A web giant put to the test: CLIP does spot the forest — but it confuses that with brightness too

Beginner track

Where we left off

Chapter 11 delivered the single most important diagnosis of this long investigation into the

cold start: the project's compass (the score that compares each story imagined by the world

model against a memory of "a tree being chopped" — see Chapter 4) does not simply switch off

once it leaves its training ground. It points backwards. A tree right in front of the agent

gets a lower score than an empty meadow. Chapter 12 dug further and showed that a good part of

this problem very probably comes from a hidden shortcut inside the project's very first vision

model (Chapters 1-2): instead of genuinely understanding "there is a tree here", the model seems

mostly to be spotting whether the image is bright or dark — a shortcut confirmed three times in

a row, in three different forms.

Before committing to the heaviest option on the menu left by Chapter 11 (building a second

brain — a separate, slower model whose only job is to "find a forest"), someone from outside the

project, consulted by the project owner, proposed a simpler and far cheaper question to test

first: does a very large, off-the-shelf AI model, never trained on Minecraft at all, already

spot a forest correctly? If so, there would be nothing new to build — it could simply be reused

as is. This chapter tells the story of that test, and what it revealed: a result that is half

encouraging, half worrying, and honestly not yet fully understood.

The idea: borrow a giant's eyes instead of building new ones

The proposed model is called CLIP. It has never seen anything from this project or from

Minecraft: it was trained by another team on hundreds of millions of photos and captions taken

from the internet — the kind of thing you find when searching for images on the web. What CLIP

can do, once trained: compare any image to any sentence and say how well they "go together",

without needing to be retrained for each new task (this is called a "zero-shot" comparison —

literally "without a single extra training attempt").

The test was simple: take exactly the same real game frames that were used to discover the

backwards-compass problem in Chapter 11, and ask CLIP to compare each one against two sentences:

"a dense forest with many trees" versus "an open grassy field with no trees". Two conditions had

to be met for the test to count as a genuine success:

1. CLIP must agree with common sense: an image with a nearby tree must score more favourably

towards "forest" than an image with no tree. (This is exactly the opposite of what the

project's broken compass was doing in Chapter 11.)

1. That score must not actually be just a roundabout way of measuring the brightness of the

scene — otherwise it would be the same old shortcut already spotted three times

(Chapter 11), simply hidden behind a much bigger and much more impressive model.

The result: half the test clearly passes, the other half fails badly

The first condition passes, and clearly: CLIP does give a better "forest" score to images

that really do show nearby trees than to images that show none. On that specific point, CLIP does

better than the project's broken compass — it does not invert.

The second condition fails, and badly. CLIP's score is very strongly tied to the overall

brightness of the image — almost as if the two were measuring nearly the same thing. And this is

not only true on the difficult frames from the crafting environment (where the project's compass

inverts): it is true even on the easy frames from the original tree-chopping environment, the

one where the project's home-made compass has been working correctly since Chapter 6.

What this means, honestly — and what remains a genuine mystery

Care is needed here, because this is an open question, not an established certainty.

This result weakens, without completely demolishing, an idea that had seemed reasonable up to

now: that the project's small model confuses brightness with forest simply because it did not see

enough varied scenes (day, night, different weather) during training. If a giant model, trained on

hundreds of millions of photos far more varied than anything this project could ever collect,

makes the same confusion — and even slightly worse — then a lack of diversity may not explain

everything.

There is even a possible explanation that would not be a "bug" at all: a genuinely dense forest,

in real life as in a game imitating it, really is darker than an open field in full sunlight,

because of the shade cast by the canopy. Picking that up could be a useful and sensible signal,

not a reasoning error.

But one precise observation stops us settling for that simple explanation: among the frames

tested, a dark cave scene with no tree whatsoever was placed by CLIP in the same group as the

bright, open scenes — "no forest" — instead of being confused with a dark forest, as a

"darkness = forest" story would have predicted. So something other than plain brightness is going

on, but exactly what remains, to this day, a genuinely unanswered question.

Phase 2: retraining the world model directly — a mixed result, not a clean one

The decision taken after the CLIP test was not to demand that any future fix become completely

independent of brightness (at the risk of stripping out a possibly legitimate signal), but to

check directly whether the real original problem — the compass pointing backwards on real random

starting points — disappears, no matter which mechanism achieves it. That work is now finished,

and here is what it produced.

The project took the existing world model (not from scratch — a simple adjustment of already

trained weights) and retrained it on a mixture: the original guaranteed-forest frames, plus real

free-spawn demonstrations, plus random exploration episodes. What was new compared with every

previous attempt: this time a photometric augmentation (the colours, brightness and contrast

of each short sequence are altered at random, but consistently across the whole sequence rather

than frame by frame, to force the model to stop leaning blindly on brightness) was applied

directly to the training of the world model itself — not to a small module bolted on top, as in

previous attempts. This is the first time in the whole investigation that this particular move has

been made directly on the main model.

The usual safety checks (no representation collapse — see Chapter 2 — and prediction quality

barely moving, which was intended: the goal was a light adjustment, not a full retraining) passed

without a hitch across all 5 training rounds.

The real test was to re-run, on each of the 5 adjusted versions, exactly the same diagnostic

that had uncovered the backwards-compass problem. The result is genuinely split, not merely

"it almost worked":

• Leaving out one specific frame — a dark scene, like a cave or an underwater passage, with no

tree at all — the fix looks clean: across the 5 adjusted versions, scenes with a nearby tree now

score several times higher than open scenes with no tree. That is a real reversal of the

original inverted direction.

• But putting that single frame back into the calculation flips the result the wrong way again

on all 5 versions, because that specific frame's score got worse after the adjustment, not

better. This is a new anomaly, tied to low brightness, appearing somewhere different from before

— but clearly belonging to the same family of problem as the brightness shortcut already spotted

earlier (Chapter 12, and the first half of this chapter with CLIP).

None of the 5 adjusted versions was promoted to replace the reference model, and no new real-game

test was run: the project's rule is to spend a live test only on a candidate that has clearly

passed its offline test first — and this one did not clearly pass.

Why this matters, stated precisely: this is the fourth time, with four genuinely different

approaches, that a brightness-linked anomaly has appeared — a small bolted-on module trained on

the model's already-frozen features; that same module retrained specifically on free-spawn data;

a giant off-the-shelf model never trained for this project (CLIP, earlier in this chapter); and

now a direct retraining of the world model itself, with real photometric augmentation — the most

direct attack yet on this precise problem. That seriously weakens the idea that "the model simply

hasn't seen enough different lighting yet" explains everything — even the most direct and most

carefully done attempt to fix exactly that still produced a new anomaly, not a clean fix.

What happens next

The condition the project had set itself for committing to the heaviest option on the menu left in

Chapter 11 — a second, slower world model dedicated to "finding a forest" before handing back to

the existing fast model to chop the tree, called H-JEPA — was: "only if the direct, cheaper fix

fails". That has now happened, with real evidence behind it rather than a merely plausible

argument. The project therefore turns towards this structurally different option as its next

direction, rather than continuing to hunt for a direct fix on the existing model.