Rafhael Oliveira · IA

The corrections from LeCun's talks: failure, passage, flag, measurement

Four failures from the queue, each answered by a passage from the two talks, turned into additive code and measured on the same budget.

1 min read· chapter: 58 min· Experiment in progress · chapter updated on 12 Sept 2026

Landed here from LinkedIn? This is chapter 13 of the experiment Antivision · Latent Rehearsal. It reads on its own; if you want the beginning, it is chapter 00.

A queue of 22 runs left four measured failures. We took each back to LeCun's text with a single question. What does the talk say that we have not done yet? The memory that persisted but did not govern. The level-2 encoder that collapsed three times, through three different doors. The world model that costs the generator. The guardrail that only acted on the real state. Three of the four had already received a trial-and-error fix, and each opened another door. The difference here is the order: first the passage, then the flag. Memory with dropout started governing: cost from 7.77% to 6.46%, and the rudder was back in charge. The house rule is to make it work, not to discard.

The three points

  1. The memory that did not govern: half the batches train without memory. Result: the rudder governs (0.301 against 0.224 for free), at lower cost.
  2. The level-2 encoder that collapsed: anti-collapse per dimension, and then an interpreter, because every measure of information is an upper bound.
  3. What changes in the theory: in text, with the action equal to the token, the world model is a predictor of transitions, not of tokens. What JEPA removes in video is already removed by the discrete action.

From the chapter, the beginning

The 22-round queue left four measured failures: the memory that did not govern, the level-2 encoder that collapsed three times, the world model that costs the generator, and the guardrail that only acted on the real state. This chapter goes back to LeCun's two talks, finds in each the passage that answers the failure, turns the passage into a flag and the flag into a round — and measures.

The house rule is to make things work, not to discard them. Each failure of the queue was taken back to LeCun's text with a single question: what does the talk say that we have not done yet? The answer became additive code (no earlier round changes its behaviour) and a new round on the same budget as the others (700 steps, seed 1337, control reg1 at 5.2706).

Three of the four failures had already received a trial correction (SigReg on level 2, the fixed norm, the cost on the real state) and each one opened another door. What is different in this chapter is the order: first the passage, then the flag. If the round fails, the failure is in the reading or in the scale — and both stay on record.

Keep reading chapter 13 →Ler em portuguêsThe PDFs by e-mail →
What the chapter covers · 12 sections · 58 min
  1. The method: failure → passage → flag → measurement
  2. The memory that did not govern
  3. The level-2 encoder: anticollapse by dimension
  4. The cost to the generator: the family that scaled
  5. The guardrail on the imagined state
  6. What changes in the project
  7. Researching further: where the unpredictable lives
  8. Two more doubts, with numbers
  9. The information of level 2, measured by an interpreter
  10. What the searches became in code
  11. Six passages, six algorithms — and what each one gave
  12. Three passages only a training run can answer
All chapters of the experiment · Antivision · Latent Rehearsal
  1. 00 · Antivision: an ELLM, a language model that rehearses in latent space before speaking
  2. 01 · The custom tokenizer: why 16 thousand tokens suit this corpus better than 50 thousand, and what the whole table shows
  3. 02 · Two architectures, one file: the fair duel between 2019 and what came after
  4. 03 · The geometry of the embedding: training starts with a perfect space and ruins it
  5. 04 · The lineage of collapse: five answers from computer vision, each measured on text
  6. 05 · The rudder and the world model: three versions of the same question, each measured against the last
  7. 06 · Method and mistakes: the rules of a legitimate comparison and the catalogue of our own errors
  8. 07 · Scale and budget: what it costs to reach GPT-3, and why VRAM is almost never the limit
  9. 08 · The workbench: serve, x-ray and train from the browser
  10. 09 · The hierarchical architecture: the whole project aligned to the four patterns of the world model
  11. 10 · What JEPA is, according to LeCun: two talks read concept by concept, and where each one applies
  12. 11 · Memory and horizon: two claims from LeCun that became experiments, and both hold for text
  13. 12 · Learned costs and inference modes: the three searches and the two sources of energy, measured
  14. 13 · The corrections from LeCun's talks: failure, passage, flag, measurement
  15. 14 · Consulted memory: a bank of abstract states read by attention
  16. 15 · What it costs and what it delivers: us, GPT-1 and GPT-2 on the only ruler valid across tokenizers
  17. 16 · The minimum was off the list: the day I measured my own search by energy
  18. 17 · ELLM: the right name for a language model that predicts in latent space and decides by energy
  19. 18 · Searching better does not fix the energy: three searches, the same text
  20. 19 · The inversion that did not repeat: the world model's cost appears once the ruler stops swinging
  21. 20 · What changed on 17/09: two new chapters and two corrections

Before this: the GPT-2 course

Rafhael Oliveira
Rafhael Oliveira

Writes the GPT-2 course and runs the Antivision experiment, measuring every run before publishing the number. The LinkedIn posts start here and end in the chapters.

Follow on LinkedIn →