Logical Intelligence

We built a demo where you can watch Kona, the world’s first Energy‑Based Reasoning Model, solve hard Sudoku puzzles side‑by‑side with frontier LLMs. After a week of public access, Kona solved 96.2% of puzzles at 313 milliseconds each. In contrast, the LLMs typically produced an incorrect answer or simply timed out. Why measure in milliseconds? Because it’s the time necessary for firmware in safety‑critical systems to react to hardware. Anything slower can mean outright failure. And the cost gap: only $4 of GPU compute for 13,000 Kona solves, versus $11,000 in API spend for the same workload on LLMs. Sudoku may seem like a toy game, but that simplicity is the point. It’s the cleanest, most accessible test of spatial constraint satisfaction, and it clearly exposes what language models can and cannot reliably do. Read about why it matters on our new blog: https://lnkd.in/eqeCWj2Y Try the demo yourself: https://lnkd.in/eXFkXecc

Energy-Based Reasoning Models will become the "brain" in a layered AI ecosystem, complementing LLMs which become the interface layer, best suited to natural language tasks Read about it here from WIRED https://lnkd.in/drdtmy-N

Some news at Logical Intelligence, as we bring Energy-Based Models (EBMs) forward to do actual reasoning, for critical systems touching real-world assets and risk: Yann LeCun joins us as Founding Chair of our Technical Research Board, guiding our work on true reasoning, world models, and optimization, and, Patrick Hillmann, former Chief Strategy Officer at Binance, joins as our Chief Strategy Officer to help scale this vision globally. EBMs and agentic systems are a critical foundation for the next era of AI.

The next step on the path to artificial super-intelligence is moving from underspecified and expensive vibe code to at-scale, low-cost, generation of formally-verified, hallucination-free, mathematically-guaranteed code from natural language specs, resulting in autonomous systems safe to deploy in critical environments. We announced last week that our Aleph agent achieved a 99.4% score on PutnamBench, the hardest formal mathematics benchmark, solving 668 of 672 problems with machine-verified proofs. Driven by OpenAI ’s GPT-5.2, the only model capable of delivering cost-efficient, at-scale formal verification, Aleph’s breakthrough comes from verification-guided search: letting models propose candidates, letting compilers filter them, and using that signal to guide the search toward guaranteed correctness. Generation is now cheap; verification is the filter. Next week we'll have an announcement to share about the team, the mission, and another step forward in the ecosystem. Follow Logical Intelligence and our Founder and CEO Eve Bodnia for updates. #AI #FormalVerification #OpenAI #EnergyBasedModel #LogicalIntelligence

Just as I am getting ready to head to AWS #reinvent ... A set of game changers in AI today. Agents doing math and logic that is provable. This is critical for any mission critical agent to use AI. (Even truly optimal code generation).   First Harmonic announced they proved the Erdos #124 problem completely by itself in LEAN.   Right after that the brand new Logical Intelligence proved it in 2.5 hours (less than half the time of harmonic!).   Only 24 hours later Logical Intelligence Aleph Prover Agent became number 1 on the Putnam benchmark! Doing proofs for 500 of 660 hardest math problems all by itself. At an average cost of 23$ (I would need more than that in coffee cups to even start to work on solving one of them 😄)   We are on the cusp of a huge upheaval in AI. LLMs are fantastic, but fraught with basic design issues that limit their use for mission critical systems. The next generation are models that truly understand and use abstract math and have no need to be non deterministic. Provable AI solutions ... the new frontier! To me it is very akin to general purpose CPU vs GPU. Hmm who has the larger market cap :) https://lnkd.in/gj5gh8MN Eve Bodnia. Vlad Tenev.

Excited to build next generation of AI together!

Introducing Logical Intelligence, the company dedicated to making mission-critical software secure. We are rolling out two complementary AI agents built on our in-house research: Aleph, an AI formal verification agent that provides mathematical guarantees for code correctness, which can generate proofs that are tens of thousands lines long with 100% accuracy and provide visibility into a codebase’s current state; and Noa, an audit agent that highlights today’s vulnerabilities and demonstrates the need for stronger guarantees. In addition to these tools, we are developing a novel foundation model, Logical Intelligence 1.0 (LI-1.0), a non-autoregressive, energy-based model designed for mathematically precise reasoning, enabling the generation of formally verifiable code at scale. Our approach enables us to rapidly add support for new verification targets without spending thousands of engineering-hours, resulting in correct code in a fraction of the time of traditional methods. Most AI (like LLMs) solve problems one piece at a time—like doing a jigsaw blindfolded. If you place a piece wrong early, the whole picture suffers. LI-1.0 thinks differently: all the pieces move together, like magnets snapping into place until the puzzle is complete. Holistic, precise, scalable reasoning. Currently we support multiple major smart-contract VMs as well as Rust and Golang programming languages, and are continuously working on expanding the supported use-cases with no human supervision. We envision the future where 100% of generated code is formally verified, ensuring complete security as more value enters the digital world. Logical Intelligence has built a team of elite talent, including ICPC winners, world champions, and double world champions; Defcon CTF finalists; a Fields Medalist; IMC champions; and academic world leaders. Logical Intelligence thanks Mithril for providing scalable, dependable GPU infrastructure and exceptional support, and we are thrilled to announce our collaboration with the Solana Foundation 🚀.

Introducing Logical Intelligence, the company dedicated to making mission-critical software secure. We are rolling out two complementary AI agents built on our in-house research: Aleph, an AI formal verification agent that provides mathematical guarantees for code correctness, which can generate proofs that are tens of thousands lines long with 100% accuracy and provide visibility into a codebase’s current state; and Noa, an audit agent that highlights today’s vulnerabilities and demonstrates the need for stronger guarantees. In addition to these tools, we are developing a novel foundation model, Logical Intelligence 1.0 (LI-1.0), a non-autoregressive, energy-based model designed for mathematically precise reasoning, enabling the generation of formally verifiable code at scale. Our approach enables us to rapidly add support for new verification targets without spending thousands of engineering-hours, resulting in correct code in a fraction of the time of traditional methods. Most AI (like LLMs) solve problems one piece at a time—like doing a jigsaw blindfolded. If you place a piece wrong early, the whole picture suffers. LI-1.0 thinks differently: all the pieces move together, like magnets snapping into place until the puzzle is complete. Holistic, precise, scalable reasoning. Currently we support multiple major smart-contract VMs as well as Rust and Golang programming languages, and are continuously working on expanding the supported use-cases with no human supervision. We envision the future where 100% of generated code is formally verified, ensuring complete security as more value enters the digital world. Logical Intelligence has built a team of elite talent, including ICPC winners, world champions, and double world champions; Defcon CTF finalists; a Fields Medalist; IMC champions; and academic world leaders. Logical Intelligence thanks Mithril for providing scalable, dependable GPU infrastructure and exceptional support, and we are thrilled to announce our collaboration with the Solana Foundation 🚀.

Excited to announce that Chief of Science of Logical Intelligence, Fields Medal-winning mathematician Dr. Michael Freedman, will be delivering a captivating talk at the Millennium Prize Lectures! Abstract: The AI age requires us to re-examine what mathematics is about. The Seven Millenium Problems provide an ideal lens for doing so. Five of the seven are core mathematical questions, two are meta-mathematical – asking about the scope of mathematics. The Poincare conjecture represents one of the core subjects, manifold topology. I’ll explain what it is about, its broader context, and why people cared so much about finding a solution, which ultimately arrived through the work of R. Hamilton and G. Perelman. Although stated in manifold topology, the proof requires vast developments in the theory of parabolic partial differential equations, some of which I will sketch. Like most powerful techniques, the methods survive their original objectives and are now deployed widely in both three- and four-dimensional manifold topology. The boundaries of mathematics and its profound implications in the AI era will be explored in depth. #MillenniumPrize #Mathematics #AI #LogicalIntelligence #MathematicalAI #MathematicalIntelligence https://lnkd.in/g7AwJz8N