AI Engineering Architecture and User Feedback (Chapter 10 - AI Engineering - By Chip Huyen)

Download Book

<<< Previous Chapter Next Book >>>

Building Real AI Products: Architecture, Guardrails, and the Power of User Feedback

Why successful AI systems are more about engineering and feedback loops than models

---

Introduction: From “Cool Demo” to Real Product

Training or calling a large language model is the easy part of modern AI.

The hard part begins when you try to turn that model into a real product:

• One that real users rely on

• One that must be safe, fast, affordable, and reliable

• One that improves over time instead of silently degrading

This chapter is about that hard part.

So far, most AI discussions focus on prompts, RAG, finetuning, or agents in isolation. But real-world systems are not isolated techniques—they are architectures. And architectures are shaped not just by technical constraints, but by users.

This blog post walks through a progressive AI engineering architecture, starting from the simplest possible setup and gradually adding:

• Context

• Guardrails

• Routers and gateways

• Caching

• Agentic logic

• Monitoring and observability

• Orchestration

• User feedback loops

At the end, you’ll see why user feedback is not just UX—it’s your most valuable dataset.

---

1. The Simplest Possible AI Architecture (And Why It Breaks)

The One-Box Architecture

Every AI product starts the same way:

1. User sends a query

2. Query goes to a model

3. Model generates a response

4. Response is returned to the user

That’s it.

This architecture is deceptively powerful. For prototypes, demos, and internal tools, it works surprisingly well. Many early ChatGPT-like apps stopped right here.

But this setup has major limitations:

• The model knows only what’s in its training data

• No protection against unsafe or malicious prompts

• No cost control

• No monitoring

• No learning from users

This architecture is useful only until users start depending on it .

---

Why Real Products Need More Than a Model Call

Once users rely on your application:

• Wrong answers become costly

• Latency becomes noticeable

• Hallucinations become dangerous

• Abuse becomes inevitable

To fix these, teams don’t “replace the model.”
They add layers around it.

---

2. Step One: Enhancing Context (Giving the Model the Right Information)

Context Is Feature Engineering for LLMs

Large language models don’t magically know your:

• Company policies

• Internal documents

• User history

• Real-time data

To make models useful, you must construct context.

This can be done via:

• Text retrieval (documents, PDFs, chat history)

• Structured data retrieval (SQL, tables)

• Image retrieval

• Tool use (search APIs, weather, news, calculators)

This process—commonly called RAG—is analogous to feature engineering in traditional ML.

The model isn’t getting smarter; it’s getting better inputs .

---

Trade-offs in Context Construction

Not all context systems are equal.

Model APIs (OpenAI, Gemini, Claude):

• Easier to use

• Limited document uploads

• Limited retrieval control

Custom RAG systems:

• More flexible

• More engineering effort

• Require tuning (chunk size, embeddings, rerankers)

Similarly, tool support varies:

• Some models support parallel tools

• Some support long-running tools

• Some don’t support tools at all

As soon as context is added, your architecture already looks much more complex—and much more powerful.

---

3. Step Two: Guardrails (Protecting Users and Yourself)

Why Guardrails Are Non-Negotiable

AI systems fail in ways traditional software never did.

They can:

• Leak private data

• Generate toxic content

• Execute unintended actions

• Be tricked by clever prompts

Guardrails exist to reduce risk, not eliminate it (elimination is impossible).

There are two broad types:

• Input guardrails

• Output guardrails

---

Input Guardrails: Protecting What Goes In

Input guardrails prevent:

• Sensitive data leakage

• Prompt injection attacks

• System compromise

Common risks:

• Employees pasting secrets into prompts

• Tools accidentally retrieving private data

• Developers embedding internal policies into prompts

A common defense is PII detection and masking:

• Detect phone numbers, IDs, addresses, faces

• Replace them with placeholders

• Send masked prompt to external API

• Unmask response locally using a reverse dictionary

This allows functionality without leaking raw data .

---

Output Guardrails: Protecting What Comes Out

Models can fail in many ways.

Quality failures:

• Invalid JSON

• Hallucinated facts

• Low-quality responses

Security failures:

• Toxic language

• Private data leaks

• Brand-damaging claims

• Unsafe tool invocation

Some failures are easy to detect (empty output).
Others require AI-based scorers.

---

Handling Failures: Retries, Fallbacks, Humans

Because models are probabilistic:

• Retrying can fix many issues

• Parallel retries reduce latency

• Humans can be used as a last resort

Some teams route conversations to humans:

• When sentiment turns negative

• After too many turns

• When safety confidence drops

Guardrails always involve trade-offs:

• More guardrails → more latency

• Streaming responses → weaker guardrails

There is no perfect solution—only careful balancing.

---

4. Step Three: Routers and Gateways (Managing Complexity at Scale)

Why One Model Is Rarely Enough

As products grow, different queries require different handling:

• FAQs vs troubleshooting

• Billing vs technical support

• Simple vs complex tasks

Using one expensive model for everything is wasteful.

This is where routing comes in.

---

Routers: Choosing the Right Path

A router is usually an intent classifier.

Examples:

• “Reset my password” → FAQ

• “Billing error” → human agent

• “Why is my app crashing?” → troubleshooting model

Routers help:

• Reduce cost

• Improve quality

• Avoid out-of-scope conversations

Routers can also:

• Ask clarifying questions

• Decide which memory to use

• Choose which tool to call next

Routers must be:

• Fast

• Cheap

• Reliable

That’s why many teams use small models or custom classifiers .

---

Model Gateways: One Interface to Rule Them All

A model gateway is a unified access layer to:

• OpenAI

• Gemini

• Claude

• Self-hosted models

Benefits:

• Centralized authentication

• Cost control

• Rate limiting

• Fallback strategies

• Easier maintenance

Instead of changing every app when an API changes, you update the gateway once.

Gateways also become natural places for:

• Logging

• Analytics

• Guardrails

• Caching

---

5. Step Four: Caching (Reducing Latency and Cost)

Why Caching Matters in AI

AI calls are:

• Slow

• Expensive

• Often repetitive

Caching avoids recomputing answers.

There are two main types:

• Exact caching

• Semantic caching

---

Exact Caching: Safe and Simple

Exact caching reuses results only when inputs match exactly.

Examples:

• Product summaries

• FAQ answers

• Embedding lookups

Key considerations:

• Eviction policy (LRU, LFU, FIFO)

• Storage layer (memory vs Redis vs DB)

• Cache duration

Caching must be careful:

• User-specific data should not be cached globally

• Time-sensitive queries should not be cached

Mistakes here can cause data leaks .

---

Semantic Caching: Powerful but Risky

Semantic caching reuses answers for similar queries.

Process:

1. Embed query

2. Search cached embeddings

3. If similarity > threshold, reuse result

Pros:

• Higher cache hit rate

Cons:

• Incorrect answers

• Complex tuning

• Extra vector search cost

Semantic caching only works well when:

• Embeddings are high quality

• Similarity thresholds are well tuned

• Cache hit rate is high

Otherwise, it often causes more harm than good.

---

6. Step Five: Agent Patterns and Write Actions

Moving Beyond Linear Pipelines

Simple pipelines are sequential:

• Query → Retrieve → Generate → Return

Agentic systems introduce:

• Loops

• Conditional branching

• Parallel execution

Example:

• Generate answer

• Detect insufficiency

• Retrieve more data

• Generate again

This dramatically increases capability.

---

Write Actions: Power with Risk

Write actions allow models to:

• Send emails

• Place orders

• Update records

• Trigger workflows

They make systems vastly more useful—but vastly more dangerous.

Write actions must be:

• Strictly guarded

• Audited

• Often human-approved

Once write actions are added, observability becomes mandatory, not optional.

---

7. Monitoring and Observability: Seeing Inside the Black Box

Monitoring vs Observability

Monitoring:

• Tracks metrics

• Tells you something is wrong

Observability:

• Lets you infer why it’s wrong

• Without deploying new code

Good observability reduces:

• Mean time to detection (MTTD)

• Mean time to resolution (MTTR)

• Change failure rate (CFR)

---

Metrics That Actually Matter

Metrics should serve a purpose.

Examples:

• Format error rate

• Hallucination signals

• Guardrail trigger rate

• Token usage

• Latency (TTFT, TPOT)

• Cost per request

Metrics should correlate with business metrics:

• DAU

• Retention

• Session duration

If they don’t, you may be optimizing the wrong thing .

---

Logs and Traces: Debugging Reality

Logs:

• Record events

• Help answer “what happened?”

Traces:

• Reconstruct an entire request’s journey

• Show timing, costs, failures

In AI systems, logs should capture:

• Prompts

• Model parameters

• Outputs

• Tool calls

• Intermediate results

Developers should regularly read production logs—their understanding of “good” and “bad” outputs evolves over time.

---

Drift Detection: Change Is Inevitable

Things that drift:

• System prompts

• User behavior

• Model versions (especially via APIs)

Drift often goes unnoticed unless explicitly tracked.

Silent drift is one of the biggest risks in AI production.

---

8. User Feedback: Your Most Valuable Dataset

Why User Feedback Is Strategic

User feedback is:

• Proprietary

• Real-world

• Continuously generated

It fuels:

• Evaluation

• Personalization

• Model improvement

• Competitive advantage

This is the data flywheel.

---

Explicit vs Implicit Feedback

Explicit feedback:

• Thumbs up/down

• Ratings

• Surveys

Pros:

• Clear signal
  Cons:

• Sparse

• Biased

Implicit feedback:

• Edits

• Rephrases

• Regeneration

• Abandonment

• Conversation length

• Sentiment

Pros:

• Abundant
  Cons:

• Noisy

• Hard to interpret

Both are necessary.

---

Conversational Feedback Is Gold

Users naturally correct AI:

• “No, I meant…”

• “That’s wrong”

• “Be shorter”

• “Check again”

These corrections are:

• Preference data

• Evaluation signals

• Training examples

Edits are especially powerful:

• Original output = losing response

• Edited output = winning response

That’s free RLHF data.

---

Designing Feedback Without Annoying Users

Good feedback systems:

• Fit naturally into workflows

• Require minimal effort

• Can be ignored

Great examples:

• Midjourney’s image selection

• GitHub Copilot’s inline suggestions

• Google Photos’ uncertainty prompts

Bad feedback systems:

• Interrupt users

• Ask too often

• Demand explanation

---

Conclusion: Architecture and Feedback Are the Real AI Moats

Modern AI success is not about:

• The biggest model

• The cleverest prompt

It’s about:

• Thoughtful architecture

• Layered defenses

• Observability

• Feedback loops

• Continuous iteration

Models will commoditize.
APIs will change.
What remains defensible is how well you learn from users and adapt.

That is the real craft of AI engineering.

Tags: Artificial Intelligence,Generative AI,Agentic AI,Technology,Book Summary,

Inference Optimization: How AI Models Become Faster, Cheaper, and Actually Useful (Chapter 9 - AI Engineering - By Chip Huyen)

Download Book

<<< Previous Chapter Next Chapter >>>

Why “running” an AI model well is just as hard as building it

---

Introduction: Why Inference Matters More Than You Think

In the AI world, we spend a lot of time talking about training models—bigger models, better architectures, more data, and more GPUs. But once a model is trained, a much more practical question takes over:

How do you run it efficiently for real users?

This is where inference optimization comes in.

Inference is the process of using a trained model to produce outputs for new inputs. In simple terms:

• Training is learning

• Inference is doing the work

No matter how brilliant a model is, if it’s too slow, users will abandon it. If it’s too expensive, businesses won’t deploy it. Worse, if inference takes longer than the value of the prediction, the model becomes useless—imagine a stock prediction that arrives after the market closes.

As the chapter makes clear, inference optimization is about making AI models faster and cheaper without ruining their quality. And it’s not just a single discipline—it sits at the intersection of:

• Machine learning

• Systems engineering

• Hardware architecture

• Compilers

• Distributed systems

• Cloud infrastructure

This blog post explains inference optimization in plain language, without skipping the important details, so you can understand why AI systems behave the way they do—and how teams improve them in production .

---

1. Understanding Inference: From Model to User

Training vs Inference (A Crucial Distinction)

Every AI model has two distinct life phases:

1. Training – learning patterns from data

2. Inference – generating outputs for new inputs

Training is expensive but happens infrequently. Inference happens constantly.

Most AI engineers, application developers, and product teams spend far more time worrying about inference than training, especially if they’re using pretrained or open-source models.

---

What Is an Inference Service?

In production, inference doesn’t happen in isolation. It’s handled by an inference service, which includes:

• An inference server that runs the model

• Routing logic

• Request handling

• Possibly preprocessing and postprocessing

Model APIs like OpenAI, Gemini, or Claude are inference services. If you use them, you don’t worry about optimization details. But if you host your own models, you own the entire inference pipeline—performance, cost, scaling, and failures included .

---

Why Optimization Is About Bottlenecks

Optimization always starts with a question:

What is slowing us down?

Just like traffic congestion in a city, inference systems have chokepoints. Identifying these bottlenecks determines which optimization techniques actually help.

---

2. Compute-Bound vs Memory-Bound: The Core Bottleneck Concept

Two Fundamental Bottlenecks

Inference workloads usually fall into one of two categories:

Compute-bound

• Speed limited by how many calculations the hardware can perform

• Example: heavy mathematical computation

• Faster chips or more parallelism help

Memory bandwidth-bound

• Speed limited by how fast data can be moved

• Common in large models where weights must be repeatedly loaded

• Faster memory and smaller models help

This distinction is foundational to understanding inference optimization .

---

Why Language Models Are Often Memory-Bound

Large language models generate text one token at a time. For each token:

• The model must load large weight matrices

• Perform relatively little computation per byte loaded

This makes decoding (token generation) memory bandwidth-bound.

---

Prefill vs Decode: Two Very Different Phases

Transformer-based language model inference has two stages:

1. Prefill

   • Processes input tokens in parallel

   • Compute-bound

   • Determines how fast the model “understands” the prompt

2. Decode

   • Generates one output token at a time

   • Memory bandwidth-bound

   • Determines how fast the response appears

Because these phases behave differently, modern inference systems often separate them across machines for better efficiency .

---

3. Online vs Batch Inference: Latency vs Cost

Two Types of Inference APIs

Most providers offer:

• Online APIs – optimized for low latency

• Batch APIs – optimized for cost and throughput

---

Online Inference

Used for:

• Chatbots

• Code assistants

• Real-time interactions

Characteristics:

• Low latency

• Users expect instant feedback

• Limited batching

Streaming responses (token-by-token) reduce perceived waiting time but come with risks—users might see bad outputs before they can be filtered.

---

Batch Inference

Used for:

• Synthetic data generation

• Periodic reports

• Document processing

• Data migration

Characteristics:

• Higher latency allowed

• Aggressive batching

• Much lower cost (often ~50% cheaper)

Unlike traditional ML, batch inference for foundation models can’t precompute everything because user prompts are open-ended .

---

4. Measuring Inference Performance: Metrics That Actually Matter

Latency Is Not One Number

Latency is best understood as multiple metrics:

Time to First Token (TTFT)

• How long users wait before seeing anything

• Tied to prefill

• Critical for chat interfaces

Time Per Output Token (TPOT)

• Speed of token generation after the first token

• Determines how fast long responses feel

Two systems with the same total latency can feel very different depending on TTFT and TPOT trade-offs.

---

Percentiles, Not Averages

Latency is a distribution.

A single slow request can ruin the average. That’s why teams look at:

• p50 (median)

• p90

• p95

• p99

Outliers often signal:

• Network issues

• Oversized prompts

• Resource contention .

---

Throughput and Cost

Throughput measures how much work the system does:

• Tokens per second (TPS)

• Requests per minute (RPM)

Higher throughput usually means lower cost—but pushing throughput too hard can destroy user experience.

---

Goodput: Throughput That Actually Counts

Goodput measures how many requests meet your service-level objectives (SLOs).

If your system completes 100 requests/minute but only 30 meet latency targets, your goodput is 30, not 100.

This metric prevents teams from optimizing the wrong thing.

---

5. Hardware: Why GPUs, Memory, and Power Dominate Costs

What Is an AI Accelerator?

An accelerator is specialized hardware designed for specific workloads.

For AI, the dominant accelerator is the GPU, designed for massive parallelism—perfect for matrix multiplication, which dominates neural network workloads.

---

Why GPUs Beat CPUs for AI

• CPUs: few powerful cores, good for sequential logic

• GPUs: thousands of small cores, great for parallel math

More than 90% of neural network computation boils down to matrix multiplication, which GPUs excel at .

---

Memory Hierarchy Matters More Than You Think

Modern accelerators use multiple memory layers:

• CPU DRAM (slow, large)

• GPU HBM (fast, smaller)

• On-chip SRAM (extremely fast, tiny)

Inference optimization is often about moving data less and smarter across this hierarchy.

---

Power Is a Hidden Bottleneck

High-end GPUs consume enormous energy:

• An H100 running continuously can use ~7,000 kWh/year

• Comparable to a household’s annual electricity use

This makes power—and cooling—a real constraint on AI scaling .

---

6. Model-Level Optimization: Making Models Lighter and Faster

Model Compression Techniques

Several techniques reduce model size:

Quantization

• Reduce numerical precision (FP32 → FP16 → INT8 → INT4)

• Smaller models

• Faster inference

• Lower memory bandwidth use

Distillation

• Train a smaller model to mimic a larger one

• Often surprisingly effective

Pruning

• Remove unimportant parameters

• Creates sparse models

• Less common due to hardware limitations

Among these, weight-only quantization is the most widely used in practice .

---

The Autoregressive Bottleneck

Generating text one token at a time is:

• Slow

• Expensive

• Memory-bandwidth heavy

Several techniques attempt to overcome this fundamental limitation.

---

Speculative Decoding

Idea:

• Use a smaller “draft” model to guess future tokens

• Have the main model verify them in parallel

If many draft tokens are accepted, the system generates multiple tokens per step—dramatically improving speed without hurting quality.

This technique is now widely supported in modern inference frameworks.

---

Inference with Reference

Instead of generating text the model already knows (e.g., copied context), simply reuse tokens from the input.

This works especially well for:

• Document Q&A

• Code editing

• Multi-turn conversations

It avoids redundant computation and speeds up generation.

---

Parallel Decoding

Some techniques try to generate multiple future tokens simultaneously.

Examples:

• Lookahead decoding

• Medusa

These approaches are promising but complex, requiring careful verification to ensure coherence.

---

7. Attention Optimization: Taming the KV Cache Explosion

Why Attention Is Expensive

Each new token attends to all previous tokens.

Without optimization:

• Computation grows quadratically

• KV cache grows linearly—but still huge

For large models and long contexts, the KV cache can exceed model size itself .

---

KV Cache Optimization Techniques

Three broad strategies exist:

Redesigning Attention

• Local window attention

• Multi-query attention

• Grouped-query attention

• Cross-layer attention

These reduce how much data must be stored and reused.

---

Optimizing KV Cache Storage

Frameworks like vLLM introduced:

• PagedAttention

• Flexible memory allocation

• Reduced fragmentation

Other approaches:

• KV cache quantization

• Adaptive compression

• Selective caching

---

Writing Better Kernels

Instead of changing algorithms, optimize how computations are executed on hardware.

The most famous example:

• FlashAttention

• Fuses multiple operations

• Minimizes memory access

• Huge speedups in practice

---

8. Service-Level Optimization: Making the Whole System Work

Batching: The Simplest Cost Saver

Batching combines multiple requests:

• Improves throughput

• Reduces cost

• Increases latency

Types:

• Static batching

• Dynamic batching

• Continuous batching

The trick is batching without hurting latency too much.

---

Compilers and Kernels

Modern inference relies heavily on:

• Compilers (torch.compile, XLA, TensorRT)

• Hardware-specific kernels

These translate high-level model code into highly optimized machine instructions.

Many companies treat their kernels as trade secrets because they directly translate into cost advantages .

---

Conclusion: Inference Optimization Is the Real Production Skill

Training models gets headlines. Inference optimization keeps businesses alive.

Inference optimization:

• Determines user experience

• Dominates long-term cost

• Requires cross-disciplinary thinking

• Is where real-world AI engineering happens

As models become commoditized, efficient inference becomes a competitive moat.

The future of AI won’t be decided by who trains the biggest model—but by who can run models fastest, cheapest, and most reliably.

---

Addendum

My prompt:

Role: You an expert in AI Engineering and a prolific writer. 
Task: Spin this attached chapter 9 as a blog post in layman terms 
Rules: Organize the post in 7-8 sections with subsections as needed 
Note: Blog post should be about 6000 to 7000 words long
Note: Try not to miss any important section or details

Tags: Artificial Intelligence,Generative AI,Agentic AI,Technology,Book Summary,

Ch 7 - Finetuning - Chapter Summary using GPT 5.2 (Dec 2025)

Download Book

<<< Previous Chapter Next Chapter >>>

Prompt

Role: You are an Expert Professor of Agentic AI
Task: Explain the attached PDF (Chapter 7 - Finetuning) in layman terms
Strictly Follow: Do not miss out any details while simplying things.
Word Limit: Respond in about 5000 to 7000 words

Response

Chapter 7 — Finetuning (Explained Simply, Without Losing Depth)

---

1. What “Finetuning” Really Means (Big Picture)

Let’s start with the core idea.

Finetuning is the process of taking an already-trained AI model and teaching it to behave better for a specific purpose.

Think of it like this:

• Pre-training = going to school and learning everything (language, math, reasoning).

• Prompting = giving instructions like “please answer like a lawyer.”

• Finetuning = sending the model for professional training so it permanently learns that behavior.

Instead of just telling the model what to do at runtime (prompting), finetuning changes the model’s internal parameters (weights) so it naturally behaves the way you want.

This is fundamentally different from prompt-based methods discussed in earlier chapters, which do not change the model internally .

---

2. Why Finetuning Exists at All

Large language models already know a lot. So why finetune?

Because:

1. They don’t always follow instructions reliably

2. They may not produce output in the exact format you need

3. They may not specialize well in niche or proprietary tasks

4. They may behave inconsistently across prompts

Finetuning helps with:

• Instruction following

• Output formatting (JSON, YAML, code, DSLs)

• Domain specialization (legal, medical, finance)

• Bias mitigation

• Safety alignment

In short:

Prompting changes what you ask.
Finetuning changes how the model thinks.

---

3. Finetuning Is Transfer Learning

Finetuning is not a new idea. It’s a form of transfer learning, first described in the 1970s.

Human analogy:

If you already know how to play the piano, learning the guitar is easier.

AI analogy:

If a model already knows language, learning legal Q&A requires far fewer examples.

This is why:

• Training a legal QA model from scratch might need millions of examples

• Finetuning a pretrained model might need hundreds

This efficiency is what makes foundation models so powerful .

---

4. Types of Finetuning (Conceptual Overview)

4.1 Continued Pre-Training (Self-Supervised Finetuning)

Before expensive labeled data, you can:

• Feed the model raw domain text

• No labels required

• Examples:

  • Legal documents

  • Medical journals

  • Vietnamese books

This helps the model absorb domain language cheaply.

This is called continued pre-training.

---

4.2 Supervised Finetuning (SFT)

Now you use (input → output) pairs.

Examples:

• Instruction → Answer

• Question → Summary

• Text → SQL

This is how models learn:

• What humans expect

• How to respond politely

• How to structure answers

Instruction data is expensive because:

• It needs domain expertise

• It must be correct

• It must be consistent

---

4.3 Preference Finetuning (RLHF-style)

Instead of a single “correct” answer, you give:

• Instruction

• Preferred answer

• Less-preferred answer

The model learns human preference, not just correctness.

---

4.4 Long-Context Finetuning

This extends how much text a model can read at once.

But:

• It requires architectural changes

• It can degrade short-context performance

• It’s harder than other finetuning methods

Example: Code Llama extended context from 4K → 16K tokens .

---

5. Who Finetunes Models?

Model developers:

• OpenAI, Meta, Mistral

• Release multiple post-trained variants

Application developers:

• Usually finetune already post-trained models

• Less work, less cost

The more refined the base model, the less finetuning you need.

---

6. When Should You Finetune?

This is one of the most important sections.

Key principle:

Finetuning should NOT be your first move.

It requires:

• Data

• GPUs

• ML expertise

• Monitoring

• Long-term maintenance

You should try prompting exhaustively first.

Many teams rush to finetuning because:

• Prompts were poorly written

• Examples were unrealistic

• Metrics were undefined

After fixing prompts, finetuning often becomes unnecessary .

---

7. Good Reasons to Finetune

7.1 Task-Specific Weaknesses

Example:

• Model handles SQL

• But fails on your company’s SQL dialect

Finetuning on your dialect fixes this.

---

7.2 Structured Outputs

If you must get:

• Valid JSON

• Compilable code

• Domain-specific syntax

Finetuning helps more than prompting.

---

7.3 Bias Mitigation

If a model shows bias:

• Gender bias

• Racial bias

Carefully curated finetuning data can reduce it.

---

7.4 Small Models Can Beat Big Models

A finetuned small model can outperform a huge generic model on a narrow task.

Example:

• Grammarly finetuned Flan-T5

• Beat a GPT-3 variant

• Used only 82k examples

• Model was 60× smaller

---

8. Reasons NOT to Finetune

8.1 Performance Trade-offs

Finetuning for Task A can degrade Task B.

This is called catastrophic interference.

---

8.2 High Up-Front Cost

You need:

• Annotated data

• ML knowledge

• Infrastructure

• Serving strategy

---

8.3 Ongoing Maintenance

New base models keep appearing.

You must decide:

• When to switch

• When to re-finetune

• When gains are “worth it”

---

9. Finetuning vs RAG (Critical Distinction)

This chapter makes a very important rule:

RAG is for facts.
Finetuning is for behavior.

Use RAG when:

• Model lacks information

• Data is private

• Data is changing

• Answers must be up-to-date

Use finetuning when:

• Model output is irrelevant

• Format is wrong

• Syntax is incorrect

• Style is inconsistent

Studies show:

• RAG often beats finetuning for factual Q&A

• RAG + base model > finetuned model alone .

---

10. Why Finetuning Is So Memory-Hungry

This is the core technical bottleneck.

Inference:

• Only forward pass

• Need weights + activations

Training (Finetuning):

• Forward pass

• Backward pass

• Gradients

• Optimizer states

Each trainable parameter requires:

• The weight

• The gradient

• 1–2 optimizer values (Adam uses 2)

So memory explodes.

---

11. Memory Math (Intuition)

Inference memory ≈

nginx
weights × 1.2

Example:

• 13B params

• 2 bytes each → 26 GB

• Total ≈ 31 GB

---

Training memory =

nginx
weights + activations + gradients + optimizer states

Example:

• 13B params

• Adam optimizer

• 16-bit precision

Just gradients + optimizer states = 78 GB

That’s why finetuning is hard.

---

12. Numerical Precision & Quantization

Floating-point formats:

• FP32 (4 bytes)

• FP16 (2 bytes)

• BF16

• TF32

Lower precision = less memory, faster compute.

---

Quantization

Quantization = reducing precision.

Examples:

• FP32 → FP16

• FP16 → INT8

• INT8 → INT4

This dramatically reduces memory.

---

Post-Training Quantization (PTQ)

Most common.

• Train in high precision

• Quantize for inference

---

Quantization-Aware Training (QAT)

Simulates low precision during training.

• Improves low-precision inference

• Doesn’t reduce training cost

---

Training Directly in Low Precision

Hard but powerful.

• Character.AI trained models fully in INT8

• Eliminated mismatch

• Reduced cost

---

13. Why Full Finetuning Doesn’t Scale

Example:

• 7B model

• FP16 weights = 14 GB

• Adam optimizer = +42 GB

• Total = 56 GB (without activations)

Most GPUs cannot handle this.

Hence the rise of:

---

14. Parameter-Efficient Finetuning (PEFT)

Idea:

Update fewer parameters, get most of the benefit.

Instead of changing everything:

• Freeze most weights

• Add small trainable components

---

15. Partial Finetuning (Why It’s Not Enough)

Freezing early layers helps memory, but:

• You need ~25% of parameters

• To match full finetuning

Still expensive.

---

16. Adapter-Based PEFT

Houlsby et al. introduced adapters:

• Small modules inserted into each layer

• Only adapters are trained

• Base model is frozen

Result:

• ~3% parameters

• ~99.6% performance

Downside:

• Extra inference latency

---

17. LoRA (Low-Rank Adaptation)

LoRA solved adapter latency.

Core idea:

Instead of adding layers, modify weight matrices mathematically.

A big matrix:

nginx
W

Becomes:

css
W + (A × B)

Where:

• A and B are small

• Only A and B are trained

• W stays frozen

This uses low-rank factorization.

---

Why LoRA Works

• Neural weight updates are often low-rank

• You don’t need to update everything

• Rank 4–64 is usually enough

Example:

• GPT-3 finetuned with 0.0027% parameters

• Matched full finetuning performance

---

18. Where LoRA Is Applied

Mostly to:

• Query (Wq)

• Key (Wk)

• Value (Wv)

• Output (Wo)

Applying LoRA to query + value often gives best returns.

---

19. Serving LoRA Models

Two strategies:

1. Merge LoRA into base model

• Faster inference

• More storage

2. Keep adapters separate

• Slight latency cost

• Massive storage savings

• Enables multi-LoRA serving

Apple uses this approach to serve many features from one base model.

---

20. QLoRA (Quantized LoRA)

QLoRA = LoRA + 4-bit quantization.

• Base weights stored in 4-bit

• Dequantized to BF16 during compute

• Uses NF4 format

• Uses CPU–GPU paging

Result:

• 65B models finetuned on single 48 GB GPU

Downside:

• Slower training due to quantization overhead

---

21. Model Merging

Instead of finetuning one model:

• Finetune multiple models separately

• Merge them later

Benefits:

• Avoids catastrophic forgetting

• Reduces memory

• Enables multi-task models

• Useful for on-device deployment

Model merging generalizes ensembles, but merges weights instead of outputs.

---

22. Final Summary (Mental Model)

Finetuning is powerful—but costly.

• Start with prompting

• Add RAG for knowledge

• Finetune for behavior

• Use PEFT whenever possible

• Prefer LoRA or QLoRA

• Consider model merging for multi-task systems

The hard part isn’t finetuning itself.

The hard part is data, evaluation, and long-term maintenance.

That’s why the next chapter focuses entirely on data.

---