Digital asset platform Bullish announced that its $1.15 billion IPO proceeds were fully settled in stablecoins, making it the first IPO in the United States to be completed using stablecoin funding.

The stablecoins used include USDCV, EURCV, USDG, PYUSD, RLUSD, among others.

Brain-tuned speech models better reflect speech processing stages in the brain

Brain-like models would better reflect human speech processing. Yet previous work showed that popular speech models encode rich semantics in middle layers but poor semantics in early and late layers, differing significantly from the brain’s hierarchy.

Researchers addressed this limitation by utilising brain-tuning to fine-tune pretrained speech language models such as Wav2Vec 2.0 and HuBERT directly on fMRI data collected while participants listened to natural speech.

This brain-guided fine-tuning successfully aligned the models’ layers with human speech processing stages: early layers align best with primary auditory regions, while deeper layers align best with semantic brain regions.

Brain-tuned models also showed improved downstream comprehension and hierarchy, matching the brain alignment results. The downstream performance hierarchy goes from acoustic in the early layers to semantic in the late layers of brain-tuned models, unlike in pretrained models.

Impressively, brain-tuning not only changes the hierarchy to be more brain-like but also enhances brain alignment and downstream performance, leading to more effective model organisms than existing pretrained models, which lag behind in performance and semantic hierarchy.

Meet OpenCUA the first from 0 to 1 computer-use agent foundation model framework and open-source SOTA model OpenCUA-32B, matching top proprietary models on OSWorld-Verified, with full infrastructure and data.

OpenCUA — comprehensive open-source framework for computer-use agents, including:

1. AgentNet — first large-scale CUA dataset (3 systems, 200+ apps & sites, 22.6K trajectories)

2. OpenCUA model — open-source SOTA on OSWorld-Verified (34.8% avg success, outperforms OpenAI CUA)

3. AgentNetTool — cross-system computer-use task annotation tool

4. AgentNetBench — offline CUA benchmark for fast, reproducible evaluation

Proprietary CUAs like Claude or OpenAI CUA are impressive but there’s no large-scale open desktop agent dataset or transparent pipeline.

Paper
Models
Data
Code

Wow! China considering yuan-backed stablecoins to boost global currency usage.

Zhipu AI introduced ComputerRL, a framework for autonomous desktop intelligence that enables agents to operate complex digital workspaces skillfully.

ComputerRL features the API-GUI paradigm, which unifies programmatic API calls and direct GUI interaction to address the inherent mismatch between machine agents and human-centric desktop environments.

Researchers applied ComputerRL to the open-source GLM-4-9B-0414 model and evaluated its performance on the OSWorld benchmark.

DeepSeek introduced more information about DeepSeek-V3.1

API Update

1. deepseek-chat → non-thinking mode
2. deepseek-reasoner → thinking mode
3. 128K context for both

Anthropic API format supported.

Strict Function Calling supported in Beta API.

Tools & Agents Upgrades:

- Better results on SWE / Terminal-Bench
- Stronger multi-step reasoning for complex search tasks
- Big gains in thinking efficiency

Model Update:

1. V3.1 Base: 840B tokens continued pretraining for long context extension on top of V3
2. Tokenizer & chat template updated — new tokenizer config.
3. V3.1 Base Open-source weights
4. V3.1 Open-source weights.

Pricing Changes:

- New pricing starts & off-peak discounts end at Sep 5th, 2025, 16:00 (UTC Time)
- Until then, APIs follow current pricing.

MetaMask announced it will issue its native stablecoin, USD (mUSD), which is planned to launch later this year on Ethereum and Linea.

mUSD will be issued by Bridge, a Stripe-owned platform.

MetaMask also plans for mUSD to be spendable via the MetaMask Card at Mastercard-accepting merchants by year-end.

Google introduced a new agentic & personalization features in AI Mode in Search

1. Agentic web browsing capabilities from Project Mariner make it easier than ever to find and book restaurant reservations.

Tell AI Mode exactly what you’re looking for and Search will present a curated list of restaurants with available tables/timeslots for you to choose from.

All you have to do is take the last step and finalize the reservation on the booking page (also expand this soon to event tickets and local appointments).

This is rolling out for Google AI Ultra subscribers in the U.S. through the new “Agentic capabilities in AI Mode” experiment in Labs.

2. Get personalized dining recommendations, tailored to your unique taste. Now, when you search for dining-related topics in AI Mode, you'll see suggestions that are more relevant and personalized – based on your previous conversations and places you’ve searched or tapped in Search and Maps.

Available for people in the US who’ve opted into the AI Mode experiment in Labs.

3. A new link-sharing capability in AI Mode so it’s even easier to collaborate with friends & family. People who open your link will jump into the AI Mode responses where you left off and can ask follow-up questions.

You’re in control of what you share and can delete shared links at any time. This feature is available now for everyone in the US, no Labs opt-in required.

Chain-of-Agents

Interesting idea to train a single model with the capabilities of a multi-agent system. 84.6% reduction in inference cost.

This work proposes training single models to natively behave like multi‑agent systems, coordinating “role‑playing” and tool agents end‑to‑end.

They distill strong multi‑agent frameworks into CoA trajectories, then optimize with agentic RL on verifiable tasks.

This method, CAST, ends up significantly improving language following as compared to just labeling the original data.

Stanford researchers have demonstrated a BCI that decodes inner speech, silently imagined words, into text in real time.

The study involved participants with ALS and stroke, who preferred inner speech over attempted speech for being less tiring, faster, and more discreet.

The work highlights a new dimension in the BCI race: usability.

While the main BCI players have focused on attempted-speech decoding, Stanford’s results suggest that patient comfort may prove just as decisive as accuracy benchmarks.

The team also introduced safeguards, such as keyword unlocking, to prevent unintended decoding, an important example of ethical design built directly into BCI technology.

OpenAI and Retro Biosciences achieved 50x increase in expressing stem cell reprogramming markers

OpenAI with the team at Retro Biosciences designed novel variants of the Yamanaka factors that achieve a 50x increase in reprogramming efficiency in vitro compared to standard OSKM proteins – a groundbreaking improvement.

The key to their success was the development of GPT-4b micro – a new experimental biology LLM that we developed to test vision that AI is able to push the frontiers of science.

New Anthropic research: filtering out dangerous information at pretraining.

The Experiment with ways to remove information about chemical, biological, radiological and nuclear (CBRN) weapons from Anthropic’s models’ training data without affecting performance on harmless tasks.

The wealth of data used in AI training contains hazardous CBRN information. Developers usually train models not to use it.

Researchers trained six different classifiers to detect and remove CBRN information from training data.

The best and most efficient results were from a classifier that used a small model from the Claude 3 Sonnet series to flag the harmful data.

XPeng has been building one of China's most advanced humanoid robot operations.

New reporting reveals they've deployed hundreds of robots in their own car factories—not for manufacturing, but as massive data collection experiments.

The team is run by ex-NVIDIA veteran Mi Liangchuan, who once managed 100+ person teams at the chip giant.

XPeng shares 70% of the robot's tech stack with their cars. Same EEA architecture, same sensors, same AI infrastructure. It's industrial symbiosis at work.

While Tesla pitches Optimus for homes by 2026, XPeng is taking the opposite approach—master the controlled factory environment first, then expand. Very Chinese strategy: practical deployment beats perfect prototypes.

Sakana AI introduced M2N2 an evolutionary algorithm that combines pre-trained models without retraining from scratch

M2N2's Approach. 3 key innovations:

Dynamic merging boundaries — Instead of fixed layer-wise mixing, the algorithm evolves optimal "split points" for combining models

Competition for resources — Models compete for training examples, naturally maintaining population diversity without hand-crafted metrics

Attraction-based mate selection — Smart pairing based on complementary strengths rather than just performance

Results:

MNIST from scratch: M2N2 matched CMA-ES performance while being more computationally efficient.

LLM combination: Merged WizardMath-7B (math specialist) with AgentEvol-7B (web tasks):
Math (GSM8k): 40.16% vs 74.22% (pure WizardMath)
Web tasks (WebShop): 86.81% vs 88.88% (pure AgentEvol)
Created a generalist model without catastrophic forgetting
Diffusion models: Merged Japanese JSDXL with English models. Result understands both languages despite training only on Japanese captions.

Code

Today at Hot Chips Meta launched it's new puck + wrist + glasses compute platform called Orion

Orion uses "WLR" or world locked rendering.

There's quite a bit of compute, and the display processor, glasses processor, application processor, and compute coprocessor all use 5nm and has over 10b transistors with multiple processing units.

DeepSeek V3.1 has a serious bug: it randomly outputs “extreme” / “极” / “極” in unexpected places. This breaks code compilation, corrupts JSON, and more. Many suspect the root cause is data contamination.

Fine-tuning LLM Agents without Fine-tuning LLMs

Catchy title and very cool memory technique to improve deep research agents. Great for continuous, real-time learning without gradient updates.

Proposes a memory‑based learning framework that lets deep‑research agents adapt online without updating model weights.

The agent is cast as a memory‑augmented MDP with case‑based reasoning, implemented in a planner–executor loop over MCP tools.

Practical takeaways for agent builders:

• Use a compact, curated case memory with adaptive retrieval rather than growing prompts.

• Keep planning concise. A fast planner outperforms slow‑think planners for multi‑step tool use on GAIA by avoiding verbose or shortcut plans.

• Separate planning and execution with explicit Subtask and Tool memories to coordinate long‑horizon work and reduce hallucinations.

Google just upgraded Gemini 2.5 Flash image generation & editing

Give the model reference images and it can produce new visuals that maintain a character, subject or object’s likeness across different poses, lighting, environments or styles - helping you create more compelling, narrative-driven work.

Looking to apply a specific artistic style, design, or texture? 2.5 Flash can now easily transfer this from one image to another while preserving the previous subject's form and details.

Combine creative elements from multiple images with a single prompt. With 2.5 Flash, you can start blending different elements from up to three inputs to create a unique, unified composition

2.5 Flash can infer what happens before or after a moment shown in an image.

Claude Code's GitHub integration is now generally available with a simplified API, ready-to-deploy templates, and support for more GitHub events beyond @-claude mentions.

What’s new in GA:
- Trigger on more GitHub events (new issues, failed CI, custom conditions)
- Subagent support in actions
- Customizable templates for common workflows like code reviews.

Anthropic shipped 2 updates and developed Claude for Chrome, where Claude works directly in your browser and takes actions on your behalf.

About updates:

- 1M token context window is now the default for all Anthropic API users with Tier 4 and custom rate limits
- 1M token context is now available on Google Cloud's Vertex AI.