meet @mostik_ai! what happens when you put 12 PhDs in one room for four months? first place on the ARC-AGI leaderboard, which I can't say much about while the competition is still running. and this, which I can. everyone's arguing about whether open models will catch up to frontier models. we think it's the wrong question. here's the one we pose: why does a frontier model have to generate your answer at all, when the only thing you need from it is the reasoning? we do this by enabling models to communicate in latent space. through our protocol, hidden states pass straight from a frontier model into a small one running on your infrastructure -- no text between them, and neither model is fine-tuned. two models from different families, sharing reasoning, both left untouched. how do we know it works? we tested it on a setup where a 753B model reads the problem, and a 4B edge-class model writes the answer. with this approach, we get results 80% as accurate as the frontier model, but at 20x faster performance. we're committed to preventing frontier model lock-in and are already partnering with inference providers to accelerate open-weight adoption. we've done this between 15 of us, in four months, 12 PhDs and a Fields medalist, backed by @generalcatalyst WIRED has the first external account of the company and the work: wired.com/story/russian-star… full writeup, the setup, and all the numbers: mostik.ai/read-more

12 PhDs and a Fields Medalist in the mountains arguing about latent communication, and it happened to be my birthday 🎂 taking suggestions on how to top that 😂

Two weeks since launch, and the public benchmarks are in. Recraft is officially the #1 independent image generation lab 🚀

Vera nice indeed! As SpaceX, OpenAI, & Anthropic start getting NVIDIA Vera CPUs, they may soon discover that the only open-source codebase GitHub can find optimized for Vera is my NumKong mixed-precision library 😎 vdotq_f32_mf8 is already there: github.com/search?q=vdotq_f3… The port-mapping docstring is stale & entirely wrong, but until the hardware arrives, all of us are benchmarking vibes. Will share real numbers once Vera reaches open-source maintainers.

We’re excited to share that our team at @xmemory_ai has published our main white paper. The core idea is simple: without focus, AI systems try to "remember" everything, and that is where they fail. Schema is the best way to teach them focus. We compared xmemory against major open-source and commercial RAG and hybrid RAG systems, markdown-file memory approaches ("agentic memory"), and memory implementations in customer-facing frontier AI apps. Instead of testing whether systems can recall similar text, we evaluated whether they can store, update, deduplicate, and retrieve facts and relationships reliably. xmemory achieved 97% accuracy, compared with 87% for the strongest competitor. Our core also outperforms the latest frontier models’ one-shot APIs in extraction tasks, showing that the harness matters just as much as model quality. We’ve also open-sourced our measurement datasets and a toolkit to generate new ones synthetically. Huge thanks to the dream team of engineers at xmemory who keep pushing this vision into production. If your company is seeing memory systems fail in real workflows, we’d love to talk.

Today we’re sharing our first blog post: “Schema as the Core of Reliability.” Our view is simple: AI memory breaks down when facts are stored as unstructured text and reconstructed later. That approach can work for thematic recall, but it becomes fragile when systems need to answer exact questions, maintain state, support workflows, and drive automation. In the piece, we explore why: - Search can recover context, but memory must support facts - Graph RAG is a meaningful step forward, but not the endpoint - Reliable memory needs structure - types, relations, constraints, deduplication, provenance - Schema is not decoration - it is the contract that makes memory observable, enforceable, and dependable If AI systems are going to do more than sound coherent - if they are going to make decisions, trigger workflows, and operate over long horizons - memory has to become a governed system of facts. That is where reliability starts. Read the full post - link in the first reply.

My biggest open-source release! NumKong — 2'000+ SIMD kernels for mixed-precision numerics, from Float6 to Float118. Started in 2023. Opened the PR in 2024. Finally, merged this week! RISC-V, Intel AMX & AVX-512, Apple SME & SVE, WASM Relaxed SIMD. 200'000 lines of code in a 5 MB binary. Same scale as OpenBLAS. Available for C 99, C++ 23, Python 3, Rust, Swift, GoLang, & JavaScript. Int4 dot products via nibble algebra. Ozaki Float64 GEMMs on Float32 tile hardware. 6-bit and 8-bit floats back-ported to 10-year-old CPUs. 5'300x faster Geospatial metrics than GeoPy. 200x faster Kabsch than BioPython. 0 ULP where OpenBLAS hits 56... and a lot more! pip install numkong Or pull it from NPM, Crates, GitHub... and let me know what breaks 🤗 Links & highlights ⬇️

Jensen just announced the start of the GPU-accelerated database era at #GTC26. AI runs on GPUs. But your data still runs on CPUs. That mismatch is breaking the AI stack. For the last two months, we’ve been busy solving this problem. Excited to announce Deeplake becoming the GPU Database. Deeplake brings your database directly onto the GPU, eliminating the CPU <-> GPU bottleneck for AI workloads. The pendulum has switched. GPU-native queries are now 10× faster and an order of magnitude cheaper to run. Last week we even put up a 101 banner in San Francisco. And this is just the beginning. We’re planning a huge set of announcements starting this week. Stay tuned.

Excited to share that @lmnrai has raised $3M to build open-source observability for long-running AI agents. Laminar is how companies like @browser_use, @OpenHandsDev, and Rye see what their agents are doing, understand why they fail, and spot patterns across millions of runs.

Today is a big day for us. We've just released SOUL 2, the next generation of our flagship in-house image model. This is a product we've been working on for the past six months. A bit of backstory: when we released the first SOUL, the response was absolutely rewarding. It became the go-to model on our platform, and grew into a daily creative tool for hundreds of thousands of people. It was the first image model in the AI space that didn't have the "AI plasticity". It had taste, and we're happy that creators and brands have adopted it for real commercial work. To take that further, we've worked closely with creatives, art directors, photographers, stylists, and concept artists to build SOUL 2. They helped define what quality really means. We're very excited for beauty, fashion, and lifestyle brands to adopt Soul 2 for their ideation and photoshoot workflows. And with SOUL ID, we're opening that door even wider, now, anyone can be a model. Higgsfield built this to bring high-quality artistic taste to AI and let the two worlds elevate each other. To celebrate the launch, we're giving up to 5,000 free SOUL 2 generations per user, depending on your plan. Can't wait to see what you create.

📷𝐏𝐨𝐝𝐜𝐚𝐬𝐭𝐥𝐞 𝐢𝐬 𝐧𝐨𝐰 𝐨𝐟𝐟𝐢𝐜𝐢𝐚𝐥𝐥𝐲 𝐀𝐬𝐲𝐧𝐜! No, we weren’t acquired. No, we’re not going anywhere. Just getting smarter, faster, and fully AI-powered. Step inside Async.com unleash your ideas now! More details coming soon!

🔥 Firebird AI 👉 $4B investment in 🇦🇲 Armenia US @VP @JDVance announced approval of export licenses for Phase 2 of @FirebirdCloudAI’s project-launching the region’s largest AI supercomputer (up to 50,000 @nvidia GPUs) & marking one of the largest 🇦🇲🇺🇸 tech partnerships to date.

The best products of the 21st century have yet to be shipped. With 21st, any idea can find its aesthetic and design, even if you aren’t a designer. Meet the first vibe crafting tool for everyone…. So every product can have soul.