The loudest voices in AI keep insisting the next frontier is bigger models. Yann LeCun just raised a billion dollars to prove them wrong.

The pattern here isn't competition. It's colonization: every layer of daily life, quietly claimed.

Yann LeCun spent years at Meta telling anyone who'd listen that large language models are a dead end for reaching human-level intelligence. Now he's put $1 billion behind that opinion. His new startup, AMI, is built on a single premise: AI that can't reason about the physical world isn't really intelligent. It's autocomplete with good PR.

That's a clean thesis. Whether it survives contact with reality is another question.

Why LeCun Thinks Language Isn't Enough

The argument isn't new. LeCun has been making it publicly for at least two years, clashing with researchers who believe scaling transformers on text will eventually get you to general intelligence. His counter: language is a thin slice of human cognition. Most of what we know, we learned by interacting with objects, spaces, and consequences. A child who's never touched a glass still understands that it falls and breaks, because they've touched other things.

LLMs don't have that. They have statistical patterns over tokens. And for many tasks, that's genuinely enough. But LeCun's bet is that it hits a ceiling, and we're approaching it faster than the scaling crowd wants to admit.

(I think he's at least partially right. The brittleness of current models in physical reasoning tasks isn't a data problem. It looks more structural than that.)

What AMI Is Actually Building

The source material is thin here, so I'll be direct about what's confirmed versus inferred. According to the Wired article linked in the Reddit thread, AMI raised $1 billion and is focused on AI that understands the physical world. LeCun left his role as Meta's chief AI scientist to lead it.

That's the confirmed part. The specific architecture, research direction, and product timeline aren't detailed in the available source. My read: given LeCun's long-standing advocacy for Joint Embedding Predictive Architecture (JEPA) and world models, AMI is almost certainly building in that direction. But I won't state that as fact.

What's notable is the funding size. $1 billion for a research-first AI startup with a contrarian thesis and no obvious near-term product is a serious vote of confidence from someone. That's not seed money. That's "we believe this could be the foundation of something large" money.

Where This Gets Complicated

LeCun's critique of LLMs is sharp. His alternative is harder to execute than he sometimes makes it sound. Building AI that genuinely understands physical cause-and-effect requires either massive amounts of embodied training data (expensive, slow to collect), high-fidelity simulation (which has its own sim-to-real gap problems), or a theoretical breakthrough in how models build internal world representations.

None of those paths are easy. And the LLM camp isn't standing still. OpenAI, Google DeepMind, and others are actively working on multimodal and agentic systems that incorporate physical reasoning as a layer on top of existing architectures. They have more data, more compute, and more deployment experience.

So AMI is racing well-funded incumbents with a fundamentally different architecture. That's either visionary or expensive.

The Credibility Variable

LeCun isn't a random critic. He's a Turing Award winner. He co-invented backpropagation as we use it today. When he says the current path has limits, the AI research community at least has to engage with the argument seriously, even when they disagree.

But credibility doesn't guarantee results. And the history of AI is littered with confident bets on "the right way" to build intelligence that took decades longer than expected, or didn't pan out at all. (The original expert systems era comes to mind. Very confident. Very expensive. Eventually superseded by the exact statistical approaches their proponents dismissed.)

What I Actually Think

I find LeCun's thesis more compelling than most of the scaling-solves-everything camp will admit publicly. Current LLMs are genuinely surprising in breadth and genuinely brittle in physical reasoning. That's not a coincidence. But AMI's success depends on whether "understanding the physical world" is a tractable research goal with a $1B budget and a 5-10 year runway, or whether it's a 30-year problem in disguise.

My honest guess: AMI produces real research contributions. Whether it produces a product that competes with GPT-whatever-we're-on-by-then is much less certain. LeCun is probably right about the destination. The route might be longer than the fundraise implies.

- The AI finds the signal. We decide what it means.

Pipedrive is a sales-focused CRM that actually understands how salespeople work. It prioritizes pipeline visibility over feature bloat. The deal board interface is intuitive, automation handles repetitive tasks, and reporting gives you real insight into what's closing and what's stalling.

Worth trying if you're managing a small to mid-size sales team that needs better forecasting without learning a new system every quarter. The pricing starts at $14/user monthly, which is reasonable for what you get.

Honest limitation: it's built for sales ops, not marketing. If you need deep marketing automation, HubSpot and Zoho CRM do more here.

The real win? You'll spend less time configuring and more time selling. That matters.

Try Pipedrive

Some links are affiliate link. We earn a commission if you subscribe. We only feature tools we'd use ourselves.

Some links are affiliate link. We earn a commission if you subscribe. We only feature tools we'd use ourselves.

Both sit around the same price tier (roughly $3/MTok input, $15/MTok output). Both handle code, analysis, and long documents. Both are fast enough for production use. So which one do you actually run?

Where they split:

Speed: Both are fast in single-turn use. In long agentic chains, Sonnet's lower error rate means fewer retries, so it ends up faster in practice.

Verdict: If you're building anything that involves code generation, tool use, or complex multi-step instructions, use Sonnet. Full stop. If your workflow is primarily document understanding with messy visual inputs, or you need GPT-4o's broader plugin ecosystem, go OpenAI. For pure writing tasks, flip a coin. But most practitioners running serious workloads have quietly defaulted to Sonnet, and I think that's the right call.

Qwen3.5-35B-A3B Uncensored - 35B model with zero refusals, no capability loss, fully unlocked for local deployment.

Yann LeCun's AI startup raises $1B - Europe's largest seed round funds world model research at scale.

OpenAI acquires Promptfoo - OpenAI buys AI security platform to strengthen vulnerability detection in enterprise systems.

Google embeds Gemini deeper in Workspace - Chat window in Docs, spreadsheet generation, and AI-powered Drive search rolling out now.

What caught our attention this week.

Some links are affiliate link. We earn a commission if you subscribe. We only feature tools we'd use ourselves.

This newsletter runs on an 8-agent AI pipeline we built in-house.

Want that kind of automation for your business?

From scanning 50+ sources to drafting, fact-checking, and formatting - AI agents handle 95% of this newsletter.

The AI finds the signal. We decide what it means.

Research and drafting assisted by AI. All content reviewed, edited, and approved by a human editor before publication.