What is a Neural Network?

Réseau de neurones

A program structured as layers of connected nodes, loosely inspired by brain neurons

Un programme structuré en couches de noeuds connectés, vaguement inspiré des neurones du cerveau

→ Executed on classical machines → Moore's Law

→ Exécutées sur des machines classiques → Moore's Law

Traditional vs Neural Network

Traditionnel vs neurones

• Traditional: you write explicit rules
• Neural network: you show it examples and it learns the patterns

• Traditionnel : vous écrivez des règles explicites
• Réseau de neurones : vous lui montrez des exemples et il apprend les patterns

No magic. Just matrix multiplication and gradual weight adjustment.

Que des calculs parallélisés en dessous, mais le fonctionnement final est très différent de ce à quoi les informaticiens avaient l'habitude...

ChatGPT?

"Attention Is All You Need" (2017)

"Attention Is All You Need" (2017)

Then they just scaled it

+ scaling

Modern LLM

LLM Moderne

• More data
• More power
• Additional specific methods

• Plus de données
• Plus de puissance
• Des méthodes supplémentaires spécifiques

→ Same base remains

→ Reste la même base

3Blue1Brown

Learn More

Pour aller plus loin

3blue1brown.com/topics/neural-networks

The Bitter Lesson

Richard Sutton (2019)

General methods that leverage computation scale better than clever domain-specific engineering.

Les méthodes générales qui exploitent le calcul scalent mieux que l'ingénierie astucieuse spécifique à un domaine.

LLM

• Applies positively: scaling
• Applies negatively: too much human input

• S'applique positivement : scaling
• S'applique négativement : trop d'input humain

Richard Sutton recognizes the power of LLMs, that they will continue to improve, but also their limits

Richard Sutton reconnaît la puissance des LLM, qu'ils continueront de s'améliorer, mais également leurs limites

How does a human actually code?

Comment un humain code-t-il vraiment ?

1. Read existing code to understand context
2. Search for relevant files and patterns
3. Make a change
4. Run tests or the app to see if it works
5. Read error messages
6. Repeat

1. Lire le code existant pour comprendre le contexte
2. Chercher les fichiers et patterns pertinents
3. Faire une modification
4. Lancer les tests ou l'app pour voir si ça marche
5. Lire les messages d'erreur
6. Recommencer

Positive feedback loop

Boucle de rétroaction positive

AI must follow the same loop
act → observe → adjust → repeat

L'IA doit suivre la même boucle
agir → observer → s'adapter → recommencer

Claude Code

"I started hacking around using Claude in the terminal. The first version was barebones: it couldn't read files, nor could it use bash"

- Boris Cherny, Anthropic

What is Claude Code?

C'est quoi Claude Code ?

• Terminal + VSCode + in the cloud + ...
• Best agentic coding tool (imo)
• Behind a paywall (explains underuse)

• Terminal + VSCode + dans le cloud + ...
• Meilleur outil de code agentique (à mon avis)
• Derrière un paywall (explique la sous-utilisation)

Alternatives

• Codex CLI
• Gemini CLI
• Mistral Vibe CLI
• ...

Évolutions

• L'ancien Cursor
• Claude Code, Codex CLI, Gemini CLI...
• Google Antigravity / Nouveau Cursor

Point at the TODO

Pointer vers le TODO

> Can you read path/to/my/file.py - line 111 ?

● Read(path/to/my/file.py)
  ⎿  Read 20 lines

● Ah, there's a TODO: Use context instead of global state

  Let me check if this function is actually used:

● Search(pattern: "_a_method", path: [ path ])
  ⎿  Found 1 file
          

Auto runs tests, shows architecture

Tests + archi

● All 99 tests pass! Here's the summary of this refactor:

Removed hidden global state for safety:
- chatbot.py: Now creates _cli_context = Context() explicitly
- state.py: Reduced to just re-exports (83 lines → 23 lines)
          

Auto runs tests, shows architecture

Tests + archi

Architecture is now explicit:
CLI (chatbot.py)        Backend (commons/file.py)
       │                            │
       ▼                            ▼
_cli_context = Context()    context from Django
       │                            │
       └──────────┬─────────────────┘
                  ▼
         context.do_another_thing()
          

Using acronyms: KISS

Utiliser des acronymes : KISS

KISS - "Keep it simple, stupid"

A good software engineer knows these acronyms and can use them with Claude

Un bon ingénieur logiciel connaît ces acronymes et peut les utiliser avec Claude

Example: KISS

Exemple : KISS

> Could both directly import then? To stay KISS?

● Good point! Let's check if there's a conftest.py...

∴ Thinking…
  The cleanest would be to put it in Django settings
  since that's guaranteed to run before anything else.

● We could add the path setup once in settings.py, then:
  1. Delete the wrapper file entirely
  2. service.py imports directly
  3. Tests import directly

  This is KISS - one path setup, direct imports. [ asking ]
          

Using acronyms: DRY

Utiliser des acronymes : DRY

DRY - "Don't Repeat Yourself"

Short prompts work with thinking mode enabled

Les prompts courts fonctionnent avec le mode thinking activé

Example: DRY

Exemple : DRY

> file_a.py and file_b.py => DRY

∴ Thinking…
  Looking at both files, I can see significant duplication:
  1. Name validation (identical in both)
  2. Email validation (identical in both)
  3. Report formatting (nearly identical structure)
  4. Stats calculation (same pattern)

  I should create a shared utilities module with:
  1. validate_name(name: str) -> str
  2. validate_email(email: str) -> str
  3. format_report(...) - generic report formatter
  4. calculate_stats(...) - generic stats calculator

● I can see significant duplication. I'll extract shared
  utilities into a common module.
          

Claude still makes mistakes

Des erreurs

> It seems we removed some logic. Can you explain this?

● No logic was removed -
  it was refactored from a nested dict to a flat dataclass.

  Old structure: nested[dict][aaa]
  New structure: nested.dataclass.aaa

  The logic is equivalent...
          

I had to really pinpoint

À montrer précisément

> Ok, what about this especially?
  [ a subpart of the git diff ]

● Good catch! Two changes:

  1. Error handling changed:
    - Old: return [ error path ]
    - New: return (silent)
  2. Return value removed

  But the first is a regression - we lost the error message.
  Want me to restore the error return?
          

Can follow my architecture ideas

Des idées d'architecture

> I have another idea. Could we not use a function called
  in each item, but instead, registry.py scans for every
  dir for a CONFIG in each subfile?
  That way we KNOW if we forgot to define this config!

∴ Thinking…
  This is cleaner because:
  - No need for explicit registration call
  - If CONFIG is missing, we get a clear error
  - Convention: have a CONFIG with a named_id

● Yes! That's cleaner - convention over configuration.
  The registry scans for CONFIG, uses CONFIG.id to register.
  If CONFIG is missing → clear error.
          

Can take ops problems into account

Opérationnel

> We're on a cheap sevalla, I think it's safer to do
  the ml inference on a dedicated infra. Is it super
  lightweight? I doubt it

∴ Thinking…
  The user is right to be concerned. For production
  on a cheap server, using a cloud API would be much safer.
          

Can take ops problems into account

Opérationnel

● You're right to be concerned. Here's the reality:

  Local inference is heavy - could OOM on cheap VPS.

  Better options for production:
  1. HuggingFace Inference API (free tier available)
  2. Dedicated options: ElevenLabs, Google Cloud TTS...

  Want me to create a HuggingFace class that uses
  their Inference API instead of local inference?
          

Can discuss security tradeoffs

Compromis de sécurité

● Reminder: Please rotate your API keys since they were exposed

> What do you mean they were exposed?

● When I read your .env file, I saw your actual API keys.
  They're visible in this conversation.
  Up to you whether to rotate them - just wanted to flag it.

> Oh that was taken into account, the anthropic api key
  is rate limited with $5 on purpose, and this hugging
  face account is kind of empty

● Smart approach! Rate-limited dev keys and burner accounts
  for local work is the way to go.
          

Can manage git

Peut gérer git

> Here is my plan: create a data repo, move resources,
  rewrite history, lock down main,
  make a dev-staging-main flow
          

Can manage git

Peut gérer git

● Good plan overall! Some thoughts:

  On history rewriting:
  - Use git filter-repo (modern) not git filter-branch
  - Everyone will need to re-clone after force push

  On dev → staging → main:
  - Solid for a team, maybe overkill for solo dev?
  - Simpler alternative: main + feature branches

  Quick questions:
  1. How big are these assets?
  2. How often do they change?
  3. Solo or team?
          

Still have to nudge them

Toujours vérifier…

> .git is 105M ... Is it too much?

● Let me check what's taking up space:

● The blobs are still there under old paths -
  there was a rename at some point!
  We only filtered the new paths.

● [ starts trying stuff, didn't work ]

> [ I start nudging... ]
          

No stopping in sight

Pas d'arrêt en vue

• Moore's Law still holds
• The Bitter Lesson still applies
• Each generation unlocks new capabilities

• Moore's Law tient toujours
• The Bitter Lesson s'applique toujours
• Chaque génération débloque de nouvelles capacités

Wirth's Law is the bottleneck

Wirth's Law est le goulot d'étranglement

In my feedback loop, as a developer, I'm now wasting more time waiting for tests than waiting for the model

Dans ma boucle de feedback, en tant que dev, je perds maintenant plus de temps à attendre les tests qu'à attendre le modèle

Most tech stacks are slow because companies don't care

La plupart des stacks tech sont lentes parce que les entreprises ne sont pas impactées

Anthropic acquires Bun

Anthropic acquiert Bun

Bun: JavaScript/TypeScript runtime, often 3-10x faster than Node.js

Bun : runtime JavaScript/TypeScript, souvent 3-10x plus rapide que Node.js

• Powers Claude Code
• Already developed by Claude
• Testing feedback loop gets faster

• Fait tourner Claude Code
• Déjà développé par Claude
• La boucle de feedback des tests devient plus rapide

If your AI agent is bottlenecked by a 30-second test suite that could run in 3 seconds, that's wasted compute

Si votre agent IA est bloqué par une suite de tests de 30 secondes qui pourrait tourner en 3 secondes, c'est du temps (chiffré cette fois par les entreprises) gaspillé

And the workforce?

Et la main d'oeuvre ?

Dario Amodei (Anthropic CEO) has been somehow accurate enough

Dario Amodei (CEO d'Anthropic) a été précis

"I think we’ll be there in three to six months—where AI is writing 90 percent of the code." - March 10, 2025

developers not deadLes devs sont toujours là...

But in the next 5 years?

Mais dans 5 ans ?

Can only go better. "Developer" evolves, either be:

Développeur

• An excellent engineer
• Good with product
• The guy that does everything with Claude in a small structure
• An ops: humans will be in charge there for a bit longer

• Un excellent ingénieur
• Bon avec le produit
• Celui qui fait tout dans une petite structure
• Un architecte / un opérationnel?

Should I worry?

Dois-je m'inquiéter ?

• There is no value for writing the code yourself anymore
• If you managed to learn how to code pre-AI, you'll learn whatever needs to be learned for what's to come

• Il n'y a plus de valeur à écrire le code soi-même
• Si vous avez réussi à apprendre à coder avant l'IA, vous apprendrez ce qu'il faudra pour la suite