LFM2.5-VL-3B: Faster edge vision capabilities with a 3B multimodal model
LiquidAI’s LFM2.5-VL-3B targets real-time vision-and-language workloads by compressing capability into a compact 3B parameter model designed for speed on the edge. The key news is how the release frames deployment tradeoffs – aiming for strong multimodal performance without the compute and latency burden that typically forces larger (and more expensive) systems.
Twitch streamers can now opt out from training Amazon’s AI
Twitch is adding a clear opt-out switch so streamers can prevent their streams, VODs, clips, and other channel content from being used to train Amazon generative AI models. The policy still keeps “AI-supported” features like captions and safety tools running, but it also makes cross-stream chat usage dependent on the other channel’s opt-out preferences.
Flow-by-Flow:Content-Judgment Bypass for governing AI output in high-loss domains
A new governance proposal argues that “evaluate whether it’s correct” breaks down when oversight becomes structurally untenable at high AI output velocity. Flow-by-Flow instead reduces supervisory load using formal, countable feature scoring and institutional capacity caps, and it aims to avoid scalable content judgment while staying within human cognitive limits.
Scaling AI agents with trustworthy data
MIT Technology Review spotlights why agentic AI rollouts succeed or fail based on data quality, provenance, and feedback loops rather than just model size. The reporting centers on the infrastructure enterprises need to keep agents reliable – including how they curate datasets, measure drift, and create auditing paths when agents act on users’ behalf.
Dynamic Coalition Formation and Communication Pricing in skill-based agentic AI systems
This work models multi-agent communication as a coalition game where activating agents and sending messages costs tokens, latency, and error propagation. The researchers propose a marginal-value activation rule plus a greedy router – aiming to contact fewer, better-suited agents by estimating which connections are worth the price before and during execution.
NL2SHACL-Bench: A benchmark suite for Natural Language to SHACL translation
NL2SHACL-Bench closes a measurement gap for converting natural language constraints into SHACL shapes, a task that matters for validating RDF knowledge graphs. The benchmark shows today’s LLMs can produce syntactically valid SHACL but struggle to consistently generate semantically equivalent constraints for complex logical and structural patterns.
Of course the ChatGPT dog cancer vaccine spawned a startup
The Verge follows up on a viral story about AI-assisted cancer vaccine work for a dog – and reports that the entrepreneur behind it has launched Gamgee. The company’s pitch expands beyond pets into personalized mRNA cancer vaccines across species, using AI plus genetics as the core development engine.
From pilots to production: How enterprises put AI to work
OpenAI’s overview lays out how organizations are moving from experimentation to real deployments of agentic AI built on ChatGPT and Codex. The emphasis is on the systems around models – workflows, tooling integration, evaluation, and operational controls – that determine whether “AI assistance” becomes dependable execution.