Simon Willison tests the cost-efficient LLM, which lands on the Artificial Analysis Pareto frontier at a fraction of comparable costs (~0.44 cents default inference) with a 98% cache-hit discount seen as transformative for agent loops.
More than 1,000 employees from OpenAI, Anthropic and other labs signed the statement asking the US government to build tools deliberately designed to slow down automated AI development.
The automaker will unveil its first humanoid robots at its Di Space experience venues, positioning BYD against Tesla in the embodied AI race.
A controlled study from 1.5B to 7B models shows reflection methods that generate extra text are beaten by simple repeated sampling at equal token budgets, isolating the accuracy gain of reflection to generating more tokens.
The pipeline generates reset-able synthetic environments for computer-use agent training, targeting stateful login-gated applications and scaling beyond prior bulk pipelines.
Scales a parametric long-term memory module to much larger capacities, decoupling memory from reasoning parameters for independently scalable long-term memory.
Token-level dynamic width pruning allocates compute per token to reduce accuracy loss under aggressive sparsity versus static structured pruning.
A single learnable embedding as an explicit retrieval token lets VLMs handle long visual context with many distractors while staying within GPU memory.
An open-source AI coding agent that works across the IDE, command line and SDK, positioning itself as a cleaner alternative to Claude Code left-behinds.
A Rust-built cockpit dashboard giving users visibility and control over their Claude Code agents in one place.
A real-time data feed/wire service for AI agents, claiming 70,000+ connected sources for live agent inputs.
Tracks live per-prompt pricing across AI APIs and offers real-workload cost calculators to compare model expenses.
A gate mechanism that filters untrusted external sources before they can influence or become agent skills, adding a safety layer to agent tooling.
The Milan-based software house has lost over 13% from its $40 peak since its July 1 Nasdaq debut, and announced €500M in new medium-long term financing from HSBC, Intesa Sanpaolo and BPER backed by SACE, bringing funding lines to €1.49B as it pursues further acquisitions.
The B2B AI company listed on Euronext Growth Milan paid the first €75,000 tranche of a €150,000 capital increase in EdiBeez, taking a roughly 3.5% minority stake.
Source:
