Meta introduced StepWiser

- Reframes stepwise reward modeling as a reasoning task: outputs CoT + judgment.

- Trained by RL using relative outcomes of rollouts.

Results:
1) SOTA performance on ProcessBench!
2) Improves policy at train time.
3) Improves inference-time search.

Anthropic launched the PHP SDK for the Anthropic API. Now offer SDKs for Python, TypeScript, Java, Go, Ruby, and PHP.

Tencent: VLMs can self-reward during RL training — no visual annotations needed

-Decompose VLM reasoning into visual vs. language parts
-Prompt the same VLM without visual input for visual reward

The US Commerce Dept has published Q2 GDP data on Bitcoin, Ethereum, Solana, TRON, Stellar, Avalanche, Arbitrum, Polygon & Optimism, with distribution via Pyth and Chainlink.

OpenAI released a new Codex features:

- A new IDE extension
- Easily move tasks between the cloud and your local environment
- Code reviews in GitHub
- Revamped Codex CLI

Powered by GPT-5 and available through your ChatGPT plan.

Microsoft introduced MAI-1-preview:
- the first foundation model trained end to end in house
- in public testing on LMArena.

Microsoft has been working on MAI-1 since at least May 2024. The goal was to have an in-house alternative to OpenAI's models. If previous leaks in the press were correct we even know its size: 500B.

xAI has released their coding model. It is free to try out for a limited time on multiple platforms.

This is a guide with tips on how to get the best results from Grok Code Fast 1.

The model is generally available via the xAI API, priced at $0.20 / 1M input tokens, $1.50 / 1M output tokens, and $0.02 / 1M cached tokens.

Grok Code Fast 1 is versatile across the full stack and is particularly strong at TypeScript, Python, Java, Rust, C++, and Go.

For the next 7 days, Grok Code Fast 1 is available for free on popular agentic coding platforms including Cursor, GitHub Copilot, Cline, opencode, Windsurf, Roo Code, and Kilo Code.

Is GPT-realtime too pricey for you? Here is a free alternative - Step-Audio 2 mini (7B): new SOTA open source LALM:

- Free alternative to expensive GPT Realtime API
- End-to-end speech I/O
- Advanced speech & audio understanding
- Expressive prosody control
- Intelligent speech conversation
- Web search support
SOTA on LibriSpeech, MMAU, URO-Bench & more.

HF.

This humanoid robot can now rally over 100 consecutive shots against a human in real table tennis — fully autonomous, sub-second reaction, human-like strikes.

Google DeepMind claims: general agents contain world models

This study gives a formal answer:

- Any agent that generalizes to multi-step goals must have learned a predictive model of its environment.
- This model can be extracted from its policy.
- Higher performance & more complex goals demand increasingly accurate world models.

Implications span:

- Safer, more general AI
- Limits on agent capabilities in complex settings
- New algorithms to elicit world models directly from agents.

New open-weights Chinese model with a really detailed tech report dropped. Meituan (delivery app company DoorDash) released LongCat-Flash.

The model is a 560B passive ~27B active MoE with adaptive number of active parameters depending on the context thanks to the Zero-Computational expert.

1) New architecture
-Layers have 2 Attention blocks and both FFN and MoE, that way you can overlap the 2 all-to-all coms. (also it's only 28 layers but you have to take into account the 2 attention blocks).
- They add the zero-computational expert that tokens can choose and do nothing, kinda like a "sink" for easy tokens.
- For load balancing, they have a dsv3-like aux loss free to set the average real/fake expert per token. They apply a decay schedule to this bias update. They also do loss balance control.

2) Scaling
- They made changes to MLA/MoE to have variance alignment at init. The gains are pretty impressive in Figure 5, but i don't know to what extent this has impact later on.
- Model growth init is pretty cool, they first train a 2x smaller model and then "when it's trained enough" (a bit unclear here how many B tokens) they init the final model by just stacking the layers of the smaller model.

3) Stability
- They track Gradient Norm Ratio and cosine similarity between experts to adjust the weight of the load balancing loss (they recommend Gradient Norm Ratio <0.1).
- To avoid large activations, they apply a z-loss to the hidden state, with a pretty small coef (another alternative to qk-clip/norm).
- They set Adam epsilon to 1e-16 and show that you want it to be lower than the gradient RMS range.

4) Others
-They train on 20T tokens for phase 1, "multiple T of tokens" for mid training on STEM/code data (70% of the mixture), 100B for long context extension without yarn (80B for 32k, 20B for 128k). The long context documents represent 25% of the mixture (not sure if it's % of documents or tokens, which changes a lot here).
- Pre-training data pipeline is context extraction, quality filtering, dedup.
- They release two new benchmarks Meeseeks (multi-turn IF) and VitaBench (real-world business scenario).
- Lots of details in the infra/inference with info on speculative decoding acceptance, quantization, deployment, kernel optimization, coms overlapping, etc.

HuggingFace.

Neurofounders Startup Map

Neurofounders launched a living database of 100+ cutting-edge neuroscience startups.
Submit a startup here.

Each company is listed with:
- Logo & mission summary
- Funding stage & founding year
- Country of origin & identifying tags

You can search, sort, and filter to explore the landscape across five categories:
1. Brain-Computer Interfaces
2. Neuromodulation
3. Neuroimaging
4. Tools & Infrastructure
5. Cognitive & Mental Health

The map already includes many of the most groundbreaking ventures, but it’s just the start.

Apple just released FastVLM and MobileCLIP2 on Hugging Face.

The models are up to 85x faster and 3.4x smaller than previous work, enabling real-time vision language model (VLM) applications.

It can even do live video captioning 100% locally in your browser.

Try the live demo here.