AI Deep Dive

The Training Loop (every AI model follows this)

1. 1Forward pass: Feed input data through the network. It produces a prediction.
2. 2Loss calculation: Compare the prediction to the correct answer. Measure how wrong it was (the "loss").
3. 3Backpropagation: Calculate how each weight contributed to the error.
4. 4Weight update: Adjust weights slightly to reduce the error (using gradient descent).
5. 5Repeat: Do this billions of times across the entire dataset. Each full pass = one "epoch."

The Three Phases of Building an LLM

Phase 1 — Pre-training (costs $10M-$100M+): Feed the model trillions of tokens of text from the internet, books, code. The model learns language, facts, reasoning patterns. This produces a "base model" — it can complete text but won't follow instructions.

Phase 2 — Fine-tuning (SFT): Train on curated instruction/response pairs. "When a human asks X, a good response looks like Y." This teaches the model to be a helpful assistant instead of just an autocomplete engine.

Phase 3 — RLHF (Reinforcement Learning from Human Feedback): Human raters rank multiple model responses. A reward model learns what humans prefer. The LLM is then trained to maximize that reward. This is what makes Claude polite, safe, and genuinely useful vs. just technically capable.

• No true memory: Each conversation starts fresh unless you engineer persistence.

• Hallucinations: They confidently state false things. This is inherent to probabilistic generation. Mitigation = RAG, grounding, verification layers.

• No real-time data: Knowledge is frozen at training cutoff unless you add search/retrieval tools.

• Math and precise logic: Unreliable for complex calculations without tool use. They approximate; they don't compute.

• Determinism: Same input can produce different outputs. Temperature controls randomness but never eliminates it.

Anthropic's approach centers on "Constitutional AI" — the model is trained with a set of principles (a "constitution") rather than relying purely on human labeling. This produces more consistent, principled behavior.

Practical differences: Claude tends to follow complex instructions more precisely, handles very long documents well (200K context), and is more cautious about harmful outputs. For enterprise use: Claude is available on AWS Bedrock and GCP Vertex, so you can keep data within your cloud perimeter.

1. 1Observe: Receive user request or trigger event. Retrieve relevant context from memory.
2. 2Think: The LLM reasons about what to do next. It creates a plan or picks the next action.
3. 3Act: Call a tool — query a database, call an API, send a message, update a record.
4. 4Observe: Check the result of the action. Did it work? Was the data correct?
5. 5Loop or stop: If the goal is met, respond to the user. If not, go back to step 2 with updated context.

How RAG Works — Step by Step

1. 1Ingest documents: Take your internal docs (PDFs, Confluence pages, Slack messages, CRM data). Split them into chunks (typically 200-500 tokens each).
2. 2Create embeddings: Run each chunk through an embedding model (e.g., OpenAI text-embedding-3, Cohere Embed). This converts text to a numerical vector (a list of ~1500 numbers) that captures semantic meaning.
3. 3Store in vector database: Store those vectors in a vector DB (Pinecone, Weaviate, pgvector, Qdrant, Chroma). This enables fast similarity search.
4. 4At query time: Convert the user's question to a vector using the same embedding model.
5. 5Search: Find the top 5-20 most similar document chunks in your vector DB.
6. 6Augment: Insert those chunks into the LLM prompt as context: "Given this information: [chunks], answer the user's question: [question]."
7. 7Generate: The LLM answers based on the retrieved context, dramatically reducing hallucination.

Approach 1: Embed an LLM via API

The vendor calls Claude/GPT-4 via API and wraps it with their product context. Your CRM data is injected into prompts. The vendor manages prompt engineering, guardrails, and tool definitions. Example: Salesforce Einstein GPT uses a combination of proprietary models and partner LLMs.

Approach 2: Fine-tune on Domain Data

Take a base model, fine-tune it on industry-specific data (healthcare records, legal contracts, financial reports). This creates a specialized model that understands domain jargon and patterns. Example: Bloomberg trained BloombergGPT on financial data.

Approach 3: Build a RAG Layer Over Customer Data

The vendor builds a RAG pipeline that indexes each customer's data (behind proper isolation/permissions). When the AI answers, it's grounded in that customer's specific data. Example: Salesforce Data Cloud feeds into Einstein AI as a retrieval layer.

Approach 4: Train Custom Models

Major vendors train their own models from scratch for specific tasks — fraud detection, recommendation engines, demand forecasting. These are typically smaller, specialized models, not general-purpose LLMs.

Say you want to build: "An AI that answers customer questions using your knowledge base."

ASR (Automatic Speech Recognition) / STT (Speech-to-Text): Converts spoken audio to text. Leading models: OpenAI Whisper (open source, very good), Google Cloud Speech, Deepgram (low latency, real-time), AssemblyAI.

LLM Processing: Same as any text-based AI. The transcribed text is the input. The LLM generates a text response.

TTS (Text-to-Speech): Converts the LLM's text response back to natural-sounding speech. Leading options: ElevenLabs (most natural), OpenAI TTS, Google Cloud TTS, Play.ht, Cartesia (ultra-low latency).

Zero Data Retention (ZDR): Enterprise API tiers from Anthropic and OpenAI guarantee your data is not used for training and is not retained after the request. Verify this in your contract.

Data residency: Run Claude via AWS Bedrock in your preferred AWS region. Data never leaves your VPC. Same with Google Vertex AI.

PII handling: Either redact PII before sending to the LLM, or use enterprise tiers with appropriate data processing agreements.

SOC2 / HIPAA / GDPR: Check provider compliance certifications. Anthropic and OpenAI have SOC2 Type II. HIPAA requires BAA agreements.

LLM API Pricing Model

You pay per token — both input (prompt) and output (response). A 2000-word document as input + a 500-word response ≈ 3500 tokens ≈ $0.01-0.05 depending on model.

Cost Optimization Strategies

Model routing: Use a cheap/fast model (Claude Haiku, GPT-4o mini) for simple queries. Route complex queries to the expensive model (Claude Opus, GPT-4). This alone can cut costs 60-80%.

Caching: Cache frequent queries. Anthropic offers prompt caching — if the same system prompt is reused, you pay a fraction of the cost.

Shorter prompts: Every token in your system prompt is charged on every request. Optimize for brevity.

Batch processing: For non-real-time tasks, use batch APIs at 50% discount.

Agentforce = a platform for building and deploying AI agents inside Salesforce. Think of it as:

LLM (multi-model — they route to different providers) + Salesforce data (CRM, Data Cloud, Knowledge Base) + Tools (Salesforce actions: create case, update opportunity, send email) + Guardrails (Trust Layer + your business rules) + Deployment (embed in Service Cloud, Sales Cloud, web, Slack, etc.)

You can build agents using low-code tools (Agent Builder) or code (Apex, LWC). Salesforce provides pre-built agent templates for: Service Agent, Sales Coach, Merchant Agent, Buyer Agent, Campaign Agent.

• "How does Agentforce handle multi-step tool failures and retries?"

• "What's the latency overhead of the Trust Layer on each LLM call?"

• "Can I bring my own model (BYOM) and still use the orchestration layer?"

• "How does Data Cloud grounding work — is it RAG under the hood? What embedding model?"

• "What's the pricing model — per agent, per conversation, per action?"

• "How do I evaluate agent quality? Is there built-in testing/evaluation tooling?"

• "What observability do I get — can I see every reasoning step, tool call, and retrieval?"

The Pragmatic Adoption Ladder

1. 1Internal productivity (low risk, high value): Deploy an LLM chatbot connected to your internal knowledge base. Let employees ask questions about policies, docs, procedures. This is the quickest win with the least risk.
2. 2Customer-facing copilot (medium risk): AI that helps customers with common questions, grounded in your help center. Always with a "talk to human" escape hatch.
3. 3Process automation agents (higher risk): Agents that take actions — update records, send emails, process refunds. Requires guardrails, approval workflows, thorough testing.
4. 4Autonomous agents (highest complexity): Multi-step agents that handle end-to-end workflows with minimal human oversight. Only after you have robust evaluation, monitoring, and rollback capabilities.