Key points
- Harmonic uses its Aristotle agent to verify C libraries and hardware, fixing bugs and proving equivalence.
- SF-Tensor announced efforts to formalize NVIDIA’s PTX language in Lean to prove GPU kernel correctness.
- &Truth plans to use Lean for proving the correctness of AI-generated software with industry experts.
Harmonic announced last week that it uses its AI agent, Aristotle, for scalable verification of both software and hardware systems. This represents a shift toward full-stack verification using artificial intelligence.Companies are using AI to mathematically prove that their software and hardware work correctly across entire systems.Harmonic’s Aristotle agent demonstrates scalable verification for C libraries and hardware, while SF-Tensor formalizes NVIDIA’s PTX in Lean for kernel correctness.
What it does
Harmonic demonstrated the tool on a C library for big-number arithmetic, where it fixed bugs and proved that a faster version was equivalent to the original. SF-Tensor announced it was formalizing NVIDIA’s PTX language in Lean to prove the correctness of GPU kernels.
How to get it
These tools are emerging from various startups. &Truth recently came out of stealth with plans to use Lean for proving AI-generated software correctness. The team includes Clark Barrett, Ilya Sergey, and Sebastian Ullrich.
Limits
The source material does not specify pricing, availability or current user counts for these verification tools. It is unclear whether these are production-ready services or research prototypes.
The takeaway
The common factor across these announcements is the goal of achieving full-stack formalization, either across the software stack or across a combination of software and hardware. Past verification was limited to small cores or simplified models.
Building on this
Developers can build systems that prove equivalence between optimized and reference implementations. Using Lean, teams can formalize languages like PTX or develop entire systems such as AWS’s Cedar in Lean, compile to C, and deploy verified code.
Why it matters
- Data science teams can use formal verification to ensure AI-generated code behaves correctly before deployment.
- Verifying GPU kernels in Lean allows teams to trust performance-critical computations without manual testing.
- Full-stack verification reduces reliance on simplified models, offering stronger guarantees for complex software stacks.
- Proving equivalence between fast and original versions ensures optimizations do not introduce subtle bugs.
Topics: formal verification, Lean, AI agents, GPU kernels, software correctness
Updated
The strongest doubt
Counterpoint from the sources
The source notes that past verification was limited to small cores or simplified models, implying current full-stack approaches may still be nascent or complex to implement broadly.
For readers
Developers
- Use Lean to formalize languages like PTX or develop entire systems and compile to C for verified deployment.
- Apply equivalence proofs to ensure optimized code versions match the behavior of original implementations.
Researchers
- Explore how AI agents like Aristotle can scale verification across both software and hardware domains.
- Study the application of separation logic and tools like Veil in verifying AI-generated software.
Investors
- Monitor startups like &Truth and SF-Tensor that are emerging from stealth with formal verification solutions.
- Assess the market for tools that enable full-stack formalization rather than partial or simplified verification.
Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
Can AI agents fix bugs in hardware designs?
Harmonic’s Aristotle agent verifies both software and hardware, but the specific bug-fixing example cited is for a C library, not hardware.
Who is involved in &Truth’s verification efforts?
Key people include Clark Barrett, known for cvc5; Ilya Sergey, known for Veil; and Sebastian Ullrich, Lean’s second-in-command.
What did Theorem. and Anthropic verify?
They verified open-source software running on a RISC-V processor, covering encode-decode round-tripping, idempotence, and isolation properties of an xv6-based sandbox.
Connected stories
Reading receipt
- Harmonic uses its Aristotle agent to verify C libraries and hardware, fixing bugs and proving equivalence.
- SF-Tensor announced efforts to formalize NVIDIA’s PTX language in Lean to prove GPU kernel correctness.
- &Truth plans to use Lean for proving the correctness of AI-generated software with industry experts.