1/ We made verifiable AI (zkML) 3-7x faster than everyone else! Here's how. We asked a simple question. What if we modified a16z Crypto's sumcheck based & lookup-centric zkVM "JOLT" for ML operations?👇🧵.

I am becoming more and more a webGPU maxi. Why?. WebGPU works on many devices. now its even on by default in iOS 26. We open-sourced a MSM webGPU lib and will expand support in the coming months. On-device proving — needs this.

Introducing Grok Imagine.

As 'decentralized' projects' legal framework become clear in the USA — there will be a wave of 'offshored' projects that come back. via DUNA, Clarity. This will accelerate progress on frameworks EVERYWHERE, no one wants to be the last one holding the 'crypto-unfriendly' straw.

12/ What's next? 🔮. We are standing on the shoulders of giants (who happen to work at @a16zcrypto ). Thank you🙏 . As their team continues advancing JOLT with streaming, GPU acceleration, and more order-of-magnitude improvements. We're contributing back while building the.

11/ Most use-case of zkML are privacy centric. We get true zero-knowledge through HyperNova's folding zk 🔐 . While other zkML projects are succinctly verifiable they are often not privacy preserving. Our architecture enables privacy via folding + dynamic step sizing for.

10 / But speed isn't everything - completeness and audibility matters 🛡️. Many frameworks claiming ONNX support lack memory consistency checks. Without these safeguards, you can't legitimately claim complete ONNX compatibility. JOLT-Atlas prioritizes performance AND verifiable.

9/ Benchmark results speak for themselves 📊.Multi-class model (preprocessing + proving + verification):. jolt-atlas: ~0.7s 🥇.mina-zkml: ~2.0s.ezkl: 4-5s.Others: A lot of zkML projects simply cannot support most ONNX opcodes. *Orders of magnitude improvements across all.

8/ With the introduction of Twist and Shout — the required trait for table 'decomposability' was dropped. Now we just need tables to be MLE-structured. This means JOLT 'Atlas' can support more zkML opcodes and precompiles.

7/ Another key advantage: JOLT never materializes full lookup tables (they're structured, not stored) 💾. While other zkML projects get locked into specific quantization schemes, our approach enables flexible quantization + potential floating-point support without rigid.

6/ For matrix-vector multiplication (a dominant cost in ML), we leverage JOLT's batched sumcheck protocol for exceptional performance. Unlike retrofitted approaches, JOLT Atlas benefits from sparsity natively through lookup-centric architecture - no additional optimization.

5/ JOLT eliminates complexity that plagues other approaches: .❌ No quotient polynomials .❌ No byte decomposition.❌ No grand products .❌ No permutation checks .✅ Just efficient lookups & sumcheck.

4/ Traditional circuit-based approaches are prohibitively expensive for non-linear functions like ReLU and SoftMax 📈. Lookups eliminate circuit representation entirely. JOLT was designed from first principles to use ONLY lookup arguments - giving us a fundamental advantage over.

3/ But what if you optimized JOLT specifically for neural network inference? 🧠 You get JOLT 'Atlas'; a machine learning centric (ONNX) variant of the JOLT zkVM (RISCV). ML workloads have specific patterns (non-linearities, MatMul ops) that play exceptionally well with JOLT's.

2/ A few weeks back @a16zcrypto's team dropped serious heat with their 6x speedup, proving that when you stop arithmetizing everything and embrace lookups + sparsity, beautiful things happen:. Less code to audit / Faster prover / Plausibly "Prover killer" proof.

RT @YourBuddyConner: In the face of an optimistic Jerome Powell pump, its important to remember this: . The real exit liquidity is the frie….

Whatever they call Backtrack Linux now, it needs a module like “Social Engineering Toolkit” but for AI tolls and vibe coding “Educational purpose only”. The attack inject vectors with text to code are insanely bad for normies.

Just wrapped a panel with AI experts on a nonchalant podcast. When I mentioned verifiable compute and ZK, got the classic: "What's that? I have not been following web3 much." . Cue 20-minute deep dive on verifiable compute, trustless AI workflows, ZK, and privacy-preserving.