Sakana AI introduced Inference-Time Scaling and Collective Intelligence for Frontier AI

AB-MCTS, a new inference-time scaling algorithm that enables multiple frontier AI models to cooperate, achieving promising initial results on the ARC-AGI-2 benchmark.

AB-MCTS combination of o4-mini + Gemini-2.5-Pro + DeepSeek-R1-0528, current frontier AI models achieves strong performance on the ARC-AGI-2 benchmark, outperforming individual o4-mini, Gemini-2.5-Pro, and DeepSeek-R1-0528 models by a large margin.

Many ARC-AGI-2 examples that were unsolvable by any single LLM were solved by combining multiple LLMs. In some cases, an initially incorrect attempt by o4-mini is used by R1-0528 and Gemini-2.5-Pro as a hint to get to the correct solution.

ARC-AGI-2 code.

The Multi-LLM AB-MCTS combination of o4-mini + Gemini-2.5-Pro + DeepSeek-R1-0528, current frontier AI models, achieves strong performance on the ARC-AGI-2 benchmark, outperforming individual models by a large margin.

Implementation of AB-MCTS on GitHub.

Meanwhile, Krea.ai has released a Video Restyle tool

It looks fancy, but under the hood is just a new video-to-video model from LumaLabs.

This is the Minimax Hailuo 02. It looks like it's been fed every sports report since 1896.

Baidu MuseSteamer

Another video generator.

It is deeply hidden in the Chinese jungle of a Chinese website and application.
It is called MuseSteamer - it can do 10 seconds (small ones on the website are 5), 1080P, any aspects, it is available in Turbo (available in beta now), Lite, Pro variations.

The main feature is Chinese lip sync (although the videos on the website are silent).

This is a development from Baidu

Judging by the examples, the quality of the model is the last generation. But on their website they also have a note that this is Model 1.0.

In general, we are waiting for them to go beyond China.

P.S. Does it seem to me alone that there are already more basic video models than basic models for generating images?

Hunyuan has released a new 3D generator.

Its name is Hunyuan3D-PolyGen.

This is definitely not the original version 2.1, and most likely an upgrade of version 2.5 (which is codeless).

Judging by the video, it looks pretty killer, but I wouldn't trust the videos.

They write that they made their own autoregressive model for retop and now you can play games or go to movies.

Now there are also 10,000+ polygons per model and increased generation accuracy.

https://3d.hunyuan.tencent.com/

Well, finally Google Flow has been rolled out to almost the entire world, including Europe. Now you don't have to use VPN

https://labs.google/fx/tools/flow

In the video, it's a lipsync on the initial photo.

Musk and the xAI team unveiled Grok 4, a new AI model built on the Colossus supercomputer with 200,000 GPUs, allowing it to perform 10x more RL computations than its competitors.

It leads the Artificial Analysis Intelligence Index with a score of 73, ahead of OpenAI o3 (70), Google Gemini 2.5 Pro (70), Anthropic Claude 4 Opus (64), and DeepSeek R1 (68).

Here's everything you need to know about it:

1. Grok 4 comes in 2 versions:

- Base - a single-agent model, fast and versatile, achieves 35% on Humanity's Last Exam (HLE) and 45% with additional computation.

- Grok 4 Heavy - a multi-agent version, where several AIs work as a "training group" to improve accuracy (50.7% on HLE). Needs some work, but is more powerful.

2. Grok 4 can generate full-fledged video games based on text queries, including code, graphics, and mechanics. The model also analyzes the "fun" of games, assessing their attractiveness. Gaming capabilities are still limited to prototypes, and complex projects require further development.

Here we have another competitor to Flux Kontext

This time from Nvidia.

This is an inserter of objects into pictures.

You can see sweet examples here:
https://research.nvidia.com/labs/par/addit/

High Input Fidelity for chatGPT

Precise image editing in GPT Image-1.

Until now it was clear that, unlike Flux Kontext, chatGPT adds quite a lot of detail to faces, poses, and environments when editing images.
They have just announced the High Input Fidelity parameter, which is designed to preserve details and more accurately preserve the appearance of the original image.
It works only through API and only on the platform https://platform.openai.com/playground/images/

Suno v4.5+

More variety in genres
Thicker vocals
Better prompt understanding
Speed up generation time
More creative beats: Specific improvements for genres like metal and pop-punk aimed at reducing repetition and increasing variety.
More complex harmonic arrangements
Stronger Covers support: can imitate celebrity voices and simulate specific roles and styles
Add Vocals - can be added to generations or to external files. Resko speeds up vocal selection.
Add Instrumentals does the opposite: you provide a vocal file (voice recording or vocal fragment), and the AI can immediately generate basic accompaniment for it to create a full song. The same vocals can be quickly used in different instrumental styles.

Looks like Musk, with his avatars Anya and Rudy, has pulled the emergency brake.

Hedra has just dropped real-time avatars.

They're definitely stepping into the territory of HeyGen and Character ai here, but it seems we're in for real-time avatars that can be connected not just to Grok or custom models, but to any LLM at all.

At least, that’s how it looks with Hedra.