Anthropic builds Scion VCS as Claude pushes daily merges past 30,000

Per-task pricing under US$0.09 on ARC-AGI-1 and ARC-AGI-2 makes Haiku 5.5 one of the cheaper options to clear the 61% mark on ARC-AGI-2, a benchmark where many larger models struggle. Teams running agent or reasoning workloads at scale can use these numbers to weigh small-model cost against accuracy. Comparisons with other labs' verified scores are available directly from ARC Prize.

官方账号ARC Prize@arcprizeSource

The core shift is system-level: LES moves the bottleneck from memory capacity to flash bandwidth and expert-prediction hit rate, letting a 35B-class model live on consumer phones without a cloud round-trip. Developers building on-device apps can fine-tune with SRFT's 4:1 task-to-replay data mix while keeping general abilities largely intact. Weights and code are on Hugging Face and GitHub, so teams can experiment today.

shao__meng@shao__mengSource

This removes a practical blocker to speculative decoding: teams no longer need head-equipped drafters or per-workload tuning to get adaptive verification benefits. Developers serving MTP or EAGLE drafts in vLLM can enable it via a single flag, provided their attention backend supports variable-length verification. Built by Giancarlo Delfin (inferact) on work by Red Hat AI and NVIDIA contributors.

官方账号vLLM@vllm_projectSource

Idle compaction quietly solves a real cost problem: long sessions that sat untouched were losing their prompt cache and forcing full reprocessing on the next message. Teams running self-hosted runners or managed deployments get the most out of this release, from drain-time controls to attribution-header toggles behind a Claude apps gateway. Admins should review the new managed-settings hooks before rolling it out.

官方一手Claude Code: GitHub Releases@ashwin-antSource

The credit-assignment design matters: by rewarding strategies across many generated prompts, RISE keeps finding attacks even against heavily filtered models, a common failure point for red-teaming tools. Safety teams running image generation services can deploy it directly against their own models since both code and attacker models are public. It contrasts with closed audit tools by making the full attack search reproducible.

rohanpaul_ai@rohanpaul_aiSource

The notable shift is that document pipelines no longer need per-format templates, since the agent generalises to unseen layouts on its own. Teams processing invoices, contracts or forms can follow the tutorial to assemble the same pipeline on AWS instead of paying for specialised IDP vendors. It is a hands-on build guide rather than a product announcement.

官方一手AWS Machine Learning Blog@Erik CordsenSource

Course materials are fully open, so self-learners can rent GPUs on Modal and replicate assignments without Stanford enrollment. The exam-heavy design assumes students use AI agents for experiments, then test mastery without them — a template other educators may copy. It suits practitioners who want empirical intuition for LLM training behavior rather than another from-scratch coding course.

shao__meng@shao__mengSource

The lecture distils hard-won engineering lessons, including a Token-In-Token-Out design mistake that destabilised industry-wide training, and argues that the architecture endgame is everything-as-a-service plus online RL. Engineers building RL training loops or agentic RL infrastructure can use the free slides and videos to benchmark their own system design against CMU's framework. All materials are openly accessible online, so availability is not region-restricted.

shao__meng@shao__mengSource