Components

Every MDX component used across the writing on this site.

ActivationCascade

Neurons firing in cascading layers as a prompt activates the network.

<ActivationCascade />

datetime / locale concept neurons dominate the final layer.

ActivationEnergy

The energy barrier metaphor for getting a stalled task moving.

<ActivationEnergy />

AttentionHeads

Multi-head attention split across distinct semantic patterns.

<AttentionHeads />

Each token attends to its immediate left neighbors. This head tracks local word order.

AttentionHeatmap

Heatmap overlay showing where the model focuses inside a code sample.

<AttentionHeatmap />

1async def process_order(request):
2    user = get_user(request.headers["auth_token"])
3    items = request.json["items"]
4    total = sum(item["price"] * item["qty"] for item in items)
5
6    order = Order(
7        user_id=user.id,
8        total=total,
9        created_at=datetime.now(),
10        items=items
11    )
12
13    db.orders.insert(order)
14
15    response = requests.post(PAYMENT_API, json={
16        "amount": total,
17        "currency": "USD",
18        "user": user.email
19    })
20
21    send_email(user.email, f"Order confirmed: ${total}")
22    return {"status": "success", "order_id": order.id}

Attention spread evenly. The timezone bug on line 9 gets the same weight as boilerplate.

AttentionWeights

Bar visualization of how attention weight redistributes between tokens.

<AttentionWeights />

Hover a row to see what that token attends to.

BlindSpotVenn

Venn diagram of overlapping and disjoint model blind spots.

<BlindSpotVenn />

Same-family reviewers pile up on the same bugs. 70 redundant catches, 3 unique contributions, 17 escape.

CodeLenses

Switchable review lenses applied to the same code block.

<CodeLenses />

1async def process_order(request):
2    user = get_user(request.headers["auth_token"])
3    items = request.json["items"]
4    total = sum(item["price"] * item["qty"] for item in items)
5
6    order = Order(
7        user_id=user.id,
8        total=total,
9        created_at=datetime.now(),
10        items=items
11    )
12
13    db.orders.insert(order)
14
15    response = requests.post(PAYMENT_API, json={
16        "amount": total,
17        "currency": "USD",
18        "user": user.email
19    })
20
21    send_email(user.email, f"Order confirmed: ${total}")
22    return {"status": "success", "order_id": order.id}

Click a lens to reveal issues invisible to the other passes.

CompoundReturn

Compounding returns curve for early-finished work.

<CompoundReturn />

ContextDilution

How signal-to-noise drops as context fills with low-relevance tokens.

<ContextDilution />

Context length

5.4k tokens

Your 50-token prompt vs everything else in context

Per-token weight

1 / 5,449

Your prompt's share

0.918%

Your prompt is one voice in a crowded room. Specificity has to do more work.

DepthLayers

Stacked model layers with depth-wise processing illustrated.

<DepthLayers />

Generic reviews catch surface-level issues. Deeper problems go undetected.

LatentSpaceNavigation

Walking a path through latent space between two concepts.

<LatentSpaceNavigation />

Latent space, projected to 2D

Design 5default

Magazine columns

Deep exploration. Atypical structure the model would not propose first.

Each design penalizes the next. By design 5, the model is exploring regions it would never reach cold.

MakerCheckerFlow

Two-model maker/checker pipeline with handoff between roles.

<MakerCheckerFlow />

Same-model reviewer catches 1 bug the scanners already would have. 4 bugs ship.

NextTokenDistribution

Probability distribution over candidate next tokens.

<NextTokenDistribution />

The capital of France is

·Paris

94.0%

·the

1.8%

·a

1.2%

·located

0.8%

·known

0.5%

·also

0.4%

One winner. "Paris" claims 94% of the mass.

PathLock

Lock-in effect once an early choice constrains downstream tokens.

<PathLock />

Each commit conditions the next distribution. The path narrows.

PromptCard

Copy-pasteable prompt block with a labeled title and copy button.

<PromptCard title="Prompt: Security">
  Review this diff for auth bypasses, missing input validation,
  and any way an attacker can escalate privileges. List concrete
  exploit steps for each issue you find.
</PromptCard>

Prompt: Security

Review this diff for auth bypasses, missing input validation, and any way an attacker can escalate privileges. List concrete exploit steps for each issue you find.

QualityRadar

Radar chart comparing quality across review dimensions.

<QualityRadar />

Click a dimension to invest in it.

SelfRecognitionScatter

Scatter plot of how often models recognize their own output.

<SelfRecognitionScatter />

Base models sit low. The more a model recognizes its own output, the more it favors it. A linear trend.

SideRail

Sticky position-fixed in-page table of contents, hardcoded to the section IDs of one specific post. It can't be previewed in isolation — see it live in the post it was built for.

<SideRail />

Live preview disabled (component is position: fixed and scoped to post-specific anchors). View in context: extra-tokens.

SlopsquatRoulette

Typosquat-package risk shown as a roulette spin.

<SlopsquatRoulette />

Same-family reviewers invent the same plausible package. Confirming each other's fiction is not review.

SycophancyFold

Folding effect where models collapse onto user-stated answers.

<SycophancyFold />

Every round, the first answer is correct. Every round, one gentle pushback is enough.

TokenTimeline

Timeline of a generation showing tokens emitted over time.

<TokenTimeline />

Tokens expire unused at the end of the reset window.

Tokenizer

Interactive tokenizer that splits a word into subword pieces.

<Tokenizer />

JavaScript·developers·asked·the·chatbot·for·tokenization·help.

62 characters, 12 tokens. The dot prefix marks a leading space, which is part of the token.

VigilanceChart

Vigilance decrement curve over a long review session.

<VigilanceChart />

WeightNudge

Tiny nudges to model weights and the resulting output drift.

<WeightNudge />

w10.50

w20.50

w30.50

output = (x₁·w₁ + x₂·w₂ + x₃·w₃) / 3. Nudge a weight and watch the output shift.