Auto Added by WPeMatico
Most visual document retrievers in production today are hand-me-downs. ColPali and the models that followed it take a generative vision-language model and repurpose it as an encoder. The result still carries a separately pretrained vision tower and a causal decoder that never generates a token. That is parameter and compute overhead for a task that
AI research agents can already propose, implement and score their own machine learning experiments. Idea generation is cheap; verification is not. Training one candidate can consume hours to days of GPU time, so an agent proposes far more candidates than it can afford to run. Which ones get run is the real lever on research
A team of researchers from UC Berkeley have released CUA-Lite, an open platform for computer-use agents (CUAs). The argument behind it is infrastructural rather than model-centric: training and benchmarking a CUA requires four pieces: agents, environments, traces, and a framework to evaluate and train them and all four are currently fragmented across separate repositories with
Video has been the most expensive modality to reason over. A Gemini model handed a 90-minute lecture has, until now, ingested the whole thing at a fixed one frame per second, whether the question was ‘summarize this’ or ‘what time does the speaker switch to the pricing slide?’ That single-pass design forces a bad trade:
Multi-agent workflows have changed the shape of local inference. A lead agent decomposes a task and spawns subagents. What looked like one user request becomes dozens of independent model calls. Pointed at a single local engine, those calls compete for the same execution slots. The queue grows while a workstation, laptop, or DGX Spark on
AI weather models have spent three years closing the gap with physics-based forecasting, but two problems stayed open: resolution too coarse for local terrain, and initialization tied to numerical weather prediction (NWP) analysis that arrives about six hours late. WeatherNext 3, released by Google DeepMind and Google Research, attacks both. It takes a live global
Today, OpenAI released GPT-6 Astra. The company calls it its most intelligent and aligned model, and positions it primarily as a computer-use system rather than a chat model. The pitch is that Astra operates software the way a person does, across browsers, spreadsheets, desktop applications and terminals, and finishes multi-step jobs instead of describing how
This week, Meta Superintelligence Labs released Muse Spark 1.3. It is the fourth Muse Spark release in five months, and the target is long-horizon agentic and coding work rather than single-turn generation. The framing in Meta’s post is usability: sustaining a long thread, collaborating with the user, and knowing when it is stuck. Is it
Perplexity has open sourced Lily, the local inference engine behind Hybrid Compute in Perplexity Computer. It is a single-process runtime: a Rust layer loads the checkpoint and drives the generation loop, an OpenAI-compatible chat-completions API streams tokens, and hand-written Metal kernels execute the model. Neither PyTorch nor MLX sits in the execution path. Lily is
Coding agents spend a large share of their tool budget on search. When the target is a known symbol, ripgrep answers it exactly. When the target is a behavior described in plain language, keyword matching often misses, and the agent falls back to guessing terms, reading whole files, and assembling context by hand. Each of
