Notes by Jin

Models & Research KRAFTON releases a bilingual Korean and English speech model built on SK Telecom's A.X K2 backbone — A.X K2 Raon-Speech is a 21.2B-parameter mixture-of-experts speech language model with roughly 3.5B active parameters, pairing SK Telecom's A.X K2 Light text

Models & Research ByteDance Doubles Seedance Clip Length to 30 Seconds — ByteDance has launched Seedance 2.5, its latest audio-video generation model. Single-pass output goes from 15 to 30 seconds, with multi-round extension for multi-minute sequences. Reference input expands to 30 images, 10 video clips, and 10 audio clips per

Tools & Open Source DeepSeek-V4-Flash Puts Pro-Class Reasoning in a 13B Active Model — DeepSeek has published open weights for V4-Flash under an MIT license. The model is a mixture-of-experts design with 284B total parameters and 13B active, supporting a 1M token context. DeepSeek says the architecture pairs Compressed Sparse Attention

Models & Research OpenAI outlines plan for scaling capable, low-cost AI — OpenAI describes a broad strategy spanning infrastructure, models, platform, and products aimed at making powerful AI systems cheaper and more broadly accessible. OpenAI ↗ My takeaway: Check whether the system around your model is doing its job before you assume

Models & Research Microsoft Launches Seven MAI Models at Build 2026 and Pitches Customer-Owned Frontier Tuning — At Microsoft Build 2026, the Microsoft AI superintelligence team announced seven new MAI models covering image, voice, transcription, reasoning, and coding. The headliners are MAI-Thinking-1, the company's first reasoning model, and MAI-Code-1-Flash,

Models & Research Researchers argue LLM security flaws are structural and not fully fixable — A paper presented at ICML this month argues that large language models cannot be made fully secure. Chatbots tag text by role to track what came from the user, the system prompt, the model's

Models & Research Study probes how multi-turn planning ability develops in AI agents — Researchers examined how long-horizon planning ability emerges across pre-training and post-training stages. They used single-teacher and multi-teacher on-policy distillation inside a controlled synthetic environment. arXiv ↗ My takeaway: Pre-training data quality shapes long-horizon reliability more than the post-training

Models & Research Microsoft debuts MAI-Cyber-1-Flash, a security model built to cut defender token costs in half — Microsoft AI has announced MAI-Cyber-1-Flash, a compact cybersecurity model derived from its MAI-Thinking-1 lineage and integrated into MDASH, the company's multi-agent vulnerability identification and remediation harness. Microsoft says the model is

Models & Research Black Forest Labs Ships FLUX 3, a Multimodal Model Trained Jointly on Video, Images, and Audio — Black Forest Labs has put FLUX 3 into Early Access, a foundation model it says learns from images, video, and audio in one architecture. The video mode generates clips up to

Models & Research Google ships Gemini 3.6 Flash and 3.5 Flash-Lite, and restricts a new cyber model to governments and trusted partners — Google announced three Gemini models on July 21, 2026. 3.6 Flash is the new workhorse tier, priced below 3.5 Flash at $1.50 per

Models & Research Anthropic Ships Claude Opus 5 With Frontier Claims and Tighter Cyber Guardrails — Anthropic released Claude Opus 5 on July 24, 2026, and says it approaches the intelligence of its frontier Fable 5 tier at half the price. Pricing holds at $5 per million input tokens and $25

Models & Research Preprint maps when language models yield to social pressure on moral questions. — A preprint from University of Chicago and Yale researchers tested eight models across 78 moral dilemmas in three experiments. The authors argue that sycophancy is better understood as one surface of a broader belief-updating process,

Models & Research Upstage Ships Solar Open 2 Open Weights and Pitches It at Agent Workloads — Upstage released Solar Open 2 on July 22, 2026, an open-weight model the company says is built for agentic use. Per Upstage, it is a 250B-parameter mixture-of-experts design with 15B active per token, a

Models & Research Qwen-Image-3.0 is released — Alibaba's Qwen team released Qwen-Image-3.0 on July 21, 2026. Per Qwen, it accepts prompts up to 4,500 tokens, renders text as small as 10 pixels, and supports 12 languages and 20-plus fonts. The company pitches it for production work

Policy & Society Trump-aligned AI advisers clash publicly over Chinese models — Several current and former advisers to President Trump on AI have publicly criticized leading American AI companies over the weekend, exposing deep divisions within his AI policy circle about how to treat competition from Chinese models. MIT Tech Review

Models & Research Qwen3.8 Heads Toward Open-Weight Release — Alibaba's Qwen team says Qwen3.8 will launch soon and go open-weight, though it has announced no date and no license. The company describes it as a 2.4-trillion-parameter model that keeps evolving. Qwen has not published an active-parameter

Models & Research MobileWan, Qualcomm Runs a 5B-Parameter Video Diffusion Model on a Phone — Qualcomm AI Research introduces MobileWan, which the authors call the first 5B-parameter video diffusion model that runs on a commercial phone. Instead of shrinking the model, they shrink its runtime footprint from the server-scale Wan2.2-5B

Models & Research Study warns pretraining data may be poisoned through coordinated web manipulation — New research argues that large-scale pretraining datasets can be poisoned through coordinated influence campaigns, not just through edits to well-known sources like Wikipedia. The authors estimate that attacker-controlled text posted to public comment forms survives crawling

Models & Research Moonshot Says Kimi K3 Is the First Open 3T-Class Model, Concedes It Still Trails Top Proprietary Rivals — Moonshot AI released Kimi K3 on July 17, 2026. The company calls it a 2.8-trillion-parameter model and, by its own account, the first open model at that scale, with

Models & Research GPT-Red: OpenAI's Automated Red-Teamer for Prompt Injection Robustness — OpenAI describes GPT-Red, an internal-only model trained through self-play to attack its own systems and surface vulnerabilities before deployment. GPT-Red and defender models train together, with GPT-Red rewarded for landing a valid attack and defenders rewarded for

Models & Research Study asks whether AI agents know when a task is simple — A new preprint argues LLM agents often lack a quick sense of when a task is easy, and proposes a policy that estimates task scope before acting. Its headline efficiency gains come from a controlled simulator

Industry & Funding Nous Research reportedly nearing funding round valuing it at $1.5B — The maker of the open-source Hermes agent is said to be raising at least $75 million, with Robot Ventures leading and USV among the other backers. All figures come from three unnamed sources. Nous Research declined