: Investor Prep"
β Now the model knows your voice, not its default.
Especially for tone, format, voice, style β show is more efficient than tell.
Bullets Β· table Β· numbered list Β· JSON Β·
comparison grid Β· three-sentence exec summary Β·
email with subject line.
The model can produce almost any format β but only if you ask.
Two seconds in the prompt saves five minutes of reformatting later.
"Make it shorter." Β· "More formal." Β· "Add an example." Β· "Give me three alternatives."
"What's missing?" β criminally underused. The model surfaces gaps you didn't know existed.
RTFC prompt from Lab 1 β or start fresh with a new task.
"what's missing?" β that's where the gold is.
Biggest immediate wins for most of you.
World's most patient tutor.
Removes the blank-page problem.
"I wish a computer would do this" β it does.
The pattern: anything with a draft-then-polish shape.
Anything tedious. Anything you've been avoiding.
It generates based on training, doesn't look up β unless web search is on. Even then, verify.
Don't take medical, legal, or financial decisions on AI advice without a human professional reviewing.
Simple arithmetic, counting letters, spatial reasoning. "How many r's in strawberry" β still weak.
Trained on human text β inherits human biases. Underrepresents some cultures and viewpoints.
First draft, not final answer.
Your judgement makes the output safe to act on.
In Lab 3 β anonymise. Practise the discipline from day one.
Your value used to be "can you write a good email?" β AI can.
Your value is now "can you judge whether this email is right?"
The shift from chat to agents is happening now.
An open standard for connecting LLMs to your tools.
Released by Anthropic late 2024. Adopted across the industry.
β Pull last week's sales from the CRM and draft the weekly summary.
β Find every email about Project Apollo and summarise the thread.
β Schedule a 30-min slot with everyone in this Slack channel next week.
You'll hear this acronym a lot in 2026. Now you know what it is.
RTFC β Role, Task, Format, Constraints.
what's missing?
"My task was X. The biggest surprise was Y."
Three small uses a day. That's it.
Expand from there.
Stop asking "can AI do this?"
Start asking "what would I need to tell a brilliant assistant to help me with this?"
Then type that. That's the whole game.
Reps first, theory second. Use what we covered today for a month before you take an advanced course.
"In the next 7 days,
I will use AI to __________________________."
β DOESN'T WORK
"I'll use AI more."
β WORKS
"I'll use Claude to draft my weekly Monday update."
"I'll use ChatGPT to prep for my Wednesday 1:1."
The companion saves it locally. I'll send a check-in email in 7 days.
The best AI prompt you'll ever write
is the next one.
WHAT WE DID TODAY
In your inbox tomorrow: recording Β· slides Β· workbook Β· companion link Β· T+7 check-in
AI-COURSES.BADITA.ORG Β· THANK YOU