CORTEX Blog

JEPA is Yann LeCun's proposed path to machines that learn how the world works by watching it. We rebuilt one at small scale to see whether the idea survives without compute to hide behind. It does — the agent went from 12% to 100% goal-reaching with no rewards, no labels and no map. But four consecutive versions failed while every standard health metric said they were fine, and the fix that finally worked was published in 2018.

Invented adversarial sentences found 9 false-positive bugs last round. This round we pulled 8 real arXiv abstracts instead and found 3 more gaps — all misses this time — plus a subtle regex bug where a character class silently case-folded under IGNORECASE and briefly reintroduced an old false positive mid-fix. 60/60 tests now pass, and we have a real number: 0% fallback across 8 genuine academic abstracts.

Every calculator gives you a number and nothing else. Maith solves the problem and explains every step of the reasoning behind it — the explanation isn't a bolted-on extra, it's the product.

Drop a photo or take one live, and RecipAI reverse-engineers the dish into a full recipe — ingredients, quantities, and steps — flagging what it can see versus what it's inferring.

43/43 passing tests looked solid until we started adversarially probing for false positives instead of just confirming the happy path. Nine confirmed bugs came back, five of them the same failure mode: a bare word-root matching an unrelated word. All nine are fixed and now permanent regression tests — and the real lesson is that false matches, not misses, are the risk a rule-based Code Prompt actually has to guard against.

We showed three models agreeing on 9 of 10 extracted facts and called it a determinism win. A reader pointed out 7 of those 10 facts were still resolved by a model call, not a tested function — three LLMs agreeing isn't code. Here's the real fix: a unit test suite with zero model calls, the two actual bugs it caught, and what we got wrong about our own numbers.

Our first Code Prompt asked an LLM to score a claim holistically and diverged on all 10 criteria across three models. v1.1 rebuilds it as actual code: regex-first fact extraction, model fallback scoped to one yes/no question at a time, zero model calls in the scoring math itself. Same claim, same three models: disagreement collapsed to a single named fact.

A Code Prompt is supposed to be model-agnostic: same input, same structured output, regardless of which LLM runs it. We built one from the CRG-RIS evaluation standard and ran it against three current models on the identical claim. It diverged on all 10 of 10 scoring criteria. Here's the falsification, and the one ambiguous sentence that caused it.

Loop engineering says: stop prompting, start designing the loop. We think that's one layer short. The next shift is a standing assistant that hires and fires agents against goals, pulling real-time signals like traffic, sales, and returns — looking less like a prompt template and more like ClickUp or Calendly.

Promptest's rubric reliably separates bad, weak, good, and excellent prompts. But when four prompts spanning every tier were handed to real coding agents against the same bug, rubric score predicted almost nothing about cost or scope discipline — a CRG-RIS study reports the negative result as the headline.

Ask a model something it cannot know, and demand an answer anyway — is what comes back a hallucination or an act of imagination? The mechanism is identical either way. What changes is the verdict you render afterward. A sequel to The Illusion of Hallucination, with a live demo you can label yourself.

A "hallucination" isn't a mind losing its grip on reality — it's any deterministic algorithm executing flawlessly on a degraded input. We formalize the model-agnostic proof, define Acceptability Mapping, and build a working calculator you can operate yourself to see why $1+1=11$ is sometimes the correct answer.

The companies that win the next decade of AI won't be the ones with the most data — they'll be the ones with the most defensible data. AI is having its Napster moment; music solved this a century ago with ASCAP and BMI, and Ché is building that same collective for AI training data.

The most valuable training data left in the world isn't scraped text — it's lived experience that never made it online. Ché collects it from the ground up, anchors attribution on-chain, and pays contributors when their data is actually used to train a model.

We tried to make a coding assistant run on a 512MB budget on a decade-old Intel MacBook — fully offline. The model alone can't build a project. But wrap it in a loop that decomposes, verifies, and learns, and it builds working software overnight. The lesson: smarter loops beat bigger models.

Neuron Surgery reimagines the training of small language models by focusing on experience, judgment, and human-aligned acceptability rather than brute-force parameter expansion — sculpting models that don't just know, but understand.