Why Listen Labs Walked Away From a $1.5B Round for Salesforce Talks
Listen Labs reportedly abandoned a signed $125 million Series C term sheet at a $1.5 billion valuation as it entered acquisition discussions with Salesforce.
Read moreClaude API relay guides, detection insights and hands-on LLM API benchmarks
2320 articles
RSS feedListen Labs reportedly abandoned a signed $125 million Series C term sheet at a $1.5 billion valuation as it entered acquisition discussions with Salesforce.
Read morevLLM’s follow-up on MiniMax M3 shows how serving gains on AMD Instinct MI355X came from repeatedly locating the active bottleneck rather than relying on one breakthrough kernel. The work spans local shapes, redundant computation, metadata movement, backend dispatch, and queue capacity.
Read morevLLM’s Tiered KV Cache Offloading extends KV storage from GPU memory to host RAM, filesystems, object stores, and remote peers. The design can replace some recomputation with reloads while enabling cache sharing across instances.
Read moreOpenWAM introduces a modular research stack for studying how generative world knowledge can be converted into executable robot actions. Its controlled studies identify design principles for knowledge transfer, world-action synergy, and embodied generalization.
Read moreGander is designed to move beyond turn-based assistants by continuously processing video, speech, and text. Its Cerebellum-Brain architecture separates low-latency interaction from complex reasoning while keeping both sides connected.
Read moreAuK presents an open-source foundation model that uses natural-language instructions and audio context to handle speech generation, editing, enhancement, and acoustic manipulation. Its distilled AuK-Flash variant targets much faster inference with only four steps.
Read moreRadixArk has open-sourced Miles v0.1, a full-stack system designed to make large-scale post-training more verifiable, customizable, and production-ready. It connects rollouts, training, and model-weight synchronization in one architecture.
Read moreWeak-to-strong transfer does not have to mean copying a weak teacher. On-Policy Reverse Distillation uses teacher policy shifts to accelerate verifier-supported updates while preserving the student’s own optimization target.
Read moreA study of instruction-tuned models from 1.7B to 70B parameters suggests that many models recognize structurally unanswerable math and code prompts before generation. Their failure is better explained by a routing gap between recognition and abstention than by an inability to detect the problem.
Read moreCounter-Swarm Doctrine argues that agent security cannot be reduced to inspecting isolated runs. Shared files and persistent state can carry instructions across executions, requiring defenders to reconstruct broader coordination episodes.
Read moreMassachusetts will require data centers with peak demand above 25 megawatts to cover their electricity needs with clean generation. Developers may need to build power on-site, fund nearby generation, or contribute to a ratepayer protection fund.
Read moreSuno’s v6 family is the company’s first music model developed with support from the record industry. It improves genre matching, editing, and multimodal prompting, but still struggles to reproduce the deliberate irregularities of human performance.
Read moreOpenAI says an unreleased model found a solution to the Navier–Stokes problem in about 88 hours. The claim has also triggered questions about data provenance, academic competition, and whether AI is weakening the trust that mathematics depends on.
Read moreUS security agencies accuse six Chinese AI companies of extracting capabilities from American frontier models through bulk queries, fake accounts, and prompt injection. They also recommend quietly routing suspected attackers to weaker models, a step that could affect legitimate users.
Read moreMicrosoft has reached an agreement with the American Federation of Teachers and its New York affiliate on enforceable principles for AI in schools. The terms cover student-data training, AI companions, transparency, and human review of high-risk decisions.
Read moreOpenAI says Paul Christiano has joined the OpenAI Foundation Board and its Safety and Security Committee. His background in AI alignment, safety, and standards adds specialist experience to the foundation’s governance structure.
Read moreNeoHorse-1 uses routing traces, tool interactions, and evaluation feedback to guide the next round of agentic post-training. The work explores a feedback loop rather than relying only on larger models.
Read moreDriveZero combines closed-loop reinforcement learning with a vision foundation model to reduce autonomous driving’s dependence on human trajectory logs. Its design separates perception from action, then reunifies them through distillation into a camera-only planner.
Read moreLong chain-of-thought generation makes a model’s KV cache grow with every decoding step, creating a serious memory bottleneck. BeaconKV keeps compact representatives of recurring query patterns to predict which distant context should remain available.
Read moreGE-Act 2.0 trains a world-action model from scratch on manipulation data instead of inheriting a pretrained video generator. Its results suggest that larger and more diverse co-training data can improve zero-shot robotic manipulation across embodiments and environments.
Read moreMOLE is an open benchmark for evaluating whether security monitors can detect harmful actions by AI agents operating shared services under limited review budgets. Its results show that an agent’s refusal behavior is a poor indicator of whether it will ultimately complete a harmful objective.
Read more