Gritt Bolts AI Onto Excavators and Starts Building Solar Farms 3x Faster

Gritt Bolts AI Onto Excavators and Starts Building Solar Farms 3x Faster

Gritt, founded by Carnegie Mellon roboticists, came out of stealth in July with $32.4 million (a $26M Series A led by Obvious Ventures, plus seed money from First Round) to put AI-guided precision onto equipment that's already on job sites — skid steers, robotic arms, the works [1][2]. No humanoid theater here: their systems help an 8-person crew go from 800 panels a day to 3,000–4,000, and they've already got contracts covering 2.8 GW of solar installs with major US power construction firms [2][3].

This is physical AI doing the unglamorous, high-leverage thing: retrofitting instead of replacing, accelerating a labor-constrained industry instead of reinventing it. Solar is the beachhead; data centers and broader infrastructure are next.

Pair this with FieldAI and the pattern is clear — the most fundable "AI products" in late 2026 increasingly ship with a chassis attached. If your roadmap still assumes AI means "a better chatbot," you're reading last year's memo.

Yann LeCun Keeps Betting Against LLMs — and Open-Sourcing the Alternative

LeCun's camp dropped H-JEPA results this cycle: a hierarchical world-model architecture that hit 73% success on Visual AntMaze tasks versus 18% for single-layer baselines — and they open-sourced the full code and weights [1]. The timing lines up with LeCun's increasingly pointed public stance that auto-regressive LLMs are a dead end for human-level intelligence because they can't rapidly adapt or build real world models — they predict tokens, not consequences [2].

He's not just a critic from the sidelines anymore. The AI Alliance named him Chief Science Advisor in April for Project Tapestry, an effort to build open, federated frontier models rather than let a handful of labs own the stack [3]. His pitch to academia: stop chasing marginal LLM gains, go build grounded, physical, world-modeling systems instead.

The X debate has been sharper than usual — people pushing back on "LLMs can't search or reason" claims, others pointing out LeCun has been saying some version of this since 2022 and the architectures keep improving anyway. Worth watching regardless: if H-JEPA-style approaches scale the way FieldAI and Gritt's embodied systems need them to, open-source world models could become the default substrate under physical AI — not proprietary LLM APIs.

EU AI Act: Agents Don't Get a Pass, Deadlines Get a Breather

The EU AI Act's core obligations kicked in this August, but the Digital Omnibus (in force since July 27) pushed the high-risk Annex III compliance deadline — recruitment, credit scoring, HR systems — to December 2027, with embedded high-risk systems getting until August 2028 [2]. Transparency rules are not delayed, though: as of August 2, companies must disclose AI interactions and label synthetic content [1].

The more consequential clarification: the European Commission confirmed AI agents don't need a new legal category — they're already covered under existing AI system and GPAI definitions [3]. That means logging, human oversight, least-privilege design, and accountability chains apply to your autonomous agents right now, not in some future "agent regulation" that never arrives.

European builders got a longer runway on the heaviest compliance lift, but the message on agents is "no loopholes." If you're deploying agentic systems in the EU, the honest framing circulating on X — treat your agent like a new hire, with onboarding, permissions, and a manager who's accountable for its mistakes — is closer to the legal reality than most vendors are admitting.