Meet Neural Theorizer (NEO) a new type of world model that learns to theorize the world from observation, without language or LLM supervision.

Google Introduced TabFM, a foundation model designed specifically for tabular data classification and regression.

This approach allows generation of high-quality predictions on previously unseen tables in a single forward pass.

Huawei released open-sources OpenPangu-2.0-Flash

#OpenPangu 2.0 includes two 512K-context models:

1. Flash: 92B total,6B active Weights+inference code+training ops released

2. Pro: 505B total,18B active - flagship model, coming soon.

it's currently one of the most complete open-source MoE models running on the Ascend ecosystem.

Features like mHC, Muon and MTP all frontier techniques introduced in late 2025 have successfully landed on CANN.

That's an impressive engineering achievement.

But compared with DeepSeek V4, the gap is still roughly one to one-and-a-half generations.

Meet LeVLJEPA: the first fully non-contrastive end-to-end vision-language pretraining method competitive with CLIP and SigLIP

No negatives. No temperature. No momentum encoder. No teacher-student.

LeVLJEPA learns image to text structure by prediction: each modality predicts the other's embedding, while SIGReg keeps each embedding isotropic Gaussian.

Code
Model

Bytedance introduced EdgeBench, a benchmark designed to study how agents learn from environments over at least 12~72-hour runs.

EdgeBench is built with three ingredients:

1. Real & Diverse: 134 real-world tasks across 6 task categories, spanning scientific problems, professional knowledge work, software engineering, optimization, formal math, and games.

2. Ultra-Long-Horizon: Each task supports 12–72 hours of agent work. Recorded human effort averages 57.2 hours.

3. Informative Feedback: Agents receive real-world feedback for continuous improvement.

After 38,000 hours of agent runs on EdgeBench, a scaling law for learning from environments emerges:

- As agents interact with task environments over time, their aggregate performance is precisely fit by a log-sigmoid function.

- This phenomenon can be explained by an elegant theory of graph exploration.

Google introduced Paper Assistant Tool (PAT)

It's an AI framework built to review full scientific manuscripts, check theoretical results, validate experiments, and flag flaws.

PAT uses inference scaling to dig deeper than a single model call, achieving a 34% improvement in recall on mathematical errors (SPOT benchmark).

Already piloted at STOC and ICML, it catches critical errors early, easing referees' burden while keeping human oversight.

Meituan introduced LongCat-2.0 a fully open-source MIT licensed, no restrictions.

The model weights and inference code to everyone.

1. 1.6T MoE · ~ 48B active · 1M token context

2. Agent-native: Integrates directly with Claude Code, OpenClaw, and Hermes Agent

3. Deployment: Support both GPU and NPU platforms verified on large-scale domestic clusters

Tencent just dropped Hy3

- 295B total params, 21B active. Compare to DeepSeek/Qwen at ~1T, overseas frontier at ~10T.

- API: ¥1 / ¥4 / ¥0.25 per M tokens. Cheapest Chinese model on the market. Cheaper than DeepSeek. ~7× cheaper than GLM-5.2.

- Token efficiency vs GLM-5.2 on office tasks: docs -47%, PPT -49%.

- Hallucination rate 12.5% → 5.4%. MRCR long-context 42.9% → 75.1%.

- SWE-Bench variance across 3 scaffolds (Codebuddy/Cline/KiloCode) ≤4 pp.

- 270-expert blind eval: Hy3 2.67/4 vs GLM5.1 2.51/4.

FREE API for 2 weeks

"Striking distance of frontier" is the wrong frame. At 21B active params, Hy3 isn't trying to be frontier. It's showing the frontier isn't where the value is.

The actual thesis buried in the announcement: scaling is moving from pretraining to RL + product feedback loops. WorkBuddy (Tencent's white-collar agent, #1 in workflow by user base) is the data flywheel. Pretraining scale plateaued. The new bottleneck is real-world task environments.

If that's right, the small-model-cheap-API play isn't a price war. It's the new architecture

Last day of Fable.

Here are some prompts to run that require maximum intelligence.

DoorDash released DashBench, an internal benchmark to compare multi-model/harness code review systems against each other.

DashBench has given confidence in maximizing the intelligence per dollar they’re spending on AI tokens to ship code safely.

Anthropic extended access to Claude Fable 5 on all paid plans through July 12.

Rightnow AI introduced The AGI compiler

it watches your LLM agent work, finds the parts that are secretly deterministic, and compiles them into verified binaries that cost nothing to run

LLMs are just the first frontend, world models and new model types plug into the same toolchain, that's why it's a compiler.

AGI compiler means one narrow testable thing, a system that turns novel experience into permanent verified near free skill while measuring what it does not know.

Researchers did not build agi, they built the loop anything deserving the name must contain.

36,843 lines of rust, all open.

GitHub.

same 300 tasks, same verified answers, 6.4x less money

Meet DeepTutor v1.5: Agent-native Personalized Tutoring.

DeepTutor connects teaching, practice, behavioral traces, unified Runtime, inspectable memory, and proactive IM companions into one evolving learner model.

It is an agent-native learning workspace that connects tutoring, problem solving, quiz generation, research, visualization, and mastery practice in one extensible system.

New research from Anthropic. Some AI capabilities are both helpful and dangerous.

E.g., knowledge of virology can be used to create life-saving vaccines or deadly pathogens.

Anthropic introduced GRAM, a training method that puts dual-use capabilities (like virology) into removable modules.

Researchers tested models from 50M to 5B parameters. Knowledge separation gets better with scale: larger models forget the removed subjects more thoroughly and resist retraining better.

Real training data isn't neatly labeled. So team tested what happens when half of it has no topic labels at all. Versus alternatives, GRAM better separates dangerous knowledge from the rest of the model.

Paper.
Code.

Meta just now introduced Muse spark 1.1 is an industry-competitive agentic and coding model.

It's available through a new Meta Model API and in Meta AI.

Meet The Agentic Garden of Forking Paths

Give two AI agents the same data, the same question, and the same instruction to "analyze it rigorously"; change only one paragraph's persona stating a prior belief, and they reach opposing scientific conclusions.

Code.

Check out TRACE, a new self-improvement approach where the agent identifies the missing capabilities behind its own failures and trains itself to address them.

TRACE-trained Qwen3.6-27B reaches 73.2% on SWE-bench Verified, outperforming much larger models like Codex 5.2 and GLM 5, while beating GRPO and GEPA with <1/4 the training rollouts.

By contrasting successful and failed trajectories, TRACE identifies its own weaknesses (such as bug localization or retrieval of the correct doc) and creates new synthetic environments to fix them.

The result is a transferable and sample-efficient synthetic env / data generation + fine-tuning pipeline for agentic tasks.

Stablecoin firm Circle wins final OCC approval to open national trust bank

Circle shares rose more than 7.7% in pre-market trading after the company said it received final approval from the U.S. Office of the Comptroller of the Currency (OCC) to establish First National Digital Currency Bank, N.A., which will operate as Circle National Trust.

Once launched, the bank will initially provide fiduciary digital asset custody services to Circle and its affiliates

Wowww! Yann LeCun is raising a new AI-focused VC fund.

He raised $1bn earlier this year for his startup Ami.

Now he's raising a fund.

The firm, named Extelligence Invest, will invest in companies across:

1. AI and data infrastructure
2. Healthcare technologies and human longevity
3. Cross-border consumer platforms
4. Frontier deeptech, including robotics, new mobility and compute.

Meet MIRA a playable, multiplayer world model.

Trained on 10k hours of data collected with publicly available bots, MIRA learns the dynamics of a four-player game. The model runs in real time at 20 fps, based on the keys you and the other players press.

Play the demo, read the technical report, and explore the open-source code

Built by General Intuition and Kyutai, in collaboration with Epic Games. Not used to develop Rocket League.

Meet Soofi S an open foundation model for German and English

Trained end-to-end on Deutsche Telekom’s Industrial AI Cloud in Munich.

Soofi S 30B-A3B a Mixture-of-Experts hybrid Mamba model trained on ~27 trillion tokens with deliberately up-weighted German.