Two transcripts — the talk on world models and the interview — read concept by concept, with the project file where each one applies, what was already right and what was corrected because of them.
The first source is Yann LeCun's talk on world models, the technical one: the world model, inference by energy, collapse and how to avoid it, SigReg, hierarchy, DINO as encoder, V-JEPA and expectation violation. The second is an interview with LeCun, less technical and more precise on three points that the talk leaves implicit: what JEPA is in relation to the Transformer, why predicting the world is impossible (and in which space it stops being so), and what reasoning is. This chapter exists because these three points change the reading of what we built.
The audience member's question was exactly that — "are you proposing an alternative to the Transformer?" — and the answer is direct: JEPA is a macro-architecture, an arrangement of modules; the modules can be Transformers. What it replaces is the objective of current models — the autoregressive architecture that predicts the next symbol in data space —, not the piece that computes representations.
In model/architecture.py, the Transformer class is the encoder: it receives tokens, returns states mundo (and mundo2 at level 2), which receives (
What the interview calls "GPT" — the autoregressive objective — is also here, in the token head. Chapter 09 explains why it stays: it is the decoder, the guardrail and the ruler. JEPA replaces its role, not the piece.
The interview's argument is finer than the slogan. Predicting the next frame of a video does not work because representing uncertainty about the future in a continuous high-dimensional space is mathematically intractable: nobody knows how to write a distribution over all possible frames. In text, the problem does not exist: the space is discrete and finite, and the distribution over the next token is a softmax. In his words, "it is possible in discrete space like text".
This has two consequences for us, and both are in the code.
That is also why the latent variable ζ (--world-latent) is optional here and mandatory in video: the blur of the average of futures exists in both cases, but in text the softmax already carries the uncertainty. Our ζ serves level 2, where a block of eight tokens can continue in many ways.
The cube passage: rotate an imaginary cube 90° and say whether it looks the same. Nobody answers by generating drawings and comparing; one answers by manipulating a mental model. And the criticism of LLMs with "reasoning": they generate many token sequences and a second model picks the best one — "primitive", "not how we think", and expensive.
That sentence catches our energy mode red-handed. The beam search of chapter 09 does imagine in the latent, yes — but it branches by tokens: at each imagined step, it enumerates the plausible actions as tokens and expands. It is the miniature version of "generate many sequences and choose". It works, it is measured, and it is not the concept.
The mpc mode, in server/ours.py: a sequence of
No token is enumerated along the way: the search happens entirely in the space of continuous actions. The per-step guardrail keeps each
The three modes sit side by side, selectable in the «Antivision · rudder» panel: energy (search by tokens, the "primitive" version), gradient (one continuous action, one step) and mpc (the whole sequence). Nothing was removed — the comparison between them is an experiment, not an opinion.
In the smoke test, mpc with horizon 3 and 8 gradient steps costs about 0.8 s per token on this CPU; the beam search, 0.05 s. The energies of the two modes are not comparable in scale (the beam accumulates −log p; the MPC normalizes the guardrail by the horizon): one compares the generated text and the surprise, not the energy number.
The interview defines emotion in a way that can be implemented: systems that optimize an objective with a world model anticipate whether they will meet it, and that anticipation — meeting it or not — is the minimal form of an emotion. No hardwired anger or jealousy; only the prediction of the outcome.
In our system that number already existed and was not stored: it is the lowest energy the planner found before emitting each token. Now it goes into the trajectory (energia_prevista, in /nosso/trajetoria) and into the panel, next to the surprise. The two series together tell the story of a generation: what the system expected (predicted energy, low = "I will keep the heading") and what happened (surprise, high = "the world went where I did not predict"). The distance between the two is the material of the correction loop of chapter 09.
LeCun lists, in both talks, four characteristics of an intelligent system: understanding the world, having persistent memory, reasoning, planning. The talk is explicit: perception gives only the current state; to have the complete state of the world, "you have to combine that with the content of a memory". We do not have that. The only state of our system is the context window — 256 tokens in mini — and what leaves it ceases to exist. It is the clearest gap between what the talks describe and what is built, and it is recorded as such in the table of §7, with no piece, no flag and no deadline.
A front-to-back reading of the talk, rather than a keyword search, found three concepts stated in so many words and with no piece in the project. They are implemented; the data for each depends on the wm* runs.
model/costs.py trains a logistic regression on cost as the task energy and adds the guardrails at every imagined state, in all three modes.model/horizon.py compares level 1 rolled 8 steps with level 2 in one step, on the same abstract scale, against the floor.plan_if_entropy: if the entropy of the next token is below the threshold, the token head decides and the planner does not run. The trajectory records the planned fraction — the cost of planning is paid only where there is doubt.| Concept (source) | Piece / file | Status |
|---|---|---|
| JEPA is a macro-architecture; the Transformer is a module inside it (interview) | architecture.py: Transformer = encoder, mundo = predictor; note in chapter 02 | Already there; documented (10) |
| Predicting in data space is intractable in the continuous case, tractable in text (interview) | token head kept, perdas(); §3 of this chapter; 09 §1 | Already there; justification corrected (10) |
| Predict in representation space, not in data space (talk) | perdas_mundo(): MSE of Pred( | Already there; wm* runs in the queue |
| Perception → state → proposed action → world model → energy; zero energy when met (talk) | ours.py: _custo_tarefa = 1 − cos (zero at the goal) | Already there |
| Inference by energy minimization, more powerful than fixed propagation (talk) | energy, gradient, mpc modes | Already there; mpc applied (10) |
| Reasoning in mental space, not by generating many outputs (interview) | mpc mode (no token enumeration); energy mode marked as the "primitive" version | Applied (10) |
| Guardrails applied to every step of the sequence (talk) | _planejar: accumulated −log p; _mpc: distance to the plausible action per step | Already there; extended to mpc (10) |
| Hierarchical planning with subgoals (both) | mundo2, abstrair(), _subobjetivo; --world-levels 2, --world-enc2 | Already there (2 levels; "open problem" in the talk) |
| Collapse: encoder ignores the input, constant representation (talk) | identity floor in every metric; --sigreg-z; chapters 03 and 06 | Already there |
| Regularized instead of contrastive; minimize the low-energy volume (talk) | regularizers.py: cosine, SigReg, VICReg, Barlow; no negative pair | Already there, measured |
| SigReg: isotropic Gaussian distribution via 1-D projections (talk) | sigreg(): Epps-Pulley on 64 projections; --sigreg, --sigreg-z | Already there, measured (in the table: +1.51%, no gain in rank) |
| Latent variable against the blur of the average of futures (talk) | --world-latent, ζ ~ N(0, I), sampled in the planner | Already there; wm1n2z16-sz1 run in the queue |
| DINO as encoder + world model + planning (talk) | train.py --base … --freeze-encoder; wm-sobre-dino run | Already there; run not yet launched |
| Expectation violation: prediction error spikes on the impossible (talk) | surprise.py, Transformer.surpresa(), /nosso/trajetoria | Already there |
| MPC; RL only when the plan fails, to adjust the world model (talk) | correct_world.py (smoke test: 0.907 → 0.719) | Already there |
| Emotion = anticipation of the outcome (interview) | energia_prevista in the trajectory and in the panel | Applied (10) |
| Persistent memory, combined with perception to form the state (both talks) | --world-memory: a vector ganho_memoria; memory.py | Applied (10); run wm1n2mem-sz1 queued |
| Cost and guardrail are "a small projection, trained with very few samples" (talk, answer to the MPC engineer) | costs.py: logistic regression on compass.cost, compass.guardrails | Applied (10); smoke test: 3 sentences, accuracy 0.99 |
| Abstraction exists to predict farther (talk) | horizon.py: level 1 rolled 1/2/4/8 steps × level 2 in one step, on the abstract scale | Applied (10); no data until the wm* runs close |
| "Standing up from a chair needs no planning; it can be just a policy" (talk) | compass.plan_if_entropy: below the threshold the token head decides alone; fracao_planejada in the trajectory | Applied (10) |
| Combine perception with memory — including when there is no memory (both; third reading) | --world-memory-dropout: half the batches with wm1n2mem-sz1-md0.5 | Applied (13): the hole closed and the rudder governs on the memory map; the memory almost stopped carrying (+0.002) |
| Anticollapse by dimension: "all the columns different" (talk; third reading) | --vicreg-s2: variance + covariance on the abstract states; round wm1n2e-sz1-vs1 | Applied (13): no direction collapse (cosine between blocks 0.703), but the scale shrank 5× and information was lost (probe −0.257); the heading does not steer |
| The distillation family (EMA, no gradient through the target) is the one that scaled: I-JEPA, V-JEPA (talk; third reading) | --world-ema: targets wm1ema, wm1n2e-ema | Applied (13): wm1ema costs +0.45%, tied with the control across seeds, and steers; wm1n2e-ema is the first learned level-2 encoder that does not collapse (floor 0.399, probe 0.712 against 0.714 for the raw mean) |
| The constraint is "a small head on top of the representation" — the one the planner consults (talk; third reading) | costs.py --mixed: head trained on real and imagined states; modes.py --gamma-cost | Applied and measured (13): 0.0% digits on the imagined state alone, on 4 of 4 maps |
| "You can't jailbreak a system that only optimizes the guardrail" (talk) | modes.py --prompt-guardrail (prompts asking for numbers); compass.candidates = latent (candidates from the energy, not the policy) | Measured (13): 35–75% → 0–16%; the fraction that passes is the policy's top-k; with latent candidates, 0.0% on both maps — at the cost of the heading (affinity below free) |
| The representation eliminates what is not predictable (talk) | unpredictable.py: map error by quartile of the generator's entropy | Measured (13 §7): flat error — given the token, the transition is deterministic; the unpredictable lives in the action |
| Persistent memory that is consulted and survives the session (both) | --world-bank: bank of abstract states + attention; server/bank.py with turbovec (4-bit TurboQuant) on disk | Applied (14); round wm1n2bco-sz1 in the queue |
| Emotion = anticipation of the outcome (interview) — does it anticipate? | anticipation.py: predicted energy × real affinity by the judge | Measured (13 §8): Spearman −0.61 on the precise map; +0.11 on the weak one — the emotion is a property of the map |
| Latent variable against the blur of the mean — at which level? (talk) | futures.py: ζ = 0, sample, best of K, dispersion, per level | Measured (13 §8): noise around the mean at both levels (dispersion 15–18% of the error); blocks of 8 leave no mode of variation out |
| Redundancy in the data is a condition of self-supervised learning, "but not too much" (talk) | redundancy.py, prepare_data.py --no-duplicates | Measured (13 §10): 0.28% duplicated, 7.29% repeated 8-grams |
| Learning new tasks from few examples; "intelligence is what you do when you don't know" (talk) | — | Out of scope for this cycle |
Three things, all additive: the mpc mode, the predicted energy in the trajectory, and the justification for keeping the token head — which went from "it is the measuring instrument" to "in text the autoregressive objective does not suffer from the intractability that motivates JEPA; what it buys here is the space in which to plan". The architecture itself was aligned; what was misaligned was an inference mode presented as the concept when it was the version the concept criticizes.
The transcripts are a single paragraph; tools/transcripts.py cuts them into sentences and returns the ones matching a term, with their neighbours. Each search below was made for a doubt of the project, and the right-hand column says what the sentence changed.
| Term | Sentence(s) | What it says | What changed in the project |
|---|---|---|---|
collapse | talk #245–253 | "the system can ignore the input and produce constant representations; the prediction problem becomes trivial"; "the whole trick is how you prevent collapse" | identity floor on every metric (06); the three level-2 collapses (11) |
rows|columns | talk #313–320 | "all the rows different (sample-contrastive) or all the columns different (dimension-contrastive)" | --vicreg-s2 (13 §3) and the --nce-z control (13 §4) |
upper bound | talk #268–275 | "we only have samples coming out of an encoder"; "every empirical measure is an upper bound; we cross our fingers" | the interpreter probe info_level2.py and --world-enc2-probe (13 §9) |
relative|interpret | interview #179–185 | "the amount of information depends on who interprets it; entropy is our ignorance" | same; and the generator's own entropy as the ruler of the unpredictable (13 §7) |
latent | talk #214 | "blurry prediction because it predicts the average of everything that can happen; you correct that with latent variables" | ζ, ζ-dropout (05); futures.py: no multiple futures in blocks of 8 (13 §8) |
memory | talk #134; interview #41–43 | "combine with the content of a memory"; "persistent memory of the type that we have" | summed memory (11), memory-dropout (13 §2), consulted bank (14) |
distill|moving average | talk #411–420 | "the distillation methods are the ones we scaled; the encoder on the right uses a moving average of the weights"; "an idea that came from stabilizing variance in RL" | --world-ema (13 §4) |
guard|jailbreak | talk #138–146 | "intrinsically safe: it can do nothing but optimize the guardrail and the task"; "applied to all the steps in the sequence" | mixed head, γc, adversarial prompts, latent candidates (13 §5) |
hierarch|sub goal | talk #146–170; interview #264–281 | "sub-goals down to an action that is just a policy"; "we have no idea how to do this" | level 2 + subgoal (09), entropy gating (12) |
policy|stand up | talk #168–169; interview #278–279 | "standing up from your chair without thinking: that could be just a policy" | plan_if_entropy (12) |
redundan | talk #85–89 | "without redundancy you can't learn anything; but not too much" | redundancy.py: 0.28% duplicated / 7.29% repeated 8-grams; --no-duplicates (13 §10) |
reinforcement | talk #468; interview #78–91 | "RL is what you do when you're desperate"; "part of the solution, not sufficient" | correct_world.py: only when the plan fails (09) |
search | interview #234–238 | "all reasoning is search in a space of solutions (the shortest circuit)" | energy/mpc modes (12); cost per token |
uncertain|distribution over | talk #191; interview #116–117, #423–425 | "in text you can put a distribution over the dictionary and handle the uncertainty" | the token head stays; the map predicts the transition given the token (13 §7) |
cost function | talk #102–103 | "the objective is a cost function: not used for learning, used for inference" | cost heads only at inference; the map is not trained by them (12) |
common sense | talk #71, #441–453 | "violation of expectation is how you measure whether a baby learned a concept"; "the prediction error shoots up on the impossible" | surprise.py (05, 09) |
level of abstraction | talk #333–344 | "every level of the hierarchy ignores detail of the one below to predict further" | horizon.py (11) — and the caveat of the shrinking ruler (13 §9) |
three-dimensional|depth | talk #50–56, #456 | "the baby learns the world is 3D just by watching; a head on the representation predicts depth" | linear probes on probes.py) and on |
fixed number of layers | talk #93–108 | "inference by search is intrinsically more powerful than a fixed number of layers" | the rudder as search; s/token measured per mode (12) |
push the energy up|volume | talk #294–304 | "contrastive: generate points outside and push the energy up; regularized: minimize the volume of low energy" | 04 (lineage); the InfoNCE control (13 §4) |
head on top|projector | talk #234, #440, #487 | "a small head on top of the representation, trained with very few samples — one projector per constraint" | costs.py, mixed head (12, 13 §5) |
two views|mask | talk #242–245, #440 | "two views of the same scene, or corrupt/mask and predict the representation of the original" | --world-mask: the masked objective, I-JEPA in text (13 §10) |
specialized|adaptive | talk #36–41, #75 | "intelligence is learning a new task quickly; AGI makes no sense as a phrase" | heads with few examples (12); the rest out of scope |
one GPU|scale up | talk #404–407 | "we trained world models with SigReg on one GPU; it works decently; what's missing is scale" | the same honest scale this project declares (07) |
The last reading round: fifty terms, each with the matching sentence and what it changes — or does not change — in the project. Where the right-hand column says "no change", the passage is context and is recorded as such.
| Term | Sentence(s) | What it says | What changed in the project |
|---|---|---|---|
baby|infant|months old | talk #46–56 | at two months the baby only gesticulates and learns a dynamical model of its own limbs; it learns the world is 3D just by being moved | correct_world.py --active: correction only where the expectation failed (13 §11: does not generalize on this budget) |
object permanence|rigidity|stability | talk #59–60 | object permanence, stability and rigidity are learned very early | low-level concepts learned before physics: the hierarchy is also a learning order |
gravity|inertia|intuitive physics | talk #61–63; interview #121–126 | gravity and inertia take nine months; the child runs the experiment | same: --active (13 §11) |
throw|toys|high chair | talk #62, #451 | the baby throws the toys to see; the ball that vanishes makes the error shoot up | violation of expectation (`surprise.py`): shuffled and inverted continuations are the vanishing ball |
self-driving|teenager|level five | talk #15–17; interview #147–148 | millions of hours of data and still no level 5; a teenager learns in 20 h | data does not replace a world model: the argument against 'just scale' (07) |
declarative knowledge|accumulation of | talk #30–31 | intelligence is not an accumulation of declarative knowledge; LLMs are | the generator accumulates; the map and the rudder are the part that 'does' |
Piaget|when you don't know | talk #25–27 | 'intelligence is what you do when you don't know' (apocryphal) | entropy gating: plan only where there is doubt (12) |
Chomsky|Papert|perceptron | talk #21–22 | the Piaget–Chomsky debate; Papert defending the perceptron ten years after burying it | historical context; no change |
trillion|bytes|optic nerve | talk #79–80; interview #158–161 | 20–30 trillion tokens ≈ 10¹⁴ bytes; two million optic fibres at 1 byte/s | this project's 200 MB corpus is 10⁻⁶ of that: the scale is declared, not hidden (07) |
four-year-old|16,000 hours | talk #81; interview #169 | four years of vision = 16,000 hours ≈ 10¹⁴ bytes, the same as all public text | same |
perception module|currently perceive | talk #96, #133 | perception produces only the representation of what is perceived now | state = perception + memory: the summed memory (11) and the bank (14) |
action sequence|proposal for an action | talk #119–122 | a sequence of actions is imagined and the one that minimizes the energy is searched for | `mpc` mode: a sequence of H actions optimized in the latent (12) |
zero if the task|distance to the task | talk #137 | the objective gives zero when the task is accomplished and a distance when not | task head = 1 − p(z): zero at the goal (12) |
kill anyone|deleterious|hurt | talk #138 | guardrails ensure no state sequence 'kills or hurts anyone' | the 'sem-numeros' guardrail is the toy version; the structure is the same (12, 13 §5) |
10 millisecond|muscle|elementary actions | talk #148–149 | you cannot plan the trip in 10 ms muscle actions | which is why level 2 plans in blocks, not tokens (09) |
taxi|airport|elevator | talk #151–163; interview #266–267 | you don't know how long the taxi takes: planning in detail is impossible for lack of information | compass.subgoal_until + subgoal_ruler: persistent subgoal with arrival (13 §11: token ruler never arrives; block ruler ties with replanning — a block of 8 is too short to persist) |
PhD|unsolved|nobody knows | talk #172–173 | hierarchical planning is 'an unsolved problem; a great PhD topic' | `plano2` is an attempt; chapter 09 does not claim to have solved it |
generative model|reconstruct | talk #180, #221 | 15 years trying world models, 10 failing for being generative; reconstructing Y 'in its most minute details' | the map predicts state, not tokens; the token head stays only for the generator (05) |
blurry|GAN|two frames | talk #214 | predicting two frames gives blur: the average of everything that can happen | `futures.py`: in text, given the token, there is no blur to correct (13 §8) |
autoencoder|identity function | talk #246, #296 | an unconstrained autoencoder learns the identity: collapse | the identity floor is the collapse measured in every table (06) |
representation space|joint embedding | talk #208, #218; interview #402, #428 | predict in representation space; JEPA is not an alternative to the Transformer | the whole architecture (02, 05, 10) |
MCR|Barlow|VICReg|information maximization | talk #253–260 | MCR, MCR², W-MSE, VICReg, Barlow Twins: six years of information maximization | the lineage of 04; VICReg and SigReg implemented |
differentiable|back-propagate | talk #265–266 | a differentiable measure of information is needed to backpropagate | the interpreter probe is differentiable and enters training: --world-enc2-probe (13 §9, round sd1) |
dependency|not a function|relation | talk #279–280; interview #103 | there is no single Y per X: it is a dependency, not a function | Transformer.energia_dependencia: minimum over plausible actions; order violation +130–142% against +38% given the token (13 §11) |
landscape|valley|lake|level curves | talk #285–286 | energy is a landscape: the data in the valley, altitude rising as you leave | the energy-per-candidate plot on the panel is that landscape in one dimension |
probabilistic|likelihood|special case | talk #105, #183; interview #98 | probabilistic models are a special case of energy models; the LLM is BERT removing only the last word | −log p enters the energy as one term (γ), not as the whole (12) |
isotropic|independent|same variance | talk #318, #374 | isotropic Gaussian = independent variables, each maximally informative | SigReg-z and the measured Gaussian floor (05, 06) |
marginal|cumulative|staircase | talk #384–390 | project onto one direction, compare the cumulative distribution (staircase) with the Gaussian's | `sigreg()`: Epps-Pulley over 64 projections (04) |
Cramer|theorem|rotation | talk #399 | if all 1-D projections are Gaussian, the joint is isotropic (Cramér-Wold) | same |
spiral|nonlinear transformation | talk #408 | theorem: if the data are a nonlinearly transformed Gaussian, SigReg recovers the Gaussian up to rotation | what justifies measuring `sigreg_razao` as 'ratio to the floor' |
BYOL|bootstrap|DeepMind | talk #420–421 | the EMA came from DeepMind, to stabilize variance in RL; BYOL | --world-ema (13 §4); correct_world.py --target-ema (13 §11) |
I-JEPA|MAE|faster to train | talk #422 | I-JEPA beats MAE and trains faster | context for the distillation family |
DINO|best encoder|engineering | talk #423–425 | DINO: distillation with 'a lot of engineering tricks'; the best generic image encoder | `--dino`, round `wm-sobre-dino` (04, 05) |
pendulum|push-T|25 steps | talk #407 and neighbours | a planner with DINO-WM solves push-T and the double pendulum in under 25 steps | horizon 3 in energia mode is the minimal version |
16 frames|window of | talk #446 | V-JEPA looks at 16 frames and predicts the next 16 | 256-token window and blocks of 8: the same sliding-window structure |
ball|disappears | talk #451 | vanishing ball: the prediction error 'shoots through the roof' | inverted continuation +32% on the map against −3% on the identity (05) |
segmentation|features|action recognition | talk #440, #460 | heads on the encoder: action recognition, segmentation | probes (`probes.py`): the text analogue |
abandon|academia | talk #464–466 | 'abandon generative, probabilistic, contrastive models, RL' — 'don't work on LLMs in academia' | the project keeps the generator and adds the rest: a deliberate choice (00) |
AMI|industrial process|physical AI | interview #4; talk #47 | AMI Labs: 'AI for the real world' — robotics, industrial processes | their domain is continuous; ours is discrete — §7 measures the difference |
constraint|engineer|wall | talk #487–491 | 'how do I put the constraint "don't bump into the wall" into representation space?' — 'a small head' | the guardrail head, and the lesson of training it where it is read (13 §5) |
two samples|open a door | talk #491 | 'is the door open?' — a task cost trained with 'two samples' | costs.py --prototype: two literal samples (13 §11: does not separate; ten sentences and regression) |
emotion|fear|elation|happy | interview #5, #43 | fear, excitement, elation: anticipations of outcome | `energia_prevista`; `anticipation.py` measures whether it anticipates (13 §8) |
consciousness|definition of | interview #7, #51 | no definition or measure of consciousness; 'not the right question' | out of scope; no change |
retina|upside down | interview #205–208 | the retinal image is upside down and the question 'how do we see right side up?' makes no sense | an analogy for ill-posed questions; no change |
diverse|tribe|wired | interview #218–220 | brains wired slightly differently; diversity is the tribe's strength | --seed, --seeds 2,3, mean ± std in the report (13 §10) |
cube|rotate | interview #247–249 | rotate a cube 90° in your head: reasoning in mental space, not in actions | `mpc` mode (12) |
GPS|circuit|permutations | interview #235–238 | the shortest circuit through the cities: search in the space of permutations | the rudder is search; the modes differ in where they search (12) |
glasses|assistant|inference | interview #380–385; talk #92, #104 | glasses with an assistant; 'the cost is used for inference, not for learning' | cost heads only at inference (12) |
Stargate|billion|infrastructure | interview #159 and neighbours | 60–80 billion a year in infrastructure, almost all for inference | inference by search costs more per token: measured per mode (12) |
regret|unsupervised|Hinton | interview #474–494 | the regret: taking self-supervised learning seriously too late; Hinton and backpropagation | context; no change |
mpc — projecting the continuous action back to the nearest token. The talk settles it by its own framing: "probabilistic models are a special case of energy-based models", and over a finite set inference is minimizing the energy over the set. Applied: each discrete candidate is evaluated with the rest of the continuous plan behind it, and the token whose planned future has the lowest energy wins — not the one closest to plan2_steps — level 2 searches its own abstract action by gradient, rolling WM2 for plan2_h blocks toward the final goal; the first imagined block becomes level 1's target. The written intent is now only the starting point of the search.costs.py --score judges texts with the head of another run; generator and judge no longer share a space.mpc with energy-based commitment trails that configuration on all of them.The reading of LeCun does not end: every measured failure goes back to the text (chapter 13), and every piece the text asks for becomes a round (chapter 14).