Friday / AI Unicorns: Case for simulating 8 billion humans (Simile AI)

Sunday / Library: Frameworks Powering AI Agents

OpenAI launched ChatGPT for Academic Researchers, offering free access to GPT-5.6, Codex, ChatGPT Work, larger context windows, and more than 75 life-science tools.

The program starts with 10,000 researchers this summer and is expected to reach 100,000 scientists, mathematicians, and engineers through 2027. OpenAI has committed more than $250 million to external scientific research.

The European Union announced €10 billion in public funding for seven AI gigafactories, up from the five originally planned.

The Commission wants to attract another €20 billion from private investors. The facilities will combine processors, software, cloud infrastructure, networking, and data centers, supplementing 19 existing European AI factories.

AMD, NVIDIA, and Qualcomm have signed letters indicating their willingness to supply chips. Europe wants technological sovereignty and has concluded that sovereignty requires a procurement process.

Google introduced AI image generation inside Google Earth, allowing users to create synthetic scenes grounded in real satellite imagery and three-dimensional locations.

It paused the feature one day later after users produced content that appeared to violate its policies.

The images were watermarked and separated from Google Earth’s main experience. But synthetic events become considerably more persuasive when placed inside recognizable real locations.

MiniMax released H3, an open multimodal model that accepts text, images, video, and audio as shared context, then generates videos with native stereo sound.

H3 supports clips of up to 15 seconds at 2K resolution and can use one video as a motion reference, an image as the subject, and a separate audio file for sound or vocals.

MiniMax says 2K generation costs less than one-third as much per second as mainstream competitors and plans to release the weights, subject to regulatory restrictions.

Thinking Machines released Inkling-Small, an open-weight multimodal model with 276 billion total parameters and 12 billion active parameters.

It accepts text, images, and audio, supports a one-million-token context window, and is available under the Apache 2.0 license. A quantized version can run with 180GB of GPU memory, compared with at least 600GB for the full-precision checkpoint.

DeepSeek officially released V4-Flash, with stronger coding, tool-use, and agent capabilities and native support for the Responses API used by Codex-style systems.

It charges $0.14 per million input tokens and $0.28 per million output tokens. Artificial Analysis estimated an average cost of three cents per benchmark task, compared with $1.86 for GPT-5.6 Sol and $3.15 for Claude Fable 5.

Its performance is lower than the leading OpenAI and Anthropic models, but roughly comparable to Google Gemini 3.6 Flash. DeepSeek’s argument remains disciplined: capable enough, much cheaper.

Alibaba answered from the other direction with Qwen3.8-Max, its largest model yet, at 2.4 trillion parameters.

The mixture-of-experts model activates 95 billion parameters per request, supports text, images, video, and one million tokens of context, and became the highest-ranked Chinese text model on Arena.AI. It ranked second globally for visual understanding, behind a Claude Fable 5 variant.

Alibaba: one agent for every interface

Alibaba introduced Qwen-UI-Agent, designed to operate across phones, browsers, desktops, terminals, search engines, and APIs.

It can combine clicks, commands, and tool calls while preserving task state across environments. Alibaba trained it using thousands of sandboxes and more than 100 physical phones.

Microsoft: video models process only what changes

Microsoft released Mage-VL, a streaming multimodal model that uses motion vectors and residuals from compressed video instead of repeatedly processing complete frames.

Microsoft reports more than 75 percent fewer visual tokens and up to 3.5 times faster inference.

Mistral AI: moderation adapts to the policy

Mistral introduced Shieldstral, a three-billion-parameter multimodal safety model that evaluates content against natural-language rules.

One model can therefore support different products, jurisdictions, and definitions of harmful content without retraining a separate classifier.

MemTensor: memory moves inside the model

MemTensor released Metis in 4-billion, 9-billion, and 27-billion-parameter versions.

Metis maintains a persistent internal memory state that updates during inference without changing the model’s weights or replaying the full conversation.

Frontis AI: a model improves ML systems

Frontis AI released Frontis-MA1, a 35-billion-parameter model trained to draft, debug, improve, and combine machine-learning solutions.

Its proposals can be executed and evaluated inside a controlled environment, creating a measurable loop for iterative improvement.

CASIA: world models learn a physical language

PhiZero represents physical changes as discrete tokens before rendering them back into video.

The design separates dynamics from visual appearance, making predicted actions and transitions easier to transfer across objects and environments.

Huawei and HUST: robot control loses the LLM

TurboVLA is a 200-million-parameter vision-language-action model that predicts robot movements without routing every decision through a large language model.

It runs at 32 updates per second on an RTX 4090 while using less than 1GB of memory.

Agents begin improving the tools, skills, and rubrics around them

Evaluation becomes part of the training loop

Memory separates from both model weights and context windows

Retrieval becomes an execution policy, not just a ranking step

Evidence and provenance become persistent agent infrastructure

World models move from predicting actions to controlling them

Video generation gets faster without surrendering diversity

Open-ended research exposes scientific judgment as the remaining bottleneck

HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone
Shows that sufficiently accurate robot-free demonstrations may eliminate the real-robot post-training normally required before deployment. This could shift manipulation-data collection away from expensive robot fleets toward scalable human demonstration systems. →read the paper

🌟 Wonder: Video World Model Done Better
Constructs interactive video worlds that users can navigate, revisit, and explore over long horizons. Its combination of camera control, persistent memory, and real-time generation addresses the difference between producing a convincing clip and maintaining a coherent explorable world. →read the paper

ShadowDancer: Teaching Video World Models Any Action by Learning Unified Dynamics Representations from a Video and Its Shadow
Learns transferable actions from paired videos that preserve dynamics while changing visual appearance. This allows a demonstrated behavior to become a reusable control signal across new subjects and environments without requiring structured action labels. →read the paper

INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models
Learns a direct mapping from a desired visual change to the action needed to produce it. The approach removes the expensive test-time search usually required to turn a predictive world model into a practical controller. →read the paper

🌟 DistillAlign: Coordinating Mode Covering and Mode Seeking in Autoregressive Video Distillation
Explains why accelerated video generators can preserve visible quality while quietly losing diversity. It aligns mode-covering and mode-seeking objectives so that distillation retains more of the teacher model’s underlying distribution rather than collapsing toward a narrow set of attractive outputs. →read the paper

🌟 Parallel Decoding Distillation for Fast Image and Video Generation
Accelerates diffusion and flow models by predicting several denoising transitions within each model evaluation. It offers a simpler alternative to adversarial and variational distillation methods, whose training instability frequently produces motion loss and mode collapse. →read the paper

DecoEvo: Score-Decoupled Co-Evolution of Solver and Rubric-Generator Skills in Text Space
Co-evolves the instructions used to solve a task and the rubric used to evaluate the solution, while preventing improvement from being faked by quietly making the rubric easier. It provides a more credible mechanism for inspectable self-improvement in open-ended tasks. →read the paper

CoRT: Counterfactual Replay for Token-Level Rubric-Guided Policy Optimization
Converts rubric feedback into token-level credit by comparing the same response with and without the rubric context. This lets reinforcement learning identify which parts of an answer actually satisfied the criteria without training another token-scoring model. →read the paper

SkillRise: Agentic Reinforcement Learning for Cross-Task Skill Evolution
Trains a single policy to solve related tasks while continuously updating a reusable skill document. The important shift is from learning separately inside each episode to accumulating procedures that improve performance on later, different tasks. →read the paper

🌟 GPT-Red: Automated Red Teaming via Self-Play at Scale
Scales prompt-injection discovery through attacker-defender self-play and uses the resulting attacks to adversarially train stronger models. It turns red teaming from a periodic external audit into a continuously escalating training loop. →read the paper

🌟 Can AI Agents Conduct Open-Ended AI Research? Early Evidence from Two Case Studies
Tests autonomous agents on genuinely open-ended research rather than narrowly verifiable tasks. It finds that agents can complete much of the engineering work but still struggle with scientific judgment, backtracking, creativity, and deciding what constitutes a meaningful contribution. →read the paper

🌟 AskChem: Claim-Centered Infrastructure for Chemistry Literature Synthesis
Changes the unit of scientific retrieval from entire papers to atomic claims carrying quotations, evidence locations, and provenance. This gives research agents a more reliable substrate for synthesizing findings across publications without losing the connection to source evidence. →read the paper

CodeNib: A Multi-View Data System for Serving Repository Context to Coding Agents
Treats repository context as maintained infrastructure rather than something every coding agent must rediscover from scratch. It combines lexical, semantic, and structural views that persist across code changes and can be served within bounded context budgets. →read the paper

🌟A New Role for Relevance: Guiding Corpus Interaction in Agentic Search
Uses relevance scores to determine where an agent begins searching, which documents it explores, and which local evidence remains visible. It turns relevance from a one-time ranking signal into an execution policy for navigating large corpora. →read the paper

Memory Decoder at Scale: A Pretrained, Parametric Long-Term Memory
Separates memory capacity from the reasoning backbone and scales a dedicated parametric memory module independently. The results suggest that adding parameters to memory can be more efficient than making the complete language model larger. →read the paper

LEDGERMIND: Provenance-Constrained Multimodal Agentic Reasoning with a Structured Evidence Ledger
Constrains an agent’s reasoning to evidence explicitly recorded by its tools. Every downstream claim must cite an active ledger entry, while repair operations are prevented from introducing unsupported content, making provenance part of the agent’s execution state rather than an afterthought. →read the paper