Meta is releasing a standalone mobile app for its ChatGPT competitor, Meta AI.

There's a Discover feed that shows interactions that others (including your IG/FB friends) are having with the assistant. Meta tells the idea is to demystify AI and show “people what they can do with it."

OpenAI is working on a similar feed for ChatGPT.

Stanford and Google DeepMind released SWiRL: A synthetic data generation and multi-step RL approach for reasoning and tool use!

With SWiRL, the model’s capability generalizes to new tasks and tools. For example, a model trained to use a retrieval tool to solve multi-hop knowledge-intensive question answering tasks becomes significantly better at using Python to solve math problems (and vice versa).

As they scale the synthetic data size, the generalization gains continue to improve.

This suggests new possibilities for self improvement, where researchers use the model to synthetically generate data on multi-step tasks in more accessible (or affordable) domains and improve it on other domains.

Amazon introduces an architecture to migrate from various models to Amazon Nova models using DSPy and its MIPROv2 algorithm.

Turns out #DeepSeek does have a new release (671B math/prover model) but it's not R2.

Xiaomi MiMo-7B: a 7B reasoning model series trained from scratch, outperforms 32B+ baselines on math and code via dense RL

- pretrained on 25T tokens w/ multi-token prediction
- RL rewards from rule-verifiable math/code tasks
- cold-start RL model (MiMo-7B-RL-Zero) hits 93.6% MATH-500, 49.1% LCB v5
- SFT→RL variant matches OpenAI o1-mini
- also open: base + SFT checkpoints
- seamless rollout engine: 2.29× faster RL training
- vLLM + MTP inference ready
- strong AIME 2025 (55.4%) and LCB v6 (49.3%) results

Google DeepMind introduced the SAS prompt: LLM as Numerical Optimizers for Robot Self-Improvement

Large language models like Gemini have an inherent ability to problem solve, without needing to retrain for specific jobs.

Robots can use these models to improve how they operate over time, by interacting with the world, and learning from those interactions.

With the SAS prompt, you can now use language models like Gemini to learn from a robot's history.

This allows the model to analyze parameter effects, and suggest ways to improve - similar to a real-life table tennis coach.

Also Google released a dataset of table tennis ball throws, and a simple MuJoCo simulation environment able to replicate trajectories from the real world, with data on specific serves and rallies.

Paper.

Sam Altman's World Brings Biometric Verification and Digital Payments to US Market

World (formerly Worldcoin) has chosen six key innovation hubs for its American debut: Atlanta, Austin, Los Angeles, Miami, Nashville, and San Francisco. Americans in these cities can now:

1. Verify their unique World ID using the company's advanced biometric technology

2. Access the complete World App experience

3. Claim the Worldcoin (WLD) token airdrop.

The company's signature NVIDIA-powered Orbs — the biometric verification devices that distinguish humans from AI — will be available across the USA via standalone World Spaces and partner locations including Razer stores.

Alongside its identity verification system, World has announced the World Card a financial product that connects directly to users' World App wallets, enabling them to spend digital assets anywhere Visa is accepted.

Key features include:

1. Seamless integration with verified human identities through World ID
2. Ability to spend digital assets at over 150 million Visa-accepting locations globally
3. Merchants receive fiat currency without needing to understand crypto.

4. A rewards program specifically optimized for the AI economy, with enhanced rewards on AI subnoscriptions and services

5. Rewards paid directly in WLD tokens to connected wallets.

World emphasizes that its architecture places Americans in complete control of their digital identity:

- Personal data remains exclusively on users' devices through "Personal Custody"

- Advanced cryptographic systems, including Anonymized Multi-Party Computation and zero-knowledge proofs, ensure data privacy
- Verification of humanity without compromising personal information

Morgan Stanley plans to offer crypto trading to E-Trade clients.

Morgan Stanley is working on a plan to add cryptocurrency trading to its E-Trade platform, in what would be the most significant move by a major US bank to help everyday customers buy into the asset class since the Trump administration began removing regulatory barriers.

The project is nascent and executives envision launching the service sometime next year, according to people familiar with the matter. The firm is considering partnering with one or multiple established crypto firms as it sets up the mechanics for the brokerage’s clients to buy and sell popular tokens including Bitcoin and Ether.

Microsoft Introduced Phi-4-reasoning, adding reasoning models to the Phi family of SLMs.

The model is trained with both supervised finetuning (using a carefully curated dataset of reasoning demonstration) and Reinforcement Learning.

- Competitive results on reasoning benchmarks with much larger top-tier models up to DeepSeek R1.

- Strong performance on new tests released after data collection (AIME 2025, HMMT).

- Reasoning transfers/generalizes well to new domains even with only SFT (e.g. k-SAT, Mae Solving, Calendar Planning, etc.)

- Retains and often significantly improves general-purpose capabilities (e.g. instruction following).

HuggingFace Phi-4-reasoning
HF Phi-4-reasoning-plus
Hf Phi-4-mini-reasoning

Microsoft is getting ready to host Elon Musk’s Grok AI model. Microsoft has been in discussions with xAI to make Grok AI available on Azure's AI Foundry service.

In recent weeks Microsoft has been in discussions with xAI to host the Grok AI model and make it available to customers and Microsoft’s own product teams through the Azure cloud service.

The move could prove controversial internally and further inflame tensions with Microsoft’s partner OpenAI.

Huawei is building a 7nm fab in Shenzhen for its smartphone and Ascend chips, its first effort to manufacture its own high-end chips.

The Guanlan site is part of a sprawling network of new chip manufacturing sites all working on various elements of Huawei's push to become a semiconductor champion, from equipment to fabrication.

Huawei wasn't considered a serious player in chip manufacturing before it was sanctioned in 2019. The move kickstarted massive investment to localise chip technology, aided by state funds and led by the tech giant. The Guanlan network is part of this effort.

Cisco's Foundation AI released Foundation-Sec-8B

Built on Llama 3.1, the LLM matches Llama 3.1-70B & GPT-4o-mini on multiple security tasks

It will help with use cases like threat detection, vulnerability assessment, security automation, and more.

Carnegie Mellon University started company with only AI employees

They got OpenAI, Gemini, Anthropic, etc models and gave them job roles. They were retarded and costed a fuckton. Claude 3.5 was the best employee and only did 24% of its tasks.

Paper.

Google DeepMind presented Evaluating Frontier Models for Stealth and Situational Awareness:
- 5 evals of ability to reason about and circumvent oversight
- 11 evals for measuring a model’s ability to instrumentally reason about itself, its environment and its deployment

No SotA model currently shows concerning levels of either capabillity.

Anthropic launched a new "AI for Science" program

Under the initiative, the company will provide up to $20,000 in free API credits (for 6 months) to researchers in “high-impact” scientific fields like drug discovery, genomics, and agriculture

𝐀𝐌𝐎 is a universal whole‑body controller that unleashes the 𝐟𝐮𝐥𝐥  kinematic workspace of humanoid robots to the physical world.

AMO is a single policy trained with RL + Hybrid Mocap & Trajectory‑Opt.

Rethinking Memory in AI

Great overview of memory in AI agents with a more structured and dynamic perspective on research and ideas.

It's not just about simple storage and retrieval operations, it's also about maintaining, updating, and optimizing memory.

When building AI agents for more complex long-horizon tasks, you quickly start to see limitations in current memory solutions. The good news is that there a lot of devs and companies thinking and building around these ideas already.

Great read for devs and researchers.

Google released an updated Gemini 2.5 Pro

With this update, you can create even more complex web apps from a single prompt.

Future House launched an AI agent Finch that can do bioinformatics analysis, including repeating analysis from research papers.

It is multimodal and results in a complete jupyter notebook (python or R) that ends in a concrete conclusion. Starting with closed-beta now. Sign up here.

HuggingFace launched Computer Use in smolagents

As vision models become more capable, they become able to power complex agentic workflows. Especially Qwen-VL models, that support built-in grounding, i.e. ability to locate any element in an image by its coordinates, thus to click any item on a screenshot.