Fern 0.1 is Mirendil’s first model, pretrained from scratch with up to 10M tokens of context. It scores competitively on the agentic coding benchmark, on par with reflection Beam.
Key points
Context length: up to 10M tokens (~0.5M lines of code)
Benchmark: agentic coding benchmark scores competitive, on par with reflection Beam
Sources
Anthropic files S-1 with SEC for proposed IPO
Anthropic filed a registration statement on Form S-1 with the SEC for its proposed initial public offering. The company plans to list Class A common stock on the Nasdaq Global Select Market under the symbol “ANT”. The offering details, including share count and price range, have not been set.
Key points
Filed registration statement on Form S-1 with the U.S. Securities and Exchange Commission.
Intends to list Class A common stock on Nasdaq Global Select Market under symbol “ANT”.
Sources
The Grid robot launch by CircuitHub builds PCBs on demand
The Grid robot by CircuitHub is a factory-scale system in Massachusetts that builds circuit boards on demand. It operates with no changeovers and no per-design setup, handling prototypes and production batches fully automated.
Key points
The Grid is a factory-scale robot located in Massachusetts.
It builds circuit boards on demand with no changeovers and no per-design setup.
Sources
Devin Security Platform adds Claude Mythos and credits
Devin Security Platform now includes Claude Mythos and offers hundreds of millions of dollars in credits for vulnerability work. It claims to help engineering teams write more secure code and secure the code they already have.
Key points
Claude Mythos and frontier cyber models are available in Devin Security Platform.
Hundreds of millions of dollars in credits are offered to help teams find, prioritize, and fix urgent vulnerabilities.
Sources
Foundation4D released by NVIDIA, returns depth and more
Foundation4D released by NVIDIA returns depth, camera poses, 3D boxes, semantic occupancy and renderable Gaussians from multi-camera driving video. It runs a single forward pass using one shared backbone with a decoder for each task.
Key points
Input: multi-camera driving video
Outputs: depth, camera poses, 3D boxes, semantic occupancy, Gaussians
Sources
Latent-Foresight learns compact VFM latent, predicts
Latent-Foresight learns a compact VFM latent representation and a generative dynamics model that predicts future. It jointly learns both components.
Key points
Latent-Foresight jointly learns a compact latent representation of visual foundation model (VFM) features.
Latent-Foresight jointly learns a generative dynamics model that predicts their future.
Sources
Proximity Conjecture for F-avoiding Bases Paper Released
New paper settles the unweighted and weighted proximity conjecture of F-avoiding bases for group-labeled matroids. It also introduces FPT algorithms for finding minimum-weight F-avoiding bases.
Key points
Settled the unweighted and weighted proximity conjecture of F-avoiding bases for group-labeled matroids.
Developed FPT algorithms for finding minimum-weight F-avoiding bases.
Sources
- Original post
- arxiv.org - Linked in the post