Friday / New: I’m pondering an idea for a new type of media that I’d like to share with you.

Sunday / Library: Chinese LLMs in 2026: DeepSeek, Qwen3, Kimi K2 and More

First, effects should be reversible: a component records how its changes can be withdrawn.

Second, dependencies should be reactive: if a required service disappears, dependent components deactivate before it is removed.

Why I stand against AI watermarking by Guillaume Meyer

One more Dario Amodei’s attempt to explain what he means

Agentic profiles for effective AI governance by Atoosa Kasirzadeh and Iason Gabriel

OpenAI previewed Ultrafast, a Cerebras-powered tier that runs GPT-5.6 Sol up to 14 times faster and reaches 750 output tokens per second. Access remains limited; latency has joined intelligence as a premium feature. (OpenAI)

Anthropic said future Claude models will watermark generated text using a version of Google DeepMind’s SynthID-Text, with no hidden characters or identifying data. The EU AI Act has reached word selection. (Anthropic)

NVIDIA released Nemotron 3.5 Lightning, a 30B open Mixture-of-Experts model for high-volume agents, alongside NeMo Switchyard, an open router that selects models by quality, latency, and cost. NVIDIA says Switchyard can preserve near-frontier accuracy at roughly one-third the cost of using Opus 4.8 alone. (NVIDIA Blog)

Six days later, NVIDIA offered up to $105B in guarantees for OpenAI’s 8GW Ohio campus, invested $1.5B in its developer, and became the site’s exclusive chip supplier. NVIDIA now routes the workloads and helps finance the buildings. (Jensen’s X)

Google released Gemini 3.7 Flash for coding and agents at an introductory $0.75 per million input tokens and $3.75 per million output tokens, only three weeks after 3.6 Flash. (blog.google)

DeepMind also put its multilingual SL2T sign-language model into Gboard and Live Transcribe on Pixel 11, beginning with American Sign Language to English. (Google DeepMind)

SpaceXAI launched Grok Bot, persistent agents with their own cloud computers that can work across apps, inboxes, and websites. It followed with Grok 4.6, built for long-running coding and knowledge work, which shipped first in Cursor and then entered GitHub Copilot. Owning the editor does not prevent distribution through the incumbent. (SpaceXAI)

Mistral launched regional inference in Europe and the U.S., an SLA-backed Priority Tier, and support for third-party open models beginning with GLM-5.2.

It is also asking European customers to make multi-year commitments that could support up to 1GW of Mistral-operated compute by 2030. Sovereign AI now comes with data residency, reserved capacity, and advance purchasing. (Mistral AI)

DeepSeek released V4 Pro and raised V4 API prices by between 50% and 1,100%, depending on the model, token type, and time of use.

At peak rates, V4 Pro costs $1.32 per million input tokens and $3.96 per million output tokens. DeepSeek remains inexpensive beside U.S. frontier models, but the race to the bottom has acquired surge pricing. (DeepSeek)

Z.ai released GLM-5.3 and reported an 84.5% score on CyberGym, slightly above Mythos 5 for vulnerability discovery, while trailing it sharply on exploit development.

The company delayed the open-weight release for two weeks and will restrict its most sensitive cyber functions to verified users. These are company-reported benchmark results, but the policy shift is real: Chinese open-model labs are adopting access controls associated with closed American systems. (Z.ai)

IBM signed a $240M agreement with Together AI for a 2,000-chip NVIDIA B300 cluster dedicated to serving open models on IBM Cloud.

Two days later, IBM partnered with OpenAI to embed GPT-5.6, Codex, and ChatGPT Work into IBM Consulting. IBM chose both sides, which is often the correct enterprise answer. (Reuters)

Bloomberg reported that Stripe finalized an agreement to acquire OpenRouter for more than $7B, only months after the model-routing company raised money at a reported $1.3B valuation. The final price could still change.

The company that routes payments is buying the company that routes models. (Bloomberg Law)

Google: Gemini 3.7 Flash becomes the new workhorse

Google released Gemini 3.7 Flash on August 13 for coding, agents, complex documents, and business automation. It improves substantially over 3.6 Flash and launched at half that model’s original price per million tokens. →read the announcement

Alibaba: Qwen3.8 opens at both ends of scale

Alibaba released the open weights for Qwen3.8-2.4T-A95B on August 12 and Qwen3.8-27B on August 14. The first brings a Qwen-Max-class model into the open; the second packages native multimodality, controllable reasoning, and agent execution into a dense model that is far easier to deploy. →view the models

DeepSeek: V4-Pro becomes the production model

DeepSeek released V4-Pro on August 13 with stronger agent execution, configurable reasoning effort, and native support for the OpenAI Responses API used by Codex-style systems. It is available through DeepSeek’s app, web interface, and API. →read the announcement

xAI: Grok 4.6 targets long-running agents

xAI released Grok 4.6 on August 12 with a focus on sustained research, repository-scale coding, and interactive visual work. The model was trained across agentic environments and shows more self-testing and verification during extended trajectories. →read the announcement

Z.ai: GLM-5.3 combines coding with unexpected cyber capability

Z.ai announced GLM-5.3 on August 14 as a major coding and agent upgrade produced largely through expanded post-training. Its cybersecurity capabilities developed faster than expected, leading Z.ai to delay the open-weight release for approximately two weeks while it completes safety testing and hardening. →read the announcement

Meta: Muse Glimmer brings agents onto one GPU

Meta released Muse Glimmer on August 10, a 30-billion-parameter open-weight model optimized for local coding, tool use, multimodal work, and always-on agents. A quantized version fits within roughly 20GB, allowing the complete agent model to run on a single consumer GPU. →read the announcement

NVIDIA: Nemotron 3.5 Lightning becomes the execution model

NVIDIA released Nemotron 3.5 Lightning on August 11, a sparse 30-billion-parameter model that activates three billion parameters per token. It is designed as a fast execution layer for long-running agents, with weights, training data, and recipes released openly. →read the announcement

OpenAI: GPT-5.6-Cyber becomes a restricted specialist

OpenAI released GPT-5.6-Cyber on August 10 through its controlled Daybreak Red program. Built on GPT-5.6 Sol, it is specialized for zero-day discovery, exploit validation, and other advanced defensive-security work that general-purpose models frequently refuse. →read the announcement

Google DeepMind: sign language reaches a production translation model

Google DeepMind introduced its multilingual sign-language-to-text model on August 12. It now powers sign-to-text dictation in Gboard and Live Transcribe on Pixel 11, beginning with American Sign Language to English, moving sign-language AI from a research demonstration into a consumer product. →read the announcement

LTX: open video becomes production-ready

LTX released LTX-2.5 on August 11 with native multishot generation, persistent continuity across cuts, a diffusion video decoder, cinema-grade EXR support, and a raw pretrained checkpoint for deeper customization. The release pushes an open video model toward professional production workflows rather than isolated short clips. →view the release

MiniMax: Music 3.0 generates the complete song

MiniMax released the open weights for Music 3.0 on August 13. Given a concept and optional lyrics, it composes, arranges, performs, and produces a song lasting up to five minutes within one generation, using separate global and local models for musical structure and acoustic detail. →read the announcement

Liquid AI: vision and tool use move to the edge

Liquid AI released LFM2.5-VL-3B on August 12, an open-weight vision-language model for screens, documents, grounding, multiple images, and function calling. It is designed for low-latency local deployment and can run within approximately 3GB of memory. →read the announcement

Cohere: native-resolution vision shrinks to 2.4 billion parameters

Cohere Labs released North Micro Vision on August 12 under Apache 2.0. The model preserves the original resolution and aspect ratio of documents, charts, forms, and screenshots, providing a compact foundation for specialized visual systems. →read the announcement

Upstage: Solar Pro 4 focuses on finished enterprise work

Upstage released Solar Pro 4 on August 11 with a 512K context window, up to 128K output tokens, and support for English, Korean, and Japanese. It is trained for long documents, terminal work, tool use, and multi-stage assignments where values must remain consistent across spreadsheets, reports, and slides. →read the announcement

Writer: Palmyra X6 optimizes cost per completed task

Writer released Palmyra X6 on August 13 as its new enterprise flagship, post-trained from GLM-5.2 for grounding, tool use, delegation, and long-running marketing and revenue workflows. It is designed specifically for Writer’s agent harness and can work toward one objective for up to eight hours. →read the announcement

Sakana AI: orchestration becomes a model

Sakana AI released Fugu on August 13 as an orchestrator for complex, multi-step work and upgraded Namazu to a new generation with stronger Japanese and agent capabilities. The models are integrated with sandboxed code execution and can produce complete files and visual artifacts inside Sakana Chat. →read the announcement

Motif Technologies: extreme sparsity becomes the architecture

Motif 3, published August 10, contains 314 billion total parameters while activating 13.2 billion per token through eight of 384 experts. Its post-training consolidates several specialist teachers through on-policy distillation, making fine-grained sparsity and capability merging the central model-design bet. →read the paper

Pathway: reasoning stays latent

BDH-CQ, published August 10, is a 150-million-parameter recurrent model whose internal memory updates from examples supplied during inference. It then solves problems through iterative latent computation without producing a textual chain of thought, offering a compact alternative to long verbal reasoning traces. →read the paper

Luna-TTS: speech generation becomes parallel

Luna-TTS, published August 12, is a 0.6-billion-parameter diffusion-based speech family trained across Chinese, English, Japanese, and Korean. One version generates the complete acoustic-token grid in parallel, while the real-time version works block by block for streaming speech, voice cloning, editing, and emotional control. →read the paper

Alibaba and LiblibAI: human animation becomes continuous

LiveAnimate, published August 12, converts a 14-billion-parameter video diffusion model into a causal streaming generator. It retrieves earlier pose-relevant context from a bounded cache, allowing real-time human animation to continue for several minutes without memory use growing with the video. →read the paper

Shanghai AI Lab

DreamX and Alaya: world models become controllable and persistent

DreamX-Phi, published August 13, predicts robotic futures from a prescribed trajectory and adds geometric, depth, and object-level constraints so that visually convincing rollouts still follow the commanded action. Alaya-EVOKE stores persistent scene geometry outside the active context, retrieving only the state required for the current camera view as a generated world expands. →read DreamX-Phi →read Alaya-EVOKE

DFM: permissible data becomes a model constraint

Mimir v1, published August 13, is an open one-billion-parameter hierarchical reasoning model trained from scratch with permissible post-training datasets. Its importance is the attempt to treat data provenance as part of the model architecture and release strategy rather than a disclosure added afterward. →read the paper

Agents move their continuity into harnesses, worlds, and checkpoints

AI science shifts from writing papers to running recoverable experiments

Safety moves to runtime contracts and internal interventions

World models stop hiding state and dynamics inside pixels

Computation becomes selective, parallel, and structurally cheaper

🌟 AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses
Transfers capabilities during inference by having a stronger model construct deterministic scaffolds for a weaker one, nearly doubling average performance without changing the weaker model’s weights. →read the paper

🌟 Persistent Recursive Worlds Enable Autonomous Software Evolution
Makes the software project and its accepted history persistent while finite-lived agents recursively delegate and propose changes, moving continuity from the agent into the environment. →read the paper

🌟 AgentRewind: Recoverable Execution for Long-Horizon LLM Agents
Records aligned checkpoints of model context and environment state so agents can roll back, preserve lessons from failed attempts, and resume from recoverable points. →read the paper

🌟 Demystifying Agent Skills: Why They Work-Until They Don’t
Shows that skills mainly stabilize procedures rather than add knowledge, while retrieval precision falls and brittle assumptions accumulate as skill libraries grow. →read the paper

🌟Envs-FORGE: Frontier-Optimized Reward-Grounded Environment Synthesis for Agent RL
Uses verifier rewards to rewrite tasks, tests, fixtures, and containers together, then admits only executable environments that pass validation into agent reinforcement learning. →read the paper

Handover of In-Context Learning State Across Session Boundaries
Formalizes what a session handover must preserve, separating exact recovery of earlier material from preservation of the continuation task’s predictive behavior. →read the paper

AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design
Uses a meta-harness to revise a design agent’s workflow from rollout feedback, allowing the scaffold around the model to become an optimization target. →read the paper

Spark-to-Paper: End-to-End Research Paper Generation as a Composable Skill
Builds research as a set of composable skills that retrieve literature, run experiments, revise claims against evidence, generate editable figures, and apply deterministic integrity checks. →read the paper

Mechanist: AI as a Scientific Instrument for Discovering the Mechanisms of Intelligence
Automates mechanistic discovery by generating hypotheses, running causal interventions, and turning resulting explanations into model-control experiments. →read the paper

OmniScientist: An Omni-Modal Omni-Discipline AI Scientist
Reasons over raw multimodal scientific evidence while enforcing novelty, statistical validity, provenance, and numerical traceability in executable code. →read the paper

ScienceFlow: A Long-Horizon Agent for ML Research, Scientific Discovery and Beyond
Organizes autoresearch into recoverable executable states, re-anchors work from live or archived progress, and allocates compute according to validated evidence. →read the paper

🌟 Stealing Reasoning Traces from Proprietary LLM APIs
Shows that encrypted reasoning objects can be replayed across sessions, users, and models to recover hidden traces, transport private information, or carry invisible instructions. →read the paper

🌟 Agent Safety Should Be a Runtime Contract
Reframes agent safety around trajectory-level prevention and checkable evidence, with harnesses blocking dangerous actions and proving that required actions occurred. →read the paper

ToolHazard: Scaling Adversarial Environments for Security Evaluation and Alignment of LLM-Based Agents
Synthesizes stateful tool environments and environment-specific indirect prompt injections, then reuses the generated attacks as security-alignment data. →read the paper

Tripwire: Triggering Aligned Refusal via Statistically Certified Safety Neurons
Identifies safety-specific neurons statistically and clamps them only on harmful inputs, reducing jailbreak success without continuously perturbing benign requests. →read the paper

Twin: Playing an Unknown Game with a Test-Time Digital Twin
Has a coding agent construct and continually repair an executable world model at test time, requiring it to reproduce observed transitions before relying on it. →read the paper

Learning How the World Evolves: Extrapolative Video World Models via Latent Dynamics Reasoning
Encodes kinematic integration explicitly and learns only higher-order residual dynamics, improving extrapolation beyond trajectories represented during training. →read the paper

Marionette: Predicting World States, Rendering Geometry, Painting Appearance
Separates a compact 3D world state, a fixed geometric renderer, and a diffusion appearance model so structural errors can be repaired directly in state. →read the paper

RynnValue: Scaling Robotic Value Foundation Models with Temporal Distance
Turns temporal distance in large-scale robot video into a general value signal, converting weak timestamp supervision into dense rewards for robotic control. →read the paper

Claim-Level Reliability Assessment for Efficient Test-Time Reasoning
Verifies decision-critical claims instead of sampling more complete solutions, using semantic falsification to locate decisive errors with fewer tokens. →read the paper

Second Thought: Reasoning in Parallel as LLM Agents Act and Observe
Fills the idle time between agent actions and observations with parallel reasoning branches, reducing sequential decoding along the task’s critical path. →read the paper

Multimodal Model Diffing for Feature Discovery and Control
Aligns and compares sparse autoencoders before and after multimodal training to isolate newly learned features that can be causally removed or steered. →read the paper

Scaling Domain Data Repetition in LLM Pretraining
Shows that optimal repetition of scarce domain data increases mildly with model size and tracks domain validation loss better than the amount of unique data. →read the paper

Dion3: Full-Stack Orthogonal Updates
Reworks orthogonalization, kernels, and distributed communication to cut optimizer-step time by as much as sixfold without worsening training loss. →read the paper