Watch
Build intuition without leaving the page.
Start with a short explanation, manipulate the mechanism, then open the chapter when you need the evidence and full technical context. Every piece makes the next step explicit.
01 · Build intuition
Videos
Press play and stay in context. Video files load only when you choose to watch them.
Featured · Reasoning Models
Test-time compute: spend more inference before answering.
See what changes when a model allocates more inference to a response, and how that affects quality, cost and latency.
Open the full chapter →
The video, interaction and chapter share the same evidence and context.
More explanations
5 explanations availableArchitectures for combining text, image, audio and video
Identify the main architecture families and the trade-offs each one introduces.
Go deeper in the chapter →Scaling: data, compute and parameters
Connect the three variables behind the transition toward foundation models.
Go deeper in the chapter →Energy, heat and connectivity beyond Earth
Separate the promise from the physical limits that constrain orbital data centers.
Go deeper in the chapter →What makes generative AI different
Understand how models move from prediction toward generating text, images, audio or code.
Go deeper in the chapter →AI as an electrical technology
Connect adoption, productivity, energy and well-being without confusing impact with GDP.
Go deeper in the chapter →02 · Choose a question
Guided paths
You do not need to traverse the entire library. Choose a question and move from intuition to evidence in three steps.
Why can more compute improve an answer?
How do you distinguish a good process from a correct final answer?
How do text, image, audio and video share meaning?
03 · Test the mechanism
Interactive visuals
Change variables, compare mechanisms and return to the chapter when you need the full evidence.
01 · Reasoning
Best-of-N: generate more candidates and choose better
Best-of-N: generate multiple answers and choose the best
The simplest test-time compute lever: more independent attempts, a better expected result. Three variants use different selection criteria.
Same prompt, N independent runs with temperature > 0. The selection criterion determines the variant.
PROMPT
"Prove that the sum of a triangle's angles is 180°"
↙
↓
↘
Candidate 1
"Draw a line through vertex A parallel to side BC. Alternate interior angles give α = α', β = β'. The full straight line is 180°, so α + β + γ = 180°."
0.91 ★ Selected
Candidate 2
"Consider the sum of the exterior angles, which is always 360°. Since each exterior angle + interior angle = 180°, for the three vertices: 3×180° - (α+β+γ) = 360°, so α+β+γ = 180°."
0.78
Candidate 3
"By Euclid's postulate, in plane geometry the sum of the interior angles of any n-sided polygon is (n-2)×180°. For n=3, the result is 180°."
0.65
SELECTION CRITERION (PRM)
→ Candidate 1 selected for the highest average score across intermediate steps
The selection criterion defines the variant. Each has different advantages depending on the domain.
Best-of-N with verifier
N = 4–64
An ORM or PRM scores each candidate. The highest-scoring candidate is selected.
Ideal for: mathematics and code where an objective verifier is available.
Self-consistency (majority)
N = 8–40
N answers are generated and the one that appears most often is selected (majority vote over the result, not the reasoning chain).
Ideal for: questions with a discrete answer (multiple choice, calculations with a single result).
Model-as-judge
N = 3–8
A separate model (or the same model in a different role) evaluates the candidates and selects the best according to criteria stated in the judge prompt.
Ideal for: open-ended text, argumentation, analysis — where there is no objective verifier.
Performance improves with N sublogarithmically. The first candidates add a lot; later candidates add progressively less.
Performance (% correct)
Number of candidates (N)
Diminishing returns
Moving from N=1 to N=4 gives the largest jump. Moving from N=32 to N=64 adds much less at twice the cost.
PRM vs self-consistency
With a PRM the ceiling is higher because the verifier can identify the best answer even when it is not the most frequent.
Practical optimum
For most applications, N between 8 and 32 gives the best performance/cost ratio. Beyond that, cost grows linearly while the benefit plateaus.
02 PRM vs ORMSee what changes when you evaluate the process rather than only the outcome. Open interaction
PRMs vs ORMs: two ways to teach reasoning
Process Reward Models evaluate every step. Outcome Reward Models evaluate only the final result. The difference is deeper than it first appears.
A PRM assigns a reward to each intermediate step in the chain, not only to the end. This makes it possible to detect and penalize incorrect reasoning even when the final answer happens to be correct by luck.
Model reasoning chain
P1
"I need to factor the denominator: x²-4 = (x-2)(x+2)"
PRM
+0.95
Correct
P2
"The roots are x=2 and x=-2, which are singularities"
PRM
+0.91
Correct
P3
"I apply partial fractions: A/(x-2) + B/(x+2), solving A=1, B=1"
PRM
+0.31
Error detected at P3
P4
"I integrate each fraction: ln|x-2| + ln|x+2| + C"
PRM
+0.58
Correct form, wrong coefficients
Ans.
"ln|x-2| + ln|x+2| + C"
ORM
+0.70
ORM: "looks like an integral"
The PRM detected the error at P3 even though the final answer looks plausible. The ORM could miss it if the overall form looks correct.
The ORM observes only the final answer. If it is correct, the entire process is reinforced — including shortcuts and incorrect reasoning that reached the correct result by luck.
Case 1 — ORM works well
Correct reasoning
→
Correct answer
→
+1 ✓
Correct reasoning produces the correct answer and is reinforced. ORM works well.
Case 2 — ORM fails silently
Shortcut / error
→
Correct answer
→
+1 ✓
Incorrect reasoning reached the correct result "by luck". The ORM reinforces it anyway. The model learns the shortcut.
Case 3 — ORM penalizes unfairly
Correct reasoning
→
Final computation error
→
0 ✗
The reasoning was correct, but a final calculation error produces the wrong answer. The ORM penalizes the entire process, including the correct steps.
In all three cases, the ORM makes decisions with incomplete information. A binary correct/incorrect signal cannot distinguish processes that deserve reinforcement from those that do not.
The choice between a PRM and an ORM is not always clear. Each is better suited to different domains.
PRM — Process Reward Model
Ideal domain
Formal mathematics, logic, code with partial tests — where the correctness of each step can be verified
Main advantage
Detects shortcuts and incorrect reasoning that reach the correct result. Reinforces the process, not only the outcome
Best-of-N search
Makes it possible to select the best step at each point in the chain (beam search), not only the best final answer
ORM — Outcome Reward Model
Ideal domain
Any task with an objectively verifiable answer — basic mathematics, code, factual questions
Main advantage
Scales without requiring labels for intermediate steps. A cheap, scalable training signal
GRPO
DeepSeek R1 uses a binary ORM — correct or incorrect — and achieved frontier-level results without PRMs or human CoT data
03 Multimodal alignment with ImageBindExplore how six modalities share a representation through a visual anchor. Open interaction
Transitive alignment — from an image–text pair to six modalities
How ImageBind learns a shared space without requiring direct pairs between non-visual modalities.
TEXT
Text
IMAGE
Image
AUDIO
Audio
DEPTH
Depth
THERMAL
Thermal
IMU
IMU
VIDEO
Video
Millions of natural pairs exist for text + image: captions, product descriptions, articles with images. For audio–text: they are scarce and expensive to label. For depth–layout: they require specialized sensors and human work. Modalities outside the text–image pair remain disconnected in the representation space.
IMAGE
anchor
anchor
TEXT
Text
AUDIO
Audio
DEPTH
Depth
THERMAL
Thermal
IMU
IMU
VIDEO
Video
Transitive alignment (dashed line): ImageBind (2023) learns the space using only the five pair types that include image. The model never sees a direct audio↔text pair, but because both are aligned with image, they become aligned with each other.
A text query can retrieve audio. An image can retrieve depth or thermal data. All modalities become connected without having trained those direct relationships explicitly.
CLIP / Image–text pair
Binary alignment
Image↔text retrieval only
N modalities → N(N−1)/2 pairs to connect them all
2 connected modalities
TEXT
IMAGE
pairs required for N=6
N(N−1)/2 = 15
ImageBind / Gemini Embedding 2
Native multimodal alignment
Cross-modal retrieval between any pair of modalities
N modalities → N pairs (all through the anchor)
6+ connected modalities
Gemini Embedding 2 (2026): text, image, video, audio, and documents as a native primitive
TEXT
IMAGE
AUDIO
DEPTH
VIDEO
IMU
pairs required for N=6
N = 6
Note: Asymmetry in data availability explains asymmetry in capabilities. Models understand images better than audio, and audio better than documents with complex layout, because high-quality training pairs are far more abundant for images than for the other modalities.
Next step
From visual intuition to a complete mental model.
Once the mechanism is clear, continue with the learning series to review architecture, evidence, references and limitations.
Explore all series →