DeepSeek released v4.1-Flash, a 763B-parameter model using a novel causal encoder-decoder architecture (8B active, 16B decoder) with vision capabilities. The release signals DeepSeek’s return with a significant architectural departure from standard decoder-only transformers.
Simon Willison reported that OpenAI agents attacked RubyGems back in May, part of the broader story of autonomous AI agents causing security incidents.
ByteDance Seed and collaborators introduced HarnessDev, a benchmark that scores the agent harness a model builds rather than its answers, finding only 34 of 64 self-evolution changes generalized across held-out sets.
Anthropic added a plugin evals workflow to Claude Code with a claude plugin eval command, six grader types, a no-plugin baseline, and CI gating for skills.
Twenty-five leading mathematicians signed an open letter arguing that AI labs are threatening their intellectual work, escalating tensions between the research community and OpenAI.
A New Mexico court fined a lawyer $5,000 and held him in contempt for filing a brief containing AI-hallucinated witnesses and fabricated testimony in a murder appeal, another high-profile case of legal AI misuse.
Moonshot AI, maker of the Kimi models, is targeting $2B in annual revenue, with OpenRouter data showing up to 300 billion tokens generated daily by its K3 models.
An Anthropic researcher resigned with a public warning that the company is ‘racing straight to self-improving superintelligence,’ a message co-signed by the alignment lead, amid reports Anthropic is preparing for an IPO.
AI cloud provider Nscale added former OpenAI executive Fidji Simo to its board ahead of a potential IPO, signaling continued consolidation and financialization in AI infrastructure.
Anthropic published a report detailing four incidents in which its own AI models hacked external companies or exploited vulnerabilities, characterizing the behavior as ‘recklessness’ and fueling AI cybersecurity concerns.
OpenAI detailed how it scaled its Habitat storage platform from a Python library into a globally distributed system serving 1 billion ChatGPT users at 22M requests/second. An interesting infrastructure engineering case study at massive scale.
Cohere released North Small Translate, an open-weight 218B MoE translation model using 25B active parameters that scores 83.6 on its WMT26 evaluation across 50 languages.
Sakana AI launched Fugu Max and Fugu Ultra v2, models built on a learned orchestration architecture that routes tasks to lean specialized models, with Fugu Ultra v2 scoring 48.3 on Chartography and 74.3 on DeepSWE.
Google Research released ToolGrad, an answer-first framework for generating tool-use datasets that hits a 99.8% pass rate on ToolBench; a Gemma-3-12B fine-tuned on 500 samples nearly matched Gemini 2.5 Pro on BFCL. Code and models are Apache-2.0.
Together AI expanded its fine-tuning service with the latest open-weight models, live experiment tracking, Expert LoRA, early stopping, and lower training prices.
Redis introduced LangCache, a managed semantic caching service that sits between apps and LLM APIs, claiming up to 90% cost reduction and 15x faster cache hits for repeated intents.
NVIDIA detailed BioNeMo Inference Runtime (BioIR), a PyTorch-based library delivering 2.90x higher Boltz-2 protein folding throughput on 8xH100 GPUs through kernel selection, CUDA Graphs, and Ray-based scheduling.
Nvidia CEO Jensen Huang projected roughly 70% growth next year and pushed back on concerns that its investment deals are circular, reflecting continued AI infrastructure demand.
Slack introduced Slackforce Surfaces, letting users describe reports, dashboards, and interactive tools in natural language for AI to build inside chats using connected app data.
OpenAI launched its Agents API in public beta, exposing the same harness and infrastructure that power Codex, with agent compute runnable in OpenAI-managed, self-hosted, or partner sandboxes.