Alibaba introduced a new GPU pooling system Aegaeon that makes AI model serving much more efficient.

Claims an 82% cut in Nvidia GPU use for serving LLMs by pooling compute across models.

In a 3+ month beta on Alibaba Cloud’s marketplace, H20 GPUs dropped from 1,192 to 213 while serving dozens of models up to 72B parameters.

The regular Cloud model hubs skew toward a few hot models, so many GPUs sit idle for cold models, and Alibaba measured 17.7% of GPUs handling only 1.35% of requests.

Aegaeon addresses this with token-level auto-scaling, which lets a GPU switch between models during generation instead of waiting for a full response to finish.

By slicing work at token boundaries and scheduling small bursts quickly, the system keeps memory warm and compute busy with little waste.

With Aegaeon, a single GPU supports up to 7 models versus 2 to 3 in other pooling systems, and switching latency drops by 97%.

Cold models load weights just in time when a request lands, then borrow a brief slice of compute without locking an entire GPU.

Hot models keep priority, so heavy traffic stays smooth while sporadic models borrow capacity in short bursts.

The wins apply to inference, not training, because generation happens token by token and fits fine-grained scheduling.

The timing suits China’s chip limits, where H20 targets inference workloads and domestic GPUs are ramping, so fewer chips can cover more traffic.

If Aegaeon generalizes, operators can lower cost per token, raise fleet utilization, and delay new GPU purchases without hurting latency for popular models.

Tradeoffs still exist, like uneven memory needs across models, long sequences that reduce preemption points, and scheduler overhead during traffic spikes.

DeepSeek released an OCR model

Their motivation is really interesting: they want to use visual modality as an efficient compression medium for textual information, and use this to solve long-context challenges in LLMs.

Of course, they are using it to get more training data for their models as well.

"DeepSeek-OCR can generate training data for LLMs/VLMs at a scale of 200k+ pages per day (a single A100-40G)."

HuggingFace.

Anthropic launched Claude for Life Sciences to support the entire life sciences process from early discovery through translation and commercialization, with Claude Sonnet 4.5 showing improved performance on protocol understanding and bioinformatics tasks, and new connectors to Benchling, BioRender, PubMed, Scholar Gateway, Synapse, and 10x Genomics

Anthropic is also developing life sciences-specific Agent Skills, beginning with single-cell-rna-qc that performs quality control and filtering on single-cell RNA sequencing data using scverse best practices

Princeton introduced Skill-Targeted Adaptive Training(STAT)

STAT uses a supervisor model and a skill catalog to construct a Missing-Skill-Profile for each student model, and then modifies training to squeeze out >=7% more performance.

The intervention can be as simple as reweighting existing training sets.

You can also think of this as a more effective distillation method.

STAT shows that leveraging skills during training can greatly help too e.g., Qwen can continue to learn new tricks from Hendrycks MATH, which it had been over-trained on.

GitHub.

Anthropic launched a sandbox within Claude Code that allows you to define exactly which directories and network hosts your agent can access.

also open sourced this sandbox tool so you can use it to sandbox other parts of your agent workflows.

In particular sandbox the bash tool with file and networking isolation to ensure that Claude only accesses files and networks you approve of.

When enabled this should significantly improve Claude's resistance to prompt injection, both in the CLI & SDK.

Nexar introduced a new AI model designed to predict and prevent car crashes — BADAS 1.0.

It beat SOTA models by learning from 10B+ real miles and 60M+ real events, not simulations

Based on Meta FAIR's V-JEPA 2.

OpenAI just dropped browser—ChatGPT Atlas.

Agent mode in Atlas completes tasks faster as you browse the web.

Available in preview for Plus, Pro, and Business users.

Available today on macOS. ChatGPT can see the page you’re on and answer your questions right there via the Ask ChatGPT sidebar.

ChatGPT can offer suggestions wherever you’re typing on the web. Ask ChatGPT to open, close, reopen, bookmark or revisit any of your tabs.

Meta showed how sparsely finetuning memory layers enables targeted updates for continual learning, w/ minimal interference with existing knowledge.

While full finetuning and LoRA see drastic drops in held-out task performance, memory layers learn the same amount with far less forgetting (-11%).

Airbnb CEO Brian Chesky: “We’re relying a lot on Alibaba’s Qwen model.

It’s very good. It’s also fast and cheap... We use OpenAI’s latest models, but we typically don’t use them that much in production because there are faster and cheaper models.”

A new Morgan Stanley report on BCI reveals a future that's closer than we think. A report from 2024 here.

Here are the key takeaways:

1. The core thesis isn't just medical. It's existential. As AI accelerates exponentially, BCI is seen as humanity's "chance to keep up." The ultimate goal is a seamless symbiosis, merging human consciousness with machine intelligence.

2. The path to mass adoption runs through medicine. With a US healthcare TAM of ~$400 Billion, BCIs will first restore sight to the blind, movement to the paralyzed, and speech to the voiceless. This addresses a dire need, creates a willing patient base, and accelerates regulatory approval.

3. Neuralink isn't just a player; it's the pacesetter. With 12 human patients already using its "Telepathy" device to control computers with their minds, the company is demonstrating a viable product.
Roadmap: From "Telepathy" (mind-control of devices) to "Blindsight" (restoring vision) by 2030.
Vertical Integration: Their secret sauce is controlling the entire stack—the chip, the surgical robot, and the software.
Funding & Hype: Recently raised $650M at a $9BN valuation, backed by top-tier VCs.

4. Key competitors are taking different, less invasive approaches:
Synchron: Uses blood vessels to place its Stentrode implant (no open-brain surgery).
Precision Neuroscience: Places a thin film on the brain's surface.
Merge Labs (by Sam Altman): Exploring non-invasive sonogenetics (using ultrasound).

5. The Inevitable Challenges & Risks
The "Neuro-Elite": Will this create a new class divide between enhanced and non-enhanced humans?
Data Security: How do we protect the most personal data imaginable—our neural signals from hacking?
Ethical Quagmire: The transition from therapy to human enhancement will be the defining ethical debate of the coming decades.

Google together with UCL made a free AI Research Foundations curriculum – available now on Google Skills.

Tahoe Therapeutics has built AI bio's largest virtual cell model. And it's giving it away for free.

Tahoe-x1 is a foundation AI model trained to simulate entire cell behaviors, including gene expression, drug responses, and interactions across biological contexts. It’s touted as the largest of its kind in AI-driven biotechnology and is open-source to foster community research.

The model was trained on over 300 million unique cells, incorporating:
- Tahoe-100M а dataset of 100 million cells with 60,000 drug-cell interactions (1,200 molecules tested on 50 cancer cell lines).

This was a "world record" at the time, built with Parse Biosciences (sample prep) and Ultima Genomics (sequencing).

- Future Goals: Tahoe aims to scale to 1 billion single-cell datapoints and 1 million drug-patient interactions, likened to a "GPT moment" for biology, akin to large language models for text.

Models built on Tahoe-100M (e.g., by Arc Institute) showed 2x better accuracy than competitors. Tahoe-x1 advances this further by integrating diverse data across species, tissues, and perturbations (external stimuli like drugs).

Key Achievements and Partnerships
1. Tahoe-100M contributed to the Arc Virtual Cell Atlas (300+ million cells), a public resource launched with Arc Institute in February 2025, making data accessible to researchers worldwide.

2. Tahoe-x1 has identified new drug candidates for cancer (including "undruggable" targets) and novel therapeutic targets. The company is advancing its pipeline toward clinical trials and seeking one strategic partner (pharma or AI company) for co-development.

3. Tahoe competes with initiatives like the Chan Zuckerberg Initiative, which also aims for 1 billion cell datasets for AI modeling.

Ant group introduced Ring-1T, a 1T-parameter MoE reasoning model with ~50B params active per token.

It’s trained with a long-CoT SFT phase, a verifiable-rewards reasoning RL phase, then a general RLHF phase, and introduces three pieces that make trillion-scale RL actually run:

- IcePop to stabilize updates

- C3PO++ to keep GPUs busy under a token budget

- ASystem to unify high-throughput RL stack

On benchmarks, it leads open weights on AIME-25, HMMT-25, ARC-AGI-1, LiveCodeBench, CodeForces, and ArenaHard v2.

It reaches silver-medal level on IMO-2025 using only natural-language reasoning.

Microsoft introduced 12 new Copilot features. With Groups, Copilot goes multiplayer.

You can now collaborate in real time with your team + Copilot to brainstorm, co-write, plan, or study together.

Plan and collaborate with the new Groups feature—perfect for organizing study sessions, family vacations, or a night out. Copilot makes your crew more productive by answering questions, assigning tasks, and getting everyone moving.

More engaging conversations with Copilot through the new Mico appearance. It brings that much more personal experience to the notion of your AI companion.

And, shop smarter online with the new Edge, AI browser. Copilot, with your permission, can evaluate your open tabs to make more confident decisions and then take action to help you book reservations, travel, or make smarter shopping decisions.

You spend $1B training a model A.
Someone on your team leaves and launches their own model API B.

You're suspicious. Was B was derived (e.g., fine-tuned) from A?
But you only have blackbox access to B.

With this paper, you can still tell with strong statistical guarantees (p-values < 1e-8).

Idea: test for independence of A's training data order with likelihoods under B.

There are crazy amounts of metadata about training process baked into the model that can't be washed out, like a palimpsest.

Stanford and Tsinghua University presented Ctrl-World is a controllable world model that generalizes zero-shot to new environments, cameras, and objects.

Build on pre-trained video model, Ctrl-World adds key designs to make it compatible with modern VLA:

1) Fully controllable via low-level action conditioning.
2) Multi-view prediction including wrist-view.
3) Context as memory for consistency.

Despite these promising progress, also notice WM can still:

- Fail in modeling complex physical interactions.
- Sensitive to initial observations.

Tahoe introduced Tahoe-x1 (Tx1) a 3B parameter, single-cell foundation model that learns unified representations of genes, cells, and drugs, achieving state-of-the-art performance across cancer-relevant cell biology benchmarks, open-sourced on HuggingFace.

Tx1 is the first billion-parameter, compute-efficient foundation model trained on perturbation-rich single-cell data. And it is fully open-source with open weights.

Tx1 is a Transformer (scGPT-inspired self-supervised objective) that stays practical to train at 3B parameter scale, enabling empirical search for optimal architectures and hyperparameters for modeling cells. It is 3-30x more compute efficient than other cell state models.

To make it so, researchers borrowed the best tricks from LMs (FlashAttention v2, FSDP, Dataset streaming, Mixed-precision + multi-node scaling). But even cooler: researchers improved the attention operation at the heart of these models.

Researchers designed new benchmarks to evaluate its performance in cancer-relevant discovery and translational tasks.

Tx1 achieves SOTA on key discovery tasks. It outperforms other models (and for the first time matches with linear baselines) in predicting gene essentiality as measured by the landmark DepMap dataset, a key piece of data in identifying subtype-specific targets.