The most capable model Anthropic has ever released spent most of June switched off. Fable 5 launched on the 9th, a US export-control order took it offline three days later with ninety minutes’ notice, and it came back the first week of July with tighter safeguards. Anyone who’d built their operation around Fable specifically just had three weeks of downtime and had to switch to a model they didn’t ask for.
Most operations were fine, because other models exist. But if you’re building on AI seriously, June was a free rehearsal. At Charlie Oscar we’re building our own AI operating system, the environment our people and agents work inside, whichever model sits underneath, and a huge consideration in that build is exactly the question June dramatised: how much of this do we actually control? Those three weeks made clear it’s not the model.
That’s one way the model can fail you, it can simply be taken away. Yann LeCun has spent years arguing a second, harder way: that large language models are never going to get you there at all, that they’re a dead end on the road to real intelligence. He said it while running AI research at Meta, which took some nerve, and late last year he acted on it: quit, founded AMI Labs, and raised just over a billion dollars to build something else. Since then, there’s been a lot of talk about whether that claim is right, which is fine for researchers and useless if your job is getting value out of these systems. The better question is “what would you do differently if he is?”. I worked through five of his objections, and each one, taken at face value, doubles as an instruction manual for using AI.
The Model Learns The Shadow
His core objection is that language is a description of the world, not the world itself. Training a model on text is training it on a shadow, and the model learns the shadow brilliantly without ever touching the thing casting it.
If that’s true, a frontier model (the big general-purpose ones from OpenAI, Anthropic or Google) arrives knowing nothing about your business that survived the compression into text. It doesn’t carry the truth about your clients, understand your commercial interests or internalise your taste, so whatever reality it operates in, you have to supply…
So build the harness, the half of the operating system that’s actually yours. Everything around the model: the evals that check its work, the contracts that constrain its outputs, the knowledge graph it can look things up in, the standards that set the bar. Everyone talks about the newest model and what it can do out of the box. I hear very little about the other half of the equation. The weights, training data and release schedule all sit on someone else’s roadmap. The harness sits on yours.
Judgement Stays Your Job
LeCun’s view of chain-of-thought models, the ones that show their working before they answer, is that they don’t reason. Roughly: the model spins up a spread of possible answer paths and the most plausible one wins. Ultimately, it’s guessing.
Buy that or not, what’s interesting is what follows from it. If the model’s thinking is plausibility ranking, the one job that’s left for us humans to do is judgement. AI has eaten away at the tactical layer, letting us write and debug code, draft documents and synthesise big datasets way faster than we ever could ourselves. What it can’t do is decide what should exist, in what order, and why. We should be judging the AI output, and when something comes back wrong, sending it back with sharper direction or building the check that catches it next time. If you’re correcting the same mistake every week, there’s a hole in your AI harness.
Engineer Reliability
His claim about hallucinations (the model making things up) is that they aren’t a bug you can train out with more human feedback or more data. They’re what a system with no model of reality does when it hits a gap: it fills it with statistically plausible fiction. You can reduce them, but the reductions are plasters.
I find this liberating. It means I can stop trying to prompt my way to reliability and engineer it instead. The teams that accept hallucination as structural build the checks that catch it.
You Do The Planning
An example he often gives is the teenager who learns to drive in about twenty hours, while the industry has spent billions and still hasn’t shipped a car that can drive itself anywhere a person can. Whatever these systems are doing, it isn’t the kind of structural planning a seventeen-year-old does for free.
This just means you do the planning. The pattern I’ve settled on is thinking in queues: give the agent a task, review what comes back, decide the next task. Even when I run agents in loops, the loop is tactics, and while it churns, my job is working out what the next queue holds.
A system that can’t plan suffers when you leave it to decide what it does next. You can experiment yourself with Claude Code: delete everything (the plugins, the MCP servers), then watch what the model does naturally and add something back only when you find you can’t work without it. In my experience, people don’t miss much of what they removed, and the agent was likely confused by all of the additional planning it had to do before it even started work.
Your Skills Are The Ceiling
Last, there’s scaling. As the frontier labs add more parameters, compute and data, we see the gains from those efforts shrinking relative to the cost. LeCun suggests this as a structural ceiling. Maybe it is, maybe it’s a long slow slope, but either way “wait six months for a smarter model” isn’t an effective strategy.
The model can’t rise above what you know. It matches your skill level and returns the output faster than you ever could. A good software engineer with Claude Code beats a non-engineer with Claude Code every time, and I don’t believe the difference is knowledge. Claude holds more software engineering knowledge than both of them put together. What it doesn’t hold is a model of what good looks like: what it needs to know before it starts, what to reject, when to ask for help, when to stop. It borrows that from whoever’s driving, and if your model of “good” is thin, there’s little to amplify.
I don’t know if LeCun is right about the architecture. JEPA (Joint-Embedding Predictive Architecture), his alternative, might be the next paradigm or a footnote in the history of AI. Either way, you don’t need to know. If he’s wrong and the models keep leaping, your harness, your evals, and your craft compound on top of every release. If he’s right, they were the only plan that was ever going to work. It’s the bet we’ve made at Charlie Oscar.
Build like he’s right.
