01ResearchInternals2 sources agree
NovaSkyAI has added Day-0 RL support to SkyRL, including LoRA, BF16 full fine-tuning and MXFP8 RL, with significant performance improvements, and reports a 62% increase in LoRA decode throughput. The implementation also includes vLLM rollouts and Reflection's Megatron-Bridge implementation.
Provenance — who else covered this
02On-deviceInternalssingle source
AMD's MI355X delivers 24.4% more throughput than NVIDIA's HGX B300 on the AgentX agentic coding benchmark, and 87.4% more tokens per dollar, thanks to vLLM optimizations and hardware details.
Provenance — who else covered this
03On-deviceProductsingle source
Inferact and AMD are collaborating to enhance production agentic inference on AMD Instinct GPUs, with a focus on long-running sessions and tool calling, and all produced work will be open-sourced. The collaboration is already in production, shipping with recent open-model releases and used by Inferact on AMD Instinct clusters.
Provenance — who else covered this
04ModelsProductsingle source
A research team has developed a model called Glimpse, which achieves frontier-level multimodal performance, outperforming Astra on 5 of 12 benchmarks and Gemini 3.8 Flash on real-world vision workloads, using a novel visual reasoning technique. The model is smaller, cheaper, and more efficient, using 17x lower token prices.
Provenance — who else covered this
05AgentsProduct3 sources agree
The Switch project introduces an open-source coordination layer that allows humans and agents from different providers to collaborate in a shared workspace, enabling seamless communication and task coordination. The project demonstrates its capabilities by resolving two incidents, an inference API issue and a recommendation-model training job failure, using agents from various providers.
Provenance — who else covered this
06RoboticsProduct2 sources agree
The lab has open-sourced V1.0 of its ASPIRE agentic robotics framework, featuring full real robot skill discovery and self-improvement capabilities, and announced benchmark results for Astra, Opus 5.5, and Sonnet 5.5 models.
Provenance — who else covered this
07ModelsProduct2 sources agree
Perplexity has open-sourced two state-of-the-art decision models, pplx-decider-flash-v1.2 and pplx-decider-v1.2, available on the Perplexity Decisions API and as open weights on Hugging Face.
Provenance — who else covered this
08ModelsProductsingle source
Beam is now available across platforms with open weights under the Apache 2.0 license, including weights, tokenizer, and reference implementation, as well as the open source harness Mirror CLI, with weights and model card available on Hugging Face.
Provenance — who else covered this
09ModelsProductsingle source
Liquid AI's new decision model, d1-ultra, achieves the top score on the Artificial Analysis Decision Index v1.0, with a lower hallucination rate but also lower accuracy compared to its predecessor d1. The model is priced at $0.042 per 1M input tokens and offers comparable pricing to other decision models.
Provenance — who else covered this
10CodingInternals3 sources agree
Modal integrated its custom ImageFS filesystem with VM Sandboxes, enabling faster performance by leveraging a tiered distributed filesystem, and built a new filesystem stack to support this integration.
Provenance — who else covered this
11BusinessProduct2 sources agree
Muse Spark Open costs $0.84 per Intelligence Index task, compared to other models like GLM-5.3 and DeepSeek V4.1 Flash, with current API pricing from Together AI, Fireworks, and Baseten.
Provenance — who else covered this
12CodingProduct2 sources agree
Reflection introduces Mirror, an open-source coding agent optimized for Beam, available for installation via uv tool, with the project's source code hosted on GitHub.
Provenance — who else covered this
13AgentsProductsingle source
GPT-6 Astra Pro in ChatGPT has been upgraded with faster responses, improved plugin and tool usage, and the ability for multiple agents to tackle different parts of a task in parallel, now available to Pro users. This upgrade allows users to follow each agent's progress and review completed tasks.
Provenance — who else covered this
14ModelsProductsingle source
A new OCR model called Indra has been launched, boasting 15x faster performance than Tesseract on CPU and improved accuracy. The model has 1M parameters.
Provenance — who else covered this
15ResearchInternalssingle source
Zephon's TorchTitan integration preserves global batches when training resumes on a different GPU count, and a 1B-parameter model test demonstrated identical global batches across various GPU counts. The integration tokenizes, packs, shuffles, and mixes data inside the training job, avoiding a preprocessed copy of the corpus.
Provenance — who else covered this
16ResearchInternals5 sources agree
The author experimented with swapping sub-calls in Recursive Language Models for Jev, finding it to be half the cost and 6× faster than Haiku 5.5 on OOLONG, and shared findings on abstention and model routing.
Provenance — who else covered this
17ResearchProduct4 sources agree
Bridgewater created a benchmark of 200 examples to measure search improvements, as explained by Aaron Linsky.
Provenance — who else covered this
18AgentsProduct2 sources agree
The new Decisions API allows Live API apps to speed up delegation by selecting from specific tools while retaining the ability to use a full LLM for long-tail queries.
Provenance — who else covered this
19ModelsProductsingle source
Gemini 4 Argon and Muse Wonder debuted in Agent Arena with net improvements of +13.2% and +6.4% respectively, while MiMo V2.6 Pro took the #1 spot among open models, and a new model router was introduced with 37% lower cost.
Provenance — who else covered this
20ModelsProductsingle source
Muse Spark Open outperforms Nemotron 3 Ultra on agentic tasks and has been added to I for B200 with more configurations to come.
Provenance — who else covered this
21ModelsInternalssingle source
Muse Open model generates 55.5M output tokens, the fewest of any open-weight model, and averages $4.08 per Index task, despite shorter outputs not reducing costs, with most cost coming from input tokens. Muse Spark 1.3 produces 77M output tokens and averages $3.38 per task.
Provenance — who else covered this
22ModelsProductsingle source
GPT Live 1.1 updates focus on multilingual support, including French, German, Hindi, Mexican Spanish, Korean, and Japanese, with a rollout in Codex, ChatGPT, and Dots, and an upcoming API release.
Provenance — who else covered this
23CodingProduct7 sources agree
Bun's next version will include a new linter, bun lint, which is 26x faster than ESLint and has over 1,000 built-in rules, compatible with ESLint and oxlint configurations.
Provenance — who else covered this
24AI securityProduct5 sources agree
Stolen AI API keys are being used by hackers to pay for autonomous scans, according to a report from BitSight, highlighting a cybersecurity threat, and the issue is related to the dark web and threat intelligence.
Provenance — who else covered this
25AI securityInternals4 sources agree
A practitioner lists their top picks for AI evaluations in offensive cybersecurity, including the dreadindex, AikidoSecurity, BArena, and nexbench, each with unique features such as vuln category level evaluations and cost per finding.
Provenance — who else covered this
26AgentsProduct3 sources agree
A panel of top AI power users discuss and compare various AI models and tools, including Dots, Muse, GrokBot, and Instinct, and share their experiences with building personal assistants and decision models. The discussion covers topics such as model speed, writing quality, subagents, and context management.
Provenance — who else covered this
27ModelsProduct2 sources agree
Opus 5.5 has been released with a fast mode that achieves 2.5x speed.
Provenance — who else covered this