Reasoning AI

OpenAI says it disrupted a coordinated effort to extract protected model reasoning, highlighting new risks for model distillation and AI defenses.

PrismML released Bonsai 2 27B, a compressed reasoning model designed for PCs and phones, strengthening the case for private, on-device AI.

OpenAI introduces GPT-6 Astra, a business-focused model with reasoning, computer use, and stronger writing and design judgment, but few details are public.

Salesforce has introduced Koa, a reasoning model for Agentforce trained on Nvidia’s Nemotron, pointing to more controlled AI for CRM work in enterprise deployments.

Salesforce and Nvidia announced Koa, an enterprise reasoning model for Agentforce designed to lower costs, protect data, and handle business workflows.

TechCrunch’s living AI glossary explains fast-moving terms from opaque recurrence to AI agents, helping builders and buyers assess new claims.

NVIDIA’s Jetson deployment guide shows how quantization and speculative decoding can bring compact reasoning models to local edge AI workloads.

Reports say OpenAI’s Astra may conceal parts of its reasoning, raising questions about safety monitoring, auditability and deployment risk for AI builders.

A report says API weaknesses at OpenAI, Anthropic, and Google may expose stronger models’ reasoning to weaker systems, raising security questions.

MIT Technology Review argues that AI agents, rather than data-hungry models alone, could make scientific research faster, more reproducible, and broadly useful.

Thinking Machines has released Inkling Small, an Apache 2.0 open-weights reasoning model that approaches Inkling’s results with far fewer active parameters.

SK Telecom introduced a second A.X K2 Korean-language model, signaling a deeper push into domestic enterprise AI with stronger reasoning claims.

Anthropic says Claude Mythos found flaws in HAWK and reduced-round AES, signaling AI may accelerate cryptanalysis without threatening current systems.

A reported chain-of-thought spoofing attack highlights a new security risk for reasoning AI models, raising reliability concerns for AI builders and buyers.

OpenAI unveiled GeneBench-Pro, a genomics benchmark meant to measure higher-order scientific reasoning as AI labs push into biology workflows.

Google releases Gemini 3.1 Pro achieving 77.1% on ARC-AGI-2 benchmark, doubling previous model's reasoning capabilities for complex problem-solving tasks.