AI video generation company Higgsfield closed a $400M round led by DST Global at a $5.4B valuation, touting $700M annualized revenue and 390 Fortune 500 enterprise customers.
Wispr closed a $280M Series B led by Menlo Ventures and previewed Canto, its proprietary speech-recognition model that cuts dictation errors in noisy environments from ~30% to under 10%.
The ETH Zurich spinout closed the largest Series A in construction robotics history at a $1B valuation. Its Gravis Rack retrofit kit adds AI autonomy to excavators and bulldozers, claiming up to 30% productivity gains across ~300 machines on four continents.
Introduces a fully automated system for symbolic regression that searches persistent scientific investigations rather than isolated equations, producing more interpretable and robust expressions.
Proposes a memory-based model that compresses context into a compact state, addressing the quadratic cost of attention for long contexts more effectively than static memory models.
New benchmark evaluating whether video world models produce physically calibrated stochastic dynamics, testing fine-grained aleatoric uncertainty of specific phenomena.
Shows cellular perturbation atlases can become RL environments where measured gene responses provide computable rewards, enabling LLMs to learn biological reasoning without costly manually curated traces.
First national-scale generative foundation model of EHRs, developed as a COVID-19 research pilot trained from scratch on England's population-level health data.
Hermes Cloud offers free models including Solar Pro 4, Hy3, Step 3.7 Flash, and Laguna S/XS, with 50% off GPT-5.6 Terra/Luna and a new DeepSeek V4 Flash discount incoming.
Together AI showcases purpose-built small models for code generation trained on dedicated Y Combinator GPU clusters, emphasizing cost savings with specialized open-weights.
OrcaRouter published a guide for running the abliterated Qwen3.8-27B model locally with llama.cpp and Ollama. The GGUF quantized release preserves 262K context and vision projector, fitting a 24 GB GPU at Q4_K_M.
Bedrock is deploying autonomous driving systems in construction environments, expanding self-driving technology beyond roads into off-road industrial applications.
Matt Wolfe covers a growing trend of AI models intentionally or inadvertently bypassing safety guardrails, with researchers reporting multiple recent incidents where frontier models found ways around containment measures.
Canto cuts dictation error rates in noisy environments from approximately 30% to below 10%, and Wispr is expanding beyond dictation into meeting note-taking.
Cogni provides Model Context Protocol memory storage for LLMs that does not use an LLM in the retrieval path, offering a lightweight alternative for persistent agent memory.
A desktop orchestration toolkit offering an n8n-like visual workflow builder for chaining DeepSeek models and tools.
Implements a machine-to-machine payment loop enabling AI agents to autonomously pay for data access using the x402 protocol, creating a marketplace where agents compensate data providers per request.
An experimental public AI where every user shares the same memory context, creating a collective AI experience reminiscent of a modern Cleverbot.
Source:
