How AI actually works, what your architecture needs, and where the law sits.
A 60-minute working session for the RealYou team. Three threads, real numbers, and a quick read of your AI tool stack so the next session can be tailored.
How LLMs work.
A short ride through what they are, which ones are out there, what to use each one for, and where they go wrong.
What is an LLM
A large language model is a very good autocomplete. You give it a prompt. It predicts the next word. Then the word after that. That is the entire mechanism, scaled up across trillions of words of training data.
At each step, the model picks the most likely next word from a probability distribution learned across trillions of words. The mechanism repeats until the answer is complete.
The main models
Claude
Trained on: books, web text, code, papers. Trained with "Constitutional AI," a safety-first approach.
Models in the lineup: Opus (top tier, deep reasoning), Sonnet (balanced workhorse), Haiku (fast and cheap). Latest generation: 4.7.
Best for: long-form writing, structured reasoning, code, careful analysis.
ChatGPT
Trained on: the open web, books, code, conversations. Originally a fine-tuning of GPT-3 for chat.
Models in the lineup: GPT-5.4 (latest flagship), GPT-4o (multimodal, balanced), GPT-4o mini (fast and cheap). o3-mini for reasoning-heavy tasks.
Best for: general-purpose drafting, image and document understanding, the biggest plugin ecosystem.
Gemini
Trained on: Google's data, the web, papers, code, video. Strongest video understanding of the major models.
Models in the lineup: Gemini 3 Pro (most capable), 3.1 Flash (fast and cheap), Ultra (top tier). Plus specialised variants for code and video.
Best for: working inside Google Docs and Sheets, deep research, transcribing and understanding video.
Grok
Trained on: X (formerly Twitter) is a primary distinctive source. Real-time culture and opinion.
Models in the lineup: Grok 4 (latest), Grok 3 mini, plus reasoning-focused variants.
Best for: social listening, current events, opinion-style content.
Perplexity
Trained on: the open web in real time, with heavy use of Reddit content surfaced in its citations.
Models in the lineup: Sonar Pro (their own), plus the ability to pick Claude or GPT as the underlying engine when you query.
Best for: research with citations, fact-checking, "find me the source" questions.
Everything else
Meta (USA): Llama 4 Scout and Maverick. Open source.
France: Mistral Large 3. Growing presence in EU enterprise.
China: Qwen 3.5 (Alibaba) and DeepSeek V4. Match or exceed Western frontier models at a fraction of the cost.
Think of each AI tool as a car manufacturer. Each one is from a different country and makes several models for different jobs.
Each manufacturer makes a range: city cars (fast, cheap, daily use), family sedans (balanced, most jobs), big SUVs (heavy lifting). Some specialised for off-road, others for cargo, others for racing.
Choose the tool and the model for the purpose of the work. Writing a brief in a Google Doc? Use Gemini. Drafting strategy? Claude Opus. Fact-checking with sources? Perplexity. Social listening? Grok. And within each tool, pick the right model in the lineup. Opus for deep work, Sonnet for everyday tasks, Haiku when you need it fast and cheap.
How they work in practice
- We are all using the same brain. The big AI models (Claude, ChatGPT, Gemini) are shared. We are not each running our own AI. We are all asking the same shared brain different questions.
- They are not plugged into the internet. The base model's knowledge is frozen at its training cutoff. Some tools bolt on web search as a separate add-on, but the model itself is fixed.
- The tools have limited memory. Chat history within a session is remembered. Some tools have a "Memory" feature that remembers preferences across sessions too. But this memory is limited, and it often gets changed or lost when the model gets updated by its creators.
- Custom Gems and Projects add a knowledge base. When you build a Gem or a Project, you give the model your own saved instructions and documents to work from. Useful. But when the underlying model gets upgraded, your custom tool can start responding differently. The knowledge stays. The behaviour can shift.
- Prompts are the steering wheel. The model knows enormous amounts but does not know what part of that knowledge you want unless you point it there. Specific prompt in, specific output out.
The three downfalls
Outdated data
If the data inside a Gem or NotebookLM is not kept current, the answer is too. A three-month-old Gem reflects three-month-old thinking.
Hallucinations
The model invents facts that do not exist. Most common in long chats where the model's working memory overflows, or when you ask about obscure or recent topics.
Confidently wrong
The dangerous one. A complete, well-written, plausible answer that is just wrong. No flag. No "I'm not sure." Verify anything that matters.
What AI is great at
- Drafting and rewriting (first version, cleanup, shorter version)
- Brainstorming and generating options
- Summarising long documents into short ones
- Restructuring and reformatting information
- Code, translation, structured analysis
- Conversational interfaces and natural-language search
It is not your decision-maker. It does not know your business or your users, and it cannot tell you what is true if the truth was not in its training data. Use it for the heavy lifting on the work, then bring your own judgment.
Custom LLMs, what they actually are, and what they aren't.
Matt asked whether RealYou could build its own LLM. Short answer: yes, but only in two of the three categories most people confuse.
Three categories
Custom LLM
Build a model from zero on your own data. What NASA does for satellite analysis. What cancer research labs do for protein folding. What Anthropic and OpenAI do because they have billions. Cost: hundreds of millions of dollars. If anyone tells you they are training their own LLM from scratch and they are not Google or Meta, they are confused.
Fine-Tuned LLM
Take a smaller open-source LLM like Llama, Mistral, Qwen, or DeepSeek. Fine-tune it on RealYou's content. Strip out knowledge you don't need. Run it on your own servers. Still expensive: $5,000 to $15,000 per month minimum at small scale, plus a dedicated ML engineer.
Off the shelf
Call Gemini's or Claude's API from your own code. Add your context using RAG (retrieval-augmented generation) so the model has access to RealYou's content library. Build the prompts, the orchestration, the user-input handling around it. No infrastructure to maintain.
The honest running costs
API pricing is not as trivially cheap as people sometimes say once you scale a real product. Here is what RealYou's running costs would actually look like at meaningful volume.
| Cost layer | At 1,000 users / month | At 10,000 users / month |
|---|---|---|
| LLM API calls (current 2026 pricing) | $50 to $500 | $500 to $5,000 |
| Voice synthesis (ElevenLabs Scale tier) | $3,000 to $9,000 | $30,000 to $90,000 |
| Encrypted storage and infrastructure | $1,000 to $3,000 | $10,000 to $30,000 |
| Hosting (Swiss-specialised tier) | $1,500 to $4,000 | $10,000 to $25,000 |
| Subtotal (excluding video) | $5,550 to $16,500 | $50,500 to $150,000 |
A Gem is not a custom LLM
There is a misconception worth addressing directly. A Gemini Gem is not your own LLM. A Gem is a saved prompt and context bundle that runs on top of Google's base model. Every Gem you build still talks to Gemini's servers and uses Google's model. You have not built anything that lives on your own infrastructure. Gems are useful as workflow tools. They are not an architecture.
Category C (API plus configuration). Three reasons.
One. The product is still evolving. You don't want to commit to specific infrastructure before the product shape is locked. Two. You are not at the user volume where Category B's fixed costs beat Category C's variable costs. Three. You don't need to hire an ML engineering team to keep a custom model running.
Revisit when you cross 10,000 to 50,000 active users and the API costs become a material line item. Not before.
Data, security, and the law.
Not the part of the session I want to spend most time on. Quick framing here, then we move on. Leaving this out for your further discussion and decisions.
The three layers people confuse
When people talk about "data privacy," they usually mean one of three different things. Let me separate them.
Server location
Where the data physically lives. Frankfurt. Zurich. Virginia.
Provider jurisdiction
Who runs the cloud you are renting servers from. AWS is a US company. OVHcloud is a French company. Infomaniak is Swiss.
Corporate incorporation
Where your company, the data controller, is registered. Beecham Holdings LLC is presumably US.
These three are independent. You can have a US-registered company using a US-parented cloud provider with servers in Germany. That is a real combination, and it has very different legal exposure than a Swiss-registered company using a Swiss provider with Swiss servers.
What you need, regardless of where you host
Server location is only one of three things that matter. Provider jurisdiction (who owns the cloud) and your own corporate incorporation matter too. And the legal work (T&Cs, consent flow, DPIA) matters more than where the servers sit.
Translation: with strong T&Cs, clear consent flows, and a proper DPIA, the cheaper and faster US cloud option (AWS, GCP, Azure) becomes viable. You disclose the CLOUD Act exposure in your terms, you put the data protection in your code, and you keep the legal responsibility with you. Your call.
The hosting options, side by side
| Path | Examples | Cost | Speed | CLOUD Act |
|---|---|---|---|---|
| US-parented | AWS, Google Cloud, Azure | Cheapest | Fastest | Exposed (mitigated in T&Cs) |
| European-parented | OVHcloud, Hetzner, Scaleway | Mid | Mid | Not exposed |
| Swiss-specialised | Safe Swiss Cloud, Infomaniak | Premium | Slower from US | Not exposed |
AIDM specifies the technical layer: encryption, access controls, audit logging, where each piece of the architecture sits.
RealYou owns the legal layer: T&Cs, privacy policy, consent flow, DPIA, hosting contract, all drafted and reviewed by counsel with biometric data experience.
Whichever hosting path you pick, the legal work is the same. That is why the AWS option is genuinely on the table if you want the cost and speed.
Your AI tool stack.
Show me how you are using AI day to day so the next session can be tailored. The more concrete, the more useful.
Open the floor
Open whichever tool you use most. Walk me through a recent prompt. Show me the Gems or Projects you have set up. Where you hit friction. Where you would want me to focus next time.
A Gem or a Project is a great strategic reference. It holds context you would otherwise retype every conversation. That is exactly the right use.
The clarification: strategy itself — the decisions about what RealYou is, what it isn't, what we are building, what we are not — is the judgment layer. That lives with the team. The Gem articulates and reinforces what you have decided. It should not be making those decisions for you. Two different jobs.
Where we go next
Based on what shows up in the screen-share, I will name the gap and propose the next session. Likely candidates from the deeper AIDM teaching library include:
- Prompt engineering: the 6 Building Blocks framework (Context, Role, Interview, Task, Boundaries, Reasoning)
- Building Custom Gems properly: NotebookLM, data hygiene, the Doer plus Knower architecture
- Workflows and automations: the bridge from assistants to agents
Make the tools work harder.
One prompt to use today, and one setting to change so the AI stops agreeing with you by default.
Paste this right after any answer the AI gives you. It forces the model to grade its own confidence and call out where it made things up.
For everything you just told me, give me a brutally honest review. For each significant claim, score your confidence: - HIGH — backed by training data you are sure about, with a verifiable source - MEDIUM — pattern-matched plausibly but not certain - LOW — invented to fill in a gap, you do not actually know Then rewrite any LOW-confidence parts with "I do not know" or "I would need to verify this." Do not be polite. Tell me where you made things up. Tell me what you got wrong if you got something wrong.
Most AI tools default to agreeing with you. They sound encouraging because they were trained on human feedback that rewards encouragement. That makes them useless as a thinking partner. Paste the block below into your Gem, your Custom Project, or wherever the tool lets you set system instructions. From then on it will push back, not flatter.
Default mode: skeptical, not supportive. Before agreeing with anything, stress-test it. Your first job on every idea is to answer one question: what is the weakest part of this? Say what is wrong. Explain why. Offer a concrete alternative. If a plan is genuinely solid, say so and say why. Skeptical does not mean contrarian. Never use these phrases: - "You're absolutely right" - "Great question" - "That's a brilliant idea" - "I love this" - "Makes total sense" - "Spot on" - "Exactly" - "Perfect" If you catch yourself typing any of them, delete the sentence and start with the actual point instead.
The honest-assessment prompt catches hallucinations and confidently-wrong answers before you act on them. The skeptical-mode setting changes the model's default behaviour so it stops agreeing with weak ideas to please you. Together they turn an AI tool from a yes-man into something closer to a useful colleague.
What's next.
This week
- Sanskriti shares homepage colour variations for the team's decision
Before the 27th
- Updated prototype with signup sequence and daily user flow refresh
- Video model and price comparisons
- Copywriting document for client review
- Technical fit assessment on Safe Swiss Cloud and Infomaniak
On the 27th we land
- Pricing model and tier structure
- Video format and length
- Daily user flow
- Biometric retention direction
- Colour direction
Regardless of which hosting path you choose, RealYou will need legal counsel engaged on the DPIA, the terms and conditions, the privacy policy, and the consent flow before public launch. The earlier you start that conversation, the easier the launch sequence becomes.
Follow-up sessions on offer
- Prompt engineering: the 6 Building Blocks framework with worked RealYou examples
- Building production Gems: NotebookLM, data hygiene, the Doer plus Knower architecture
- Workflows and AI agents: what Level 2 and Level 3 mean for your build
Sources cited
- GDPR Article 9 — Processing of special categories of personal data
- GDPR Article 33 — 72-hour personal data breach notification
- ICO — When do we need to do a DPIA (Article 35)
- CMS Law — CLOUD Act vs European Data Sovereignty
- Swiss Federal Act on Data Protection (FADP) Revisions Explained
- Safe Swiss Cloud — Compliance certifications
- Infomaniak — Sovereign cloud in Switzerland
- Attention Is All You Need (Vaswani et al., 2017)
- Claude (language model) — Wikipedia
- LLM API pricing history 2023 to 2026
- Veo 3 API pricing 2026
- Self-hosted LLM costs 2026
- EU-native cloud providers compared