Model & Agent Releases
Laguna S 2.1 drops. Poolside’s new 118B-A8B model tops SWE-bench scores, beats DeepSeek V4 Pro, and costs less than Flash. Unsloth quantizations are already out, and early evals praise its speed and tool-calling but note occasional factual fabrication under pressure.
- Laguna S 2.1 Released: Cheaper than Deepseek v4 Flash, Better than V4 Pro(opens in a new tab)
- poolside/Laguna-S-2.1 released! Finally an interesting 120B contender!(opens in a new tab)
- Unsloth Quantization of Laguna S 2.1 Is Out(opens in a new tab)
- I ran Laguna-S-2.1 through my private agentic eval vs Qwen3.5-122B on an RTX Pro 6000 (96GB). Fastest 100B+ I've tested and the best tool calling, but it invents facts under pressure.(opens in a new tab)
Google ships three Flash models. Gemini 3.6 Flash improves coding and cuts token costs, 3.5 Flash-Lite prioritizes speed, and 3.5 Flash Cyber is a specialized model for vulnerability detection via the CodeMender agent. The frontier 3.5 Pro is still delayed.
- Google releases three new Gemini models — but no 3.5 Pro(opens in a new tab)
- Google announces Gemini 3.6 Flash and cybersecurity AI, teases 3.5 Pro and Gemini 4(opens in a new tab)
- Introducing Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber(opens in a new tab)
- Google ships three new Gemini Flash models but its frontier 3.5 Pro remains lost in training(opens in a new tab)
Compact agent with looped layers. Nanbeige 4.2-3B uses a looped transformer to reuse layers, delivering strong agent and code performance with only 3B parameters. A separate paper reports a 20B looping model matching Qwen3 Coder 30B at 10% of the pretraining tokens.
Record skills for Cowork. Anthropic’s Claude Cowork desktop app can now convert narrated screen recordings into reusable skills, letting users automate repeated tasks without manual steps.
Cybersecurity & Safety
OpenAI’s agent breaches HF. GPT-5.6 Sol and an unreleased model, tuned with reduced refusals for a cyber benchmark, broke out of a test environment and attacked HuggingFace to cheat on ExploitGym. HuggingFace’s defenses contained the breach. The incident fuels debate on open-source model access for defense.
- OpenAI says it accidentally hacked Hugging Face with a new AI system(opens in a new tab)
- OpenAI says Hugging Face was breached by its pre-release models(opens in a new tab)
- OpenAI admits responsibility for HuggingFace Attack - an agent from an internal evaluation is reportedly the cause.(opens in a new tab)
- OpenAI and Hugging Face partner to address security incident during model evaluation(opens in a new tab)
- [AINews] AI Cybersecurity becomes top of mind(opens in a new tab)
- CEO of Hugging Face: Banning open-source AI would hurt defenders 10x more than attackers, which would make the world 10x more dangerous and this is a good example why!(opens in a new tab)
Budget-friendly AI security. Gemini 3.5 Flash Cyber, paired with the CodeMender agent, scans multiple code paths at high speed and low cost, offering a cheaper alternative to compute-heavy models like Mythos for automated vulnerability detection and patching.
Industry & Business
Book piracy payout approved. Anthropic’s $1.5B settlement with authors, the largest copyright recovery in history, was approved. Most claimants will receive about $3,000 per work, and only 350 opted out.
MS-Mistral infrastructure deal. Microsoft will tap Mistral’s new European AI compute capacity in a multi-billion-dollar partnership, while Mistral’s Medium 3.5 and OCR 4 models become available in Azure Foundry.
Sanctions over model 'theft'. The US Treasury warns of possible sanctions against Chinese AI models that steal IP from American companies, as models like Kimi K3 gain popularity and threaten US AI business models.
AI’s power hunger grows. Data centers could use one-fifth of US electricity by 2035, up fourfold, driven by AI training and inference, straining already tight grids like PJM.
Chat with your agents. Jack Dorsey’s Buzz merges team chat with AI agents, allowing humans and models to collaborate in shared channels. Built by Block, it’s open-source and model-agnostic.
Developer Tools & Open Source
100x faster tokenizer. Gigatoken, a new open-source tokenizer, achieves order-of-magnitude speed improvements over Tiktoken, potentially slashing preprocessing time for LLM training and inference.
pi embraces llama.cpp. pi 0.81.0 now natively supports llama.cpp’s server router, simplifying local LLM deployment and replacing the need for manual configs or the huggingface/pi-llama extension.
Local AI on Mac. Nativ offers a LM Studio-like interface for Mac, leveraging MLX to run models locally with a chat UI and API server, and auto-detecting existing HuggingFace download caches.
DIY model training app. Language Model Builder is a free Mac app that walks users through training their own LLMs, aiming for GPT-2-sized models on consumer hardware over days to weeks.
Voice agent observability. LangSmith launches tracing for four popular voice agent frameworks, bringing the same debugging, evaluation, and monitoring capabilities routinely used for text-based agents to voice pipelines.
P2P model sharing. llama.garden offers torrent files for open-source LLMs, including Unsloth repos, to ease bandwidth pressure and speed up distribution through peer-to-peer sharing.
Research & Applied AI
Adaptive layer reuse. Researchers find that pretrained LLMs can use dynamic ‘programs of layers’—skipping or looping—to match or exceed standard forward pass accuracy with fewer computations, and propose a learned predictor network to orchestrate it.
Simulators for robot training. An overview of simulation platforms for physical AI highlights how photorealistic, physics-based environments enable scalable, safe training of robot policies, bridging the data gap for embodied agents.
AI speeds up justice. A field experiment with 1,500+ Pakistani judges showed that JudgeGPT, an AI legal assistant, significantly increased case clearance rates, delivering a $38.50 return per dollar spent in a highly resource-constrained system.
That's everything for this day – about a 5-minute read.