From 2cc8d4b88904541544fc39c73af9d775778c55da Mon Sep 17 00:00:00 2001 From: Shayan Rais Date: Mon, 4 May 2026 19:34:53 +0500 Subject: [PATCH] add slide on temperature=0 nondeterminism and Source: footers to slides 7, 9, 10, 15, 23, 46 Co-Authored-By: Claude --- .../claude-code-best-practice/index.html | 192 ++++++++++++------ 1 file changed, 135 insertions(+), 57 deletions(-) diff --git a/presentation/claude-code-best-practice/index.html b/presentation/claude-code-best-practice/index.html index bed2a97..c6ca716 100644 --- a/presentation/claude-code-best-practice/index.html +++ b/presentation/claude-code-best-practice/index.html @@ -125,6 +125,7 @@ .pillar-mini-card .pmc-body { font-size: 0.72rem; line-height: 1.35; color: #333; margin-top: 5px; display: -webkit-box; -webkit-line-clamp: 5; -webkit-box-orient: vertical; overflow: hidden; } .pillar-mini-card .pmc-badge { display: inline-block; font-size: 0.65rem; font-weight: 600; padding: 2px 7px; border-radius: 999px; margin-top: 6px; white-space: nowrap; align-self: flex-start; } .pillar-mini-card.inactive { opacity: 0.55; } + @@ -402,6 +403,7 @@ Boris Cherny (creator of Claude Code) — different teams use Claude Code completely differently.
There is no single “correct” way. But there are patterns worth understanding. +

Source: Boris Cherny on X — tweet 1 · tweet 2 · tweet 3

@@ -471,9 +473,6 @@

Each run the model samples — temperature controls how widely it samples.

- -

Bender, Gebru, McMillan-Major, Mitchell — On the Dangers of Stochastic Parrots (2021)

-
stochastic @@ -484,12 +483,86 @@ confident pattern-matching
+ + +

Source: Bender, Gebru, McMillan-Major, Mitchell — On the Dangers of Stochastic Parrots (2021)

- +
+ + +
+

Even temperature = 0 isn’t deterministic.

+

You set it to zero. You expect the same answer every time. You’re wrong.

+
+ + +
+ + +
+
The data point
+ + +
+ + +
+
Without fix
+
+
80
+
unique completions
+
out of 1,000 calls
+
+
+ + +
→
+ + +
+
With fix
+
+
1
+
unique completion
+
out of 1,000 calls
+
+
+ +
+ +

Qwen3-235B at temperature = 0 — first divergence at token 103 (“Queens, New York” vs “New York City”)

+
+ + +
+
Why it happens
+

Server load varies → batch size varies → kernel reductions reorder → numerics shift. Not GPU randomness — arithmetic order.

+
+ + +
+
The fix
+

Batch-invariant kernels → consistent reduction order → identical numerics every run.

+
+ +
+ + +

Determinism is engineered in — at every layer.

+ + +

Source: Thinking Machines — Defeating Nondeterminism in LLM Inference (2025)

+ +
+ + + + +
@@ -546,9 +619,9 @@
- + -
+

🧠 Models — e.g. Opus, GPT, Gemini

@@ -585,9 +658,9 @@
- + -
+

🧠 Limitations

The raw model has no real-time access — no internet, no files, no clock.

@@ -596,9 +669,9 @@
- + -
+
@@ -761,9 +834,9 @@
- + -
+

⚡ Tool Calling — how the harness reaches the world

Turn — one round from the user’s view: you ask, the assistant answers. The entire flow above — your request, the assistant’s tool calls, and the final reply — is one turn.

Inference — one call to the language model. The model wakes up, reads the input it was given, writes a reply, then forgets everything. Every arrow touching the “Language Model” column above is a separate inference. One turn can contain many inferences.

+ + +

Source: Anthropic — Claude Code in Action: What is a coding assistant?

- + -
+

💪 Harness — the body around the brain

@@ -826,9 +902,9 @@
- + -
+

💪 Harness — the body around the brain

@@ -863,9 +939,9 @@
- + -
+

🎉 Yayyyyy! Problem solved with harness

The harness reaches out via WebSearch and fetches a real answer from live sources.

@@ -874,9 +950,9 @@
- + -
+
?

Really?

@@ -884,9 +960,9 @@
- + -
+

💪 Non-determinism — Doesn’t always use its tools

Similar prompt — but this time the model decided not to use the tool.

@@ -895,9 +971,9 @@
- + -
+

💪 Non-determinism — Tools can fail

The model first tried one source — it failed (403) — so it fell back to another.

@@ -906,9 +982,9 @@
- + -
+

🚨 Problem Statement

  1. @@ -923,20 +999,21 @@
- + -
+

Vibe Coding

Andrej Karpathy's Feb 3 2025 tweet coining 'vibe coding' — 'fully give in to the vibes, embrace exponentials, and forget that the code even exists'

Andrej Karpathy — OpenAI founding team · former Director of AI at Tesla · founder of Eureka Labs.

+

Source: Andrej Karpathy on X (Feb 3, 2025)

- + -
+

Vibe Coding vs Agentic Engineering

@@ -1005,7 +1082,7 @@ todoapp/ -
+

👤 Agents

@@ -1063,7 +1140,7 @@ todoapp/
-
+

Create your first agent — /agents

@@ -1117,7 +1194,7 @@ todoapp/
-
+

Demo