Hermes Agent Adds Vercel AI Gateway, 200+ Models AccessAMD Acquires AI Inference Firm Taalas, Intensifying CompetitionDatabricks Cuts AI Coding Costs by 70%Claude Code v2.1.224: Self-Hosted Runners and Archived Plugin SupportNew Research: Co-Evolving Scaffolds and Parameters Improves Post-TrainingOpenAI Black Hat Talk Reveals Full Timeline of Hugging Face IncidentOracle Bans AI-Generated Code from OpenJDK Despite Ellison's ClaimWoodpecker Distillation: Weak Models Diagnose Reasoning Bugs in Strong ModelsQEvict: Recoverable Quantized KV Eviction for Long-Context DecodingPPDL: LLM Workflows as Probabilistic Programs, Making AI Confidence QuantifiablePRISM: Structured Multimodal Synthesis Helps AI Learn Instruction PrioritiesDisentangling 3D Modeling from Spatial Reasoning Improves Both TasksVercel Container Registry Now Supports Public RepositoriesVercel Audit Log Drains Adds Datadog, Splunk, and Panther IntegrationHSP GRUPPE Boosts Tax Advisory Productivity with ChatGPT EnterpriseDeveloper Recreates Raccoon Heist Game Using Codex and GPT-5.6 Sol UltraThe Tokenpocalypse: Companies Scramble to Cut AI SpendingAI Psychosis Emerges as New Leadership Blind SpotScalzi: Generative AI is 'Guitar Hero of Creativity' AI Billboards Overwhelm San Francisco Streets, Sparking BacklashHermes Agent Adds Vercel AI Gateway, 200+ Models Access

Hermes Agent now supports Vercel AI Gateway as its inference layer and can run agent commands in an isolated Vercel Sandbox microVM. Users can access 200+ models at no token markup, with all requests appearing in the AI Gateway dashboard for unified usage and spend tracking. Developers benefit from centralized multi-model management and enhanced security through isolated sandbox environments.

AMD has acquired Taalas, an AI inference company, for an undisclosed amount. The move intensifies competition in the AI inference space, positioning AMD to better rival NVIDIA. Specific product roadmaps and integration plans have not yet been disclosed.

Databricks reduced its AI coding spend by 70% through strategic optimizations, as detailed in an official blog post. Tactics include model selection refinement, improved prompt engineering, caching, and batch processing—all without sacrificing efficiency. This case demonstrates that careful management can yield substantial cost savings on AI tools, rather than simply relying on more expensive models.

A new arXiv paper introduces a post-training method that co-evolves model parameters with procedural scaffolds, unlike traditional approaches that design them independently. This strategy aims to internalize complex reasoning processes during training, reducing reliance on external scaffolds at inference time. Initial results suggest the co-optimization enables models to automatically acquire and internalize more complex skills, though further validation is needed.

OpenAI delivered a last-minute presentation at Black Hat security conference detailing the ‘Hugging Face Incident,’ with the video now available. The talk provides a complete timeline of what happened, marking OpenAI’s first public response to the security event. Essential viewing for developers concerned about AI security and enterprise deployment.

Oracle has banned AI-generated code from OpenJDK submissions, contradicting CEO Larry Ellison’s earlier claim that Oracle doesn’t write its own code. The policy sparked significant discussion (365 points, 246 comments on HN), potentially impacting Java ecosystem development workflows and AI coding tool adoption.

New arXiv paper introduces Woodpecker Distillation, showing that LLM reasoning failures often stem from localized bugs in intermediate steps rather than global incompetence. The method uses weak models to diagnose specific error locations in stronger models, offering a new approach to improving LLM reasoning.

New arXiv paper introduces QEvict, addressing KV cache memory constraints in autoregressive LLM inference. Unlike traditional attention-score-based token eviction, QEvict uses recoverable quantized eviction to handle attention drift in long-context decoding, reducing memory while maintaining performance.

PPDL is a framework that models LLM-based application flows as probabilistic programs, giving developers confidence metrics for each output. This approach uses probabilistic inference to capture output uncertainty, providing a new technical path for building reliable LLM applications. Developers no longer need to blindly trust model outputs but can rely on confidence data for decision-making and fault tolerance design.

PRISM is a method for training AI to recognize instruction priorities through structured multimodal data synthesis. Current multimodal training data often reduces complex instructions to single questions, while real-world scenarios bundle multiple requirements of unequal importance. PRISM employs a ‘rubric comprehension’ mechanism to help models distinguish which requirements matter more, improving complex instruction following.

This arXiv paper proposes explicitly disentangling 3D perception from spatial reasoning rather than jointly acquiring both through large-scale training. The study finds that when the two tasks are separated, models improve on each; joint training often causes one task to consume model capacity needed by the other. This paradigm offers new architectural directions for spatial AI models.

Vercel Container Registry now supports public repositories, allowing any Vercel user to pull and use images. Previously sharing was limited to 100 named teams; public access extends read-only access to all Vercel teams. Push and modification rights remain with owners. This simplifies cross-team image distribution for developers.

Vercel’s Audit Log Drains now stream team audit events to Datadog, Splunk, and Panther, joining existing HTTPS and S3 destinations. The feature forwards every Activity Log event plus additional metadata, giving developers more flexible security monitoring and log analysis options.

OpenAI published a case study showing how German tax advisory firm HSP GRUPPE integrated ChatGPT Enterprise into daily operations, boosting productivity and work quality while creating more capacity for client service. A practical example of enterprise AI adoption in professional services.

After demonstrating Claude Fable 5 building a complete game in one shot, Simon Willison challenged Codex + GPT-5.6 Sol Ultra with the same premise — a game idea he generated with GPT-3 and DALL-E four years ago. A head-to-head comparison of AI coding capabilities.

Simon Willison cites a 404 Media report with leaked Accenture meeting audio revealing that token costs are spiraling out of control. Companies are increasingly reevaluating AI spending and seeking strategies to reduce token consumption as the ‘Tokenpocalypse’ arrives.

Fast Company article warns that AI systems’ aberrant behavior — characterized as ‘AI psychosis’ — is becoming a blind spot for corporate leadership. With 159 points and 102 comments on HN, the topic resonates with tech professionals. The piece urges managers to acknowledge AI uncertainty and establish better monitoring and intervention mechanisms.

Sci-fi author John Scalzi published an opinion piece comparing generative AI to Guitar Hero — appearing to play but running on preset patterns. He argues current AI creation tools create an ‘illusion of creation,’ giving users engagement rather than real creative ability. The post sparked 63 Hacker News comments, representing a typical creator-community critique of generative AI.

SF Standard reports AI company billboards are flooding San Francisco streets, drawing criticism from residents who call them ‘dystopian’ and ‘unfunny.’ The topic gained traction on HN (32 points, 62 comments), reflecting divided opinions in the tech community as AI marketing budgets expand.