This work instead describes the future of critical open world evaluations.

Most agentic benchmarks center around tasks that are automatically verifiable.

Benchmarks are saturated more quickly than ever. How should frontier AI evaluations evolve?

Research team argue that the AI community is already converging on an answer: Open-world evaluations. They are long, messy, real-world tasks that would be impractical for benchmarks.

A new project, CRUX, will run open-world evaluations regularly.

In a first experiment, team tasked an agent with publishing an iOS app to the App Store. The agent made two errors, but eventually succeeded after just one manual intervention.

Team released over 1GB logs from this experiment. Also pre-processed them for easy access in Docent.

Meet a completely local agent that lives right inside your browser

Powered by Gemma 4 E2B and WebGPU, it uses native tool calling to:

1. Search browsing history
2. Read and summarize pages
3. Manage tabs

100% local. No servers needed!

The extension is on the Chrome Web Store and the source is fully open:
Chrome Web Store
GitHub
HF.

New Anthropic research: Project Deal.

Anthropic created a marketplace for employees in their San Francisco office, with one big twist. Team tasked Claude with buying, selling and negotiating on colleagues’ behalf.

Claude interviewed 69 of colleagues about what they wanted to buy and sell. Each Claude asked for any custom instructions, then went off to haggle.

Team ran 4 markets in parallel, to find out what would happen if varied the models doing the negotiating.

At the end, team revealed which of the four runs was “real” and everyone met up to exchange their actual goods.

In short, this worked. Digital barterers agreed on 186 deals, at a total transaction volume of over $4,000.

In a survey, participants said Claude’s deals seemed fair, and surprisingly to team almost half said they’d be willing to pay for a service like this in future.

But the quality of the model mattered a lot. In the simulated runs where Opus and Haiku models negotiated with one-another, the Opus models got substantially better deals.

Interestingly, though, participants in our survey didn’t pick up on this disparity.

The custom instructions didn’t matter much. Claude followed them well. But “hardballing Claudes” didn’t generally fare better than “courteous Claudes.”

To team amazement, another Claude agent modeled its human’s preferences so accurately that based on only an offhand mention of an interest in skiing Claude bought him the exact snowboard he already owned.

Markets of AI agents could provide value, but there are plenty of rough edges. Access to higher-quality models conferred a real advantage—and participants didn’t notice. There are plenty of other ways they can go wrong.

Policy and legal frameworks will need to adapt to keep up.

Xiaomi dropped 2 open-source models:

1. MiMo-V2.5-Pro (Code Agent, 1T total)

2. MiMo-V2.5 (Multimodal Agent, 310B total).

HF.

And giving devs & creators 100T Free Token Grant for Builders 100t.xiaomimimo.com

Google Quantum team ran a verifiable quantum algorithm that probes how parts of a quantum system interact, from molecules to magnets and beyond.

On Willow chip, it ran 13,000× faster than the best classical supercomputers. A first in quantum computing.

If you're looking for the next multi-million dollar business idea - look no further.

YC and a16z are both recommending these AI business ideas.

This is where the smartest investors in the world are currently looking.

Wiz Research discovered remote code execution on GitHub with a single git push

The flaw in github allowed unauthorized access to millions of repositories belonging to other users and organizations

New work by Anthropic. Can LLMs simply tell us about unwanted behaviors they’ve picked up in training?

Research team train a single Introspection Adapter (IA) that makes fine-tuned models describe their behaviors.

It generalizes to detecting hidden misalignment, backdoors and safeguard removal.

IAs achieve SOTA on AuditBench, generalize to four OOD model families, and once trained can audit many fine-tuned models cheaply. This approach could also allow developers to target specific failure modes they care about.

Models and datasets.

DeepSeek’s multimodal model is now live, and some users are already able to try it out. So far, it’s performing pretty well.

Now the question is: will this one also be open-sourced?

Tomorrow might be a good time.

Securitize, Computershare open tokenization path for Apple, Tesla, Nvidia and 25,000 stocks

Securitize said it has reached an agreement with Computershare to support U.S. listed clients in issuing equity securities in tokenized form.

Issuers can add Issuer Sponsored Tokens (ISTs) alongside traditional shares and DRS without changing their capital structure.

ISTs represent actual shares in token form, not derivatives or wrappers, and are designed to fit within existing regulatory frameworks.

Cursor introduced SDK so you can build agents with the same runtime, harness, and models that power Cursor.

Run agents from CI/CD pipelines, create automations for end-to-end workflows, or embed agents directly inside your products.

With this SDK, you can run agents locally or deploy them in Cursor’s cloud.

Also open-sourced a few starter projects for you to build on: a coding agent CLI, a prototyping tool, and an agent-powered kanban board.

Mistral Introduced remote agents in Vibe and Mistral Medium 3.5.

You can now launch remote agents in the cloud, including from the CLI or Le Chat.

Plus, new Work mode in Le Chat for complex, multi-step tasks.

Mistral Medium 3.5, a new flagship model in public preview that merges instruction-following, reasoning, and coding into a single 128B dense model with a 256k context window and configurable reasoning effort.

It's a new default model for Mistral Vibe and Le Chat. Released as open weights, under a modified MIT license.

You can now launch remote Vibe agents, in parallel. And if you're already writing code locally you can teleport the session to the cloud mid-way history, task state, approvals, and diffs all come with it and keep going.

Le Chat now has Work mode (Preview) a powerful agent for complex long-horizon tasks like research, analysis, and actions across your connected tools.

Connectors are on by default so the agent pulls context from docs, email, and calendars as it works, and sessions persist across many turns until the task is complete.

Sakana AI introduced KAME: tandem architecture that shifts the paradigm to “speak while thinking.”

A fast speech model starts replying instantly, while a backend LLM runs in parallel to inject deep knowledge on the fly.

It’s a completely different way to approach conversational AI, making it feel remarkably more alive.

DeepSeek just released "Thinking with Visual Primitives"

It's a reasoning framework that lets models “point” with visual markers (points, bounding boxes) while they think.

Instead of describing locations in words, the AI grounds each step of its chain-of-thought directly onto image coordinates.

Result: despite being far smaller and using fewer image tokens, this model matches or beats GPT-5.4, Claude-Sonnet-4.6, and Gemini-3-Flash on challenging visual QA tasks.

This demonstrates a path toward more efficient and scalable System-2-like multimodal intelligence.

Paper.

The US-based brain implant company Axoft has tested its device in China, a sign of cooperation between the two countries that are competing to develop the most advanced neurotechnology.

It's the first public disclosure of a US brain implant business studying its device in China.

Axoft, which announced a $55 million funding round yesterday, said it's tested the implant temporarily in 11 patients, including 1 in Shanghai.

It's planning more trials in China as well as other countries.

Axoft is developing the implant to treat disorders of consciousness, like comas.

Thomas Reardon, who led work on the Meta Neural Band, is raising funds for a new startup, Flourish, that aims to develop energy-efficient AI

Flourish is in discussions to raise capital at a $2.5 billion valuation, according to the people. Investors have already committed $500 million, one of the people said.

The company has been recruiting AI researchers and neuroscientists.

It turns out, the brain is incredibly energy efficient - it creates consciousness and operates the body using about the same amount of power as a lightbulb.

AI that's more closely modeled on the brain may be more energy efficient than current tools, some scientists think.

Google DeepMind introduced AI co-clinician is a new research initiative to help explore how multimodal agents could better support healthcare workers and patients.

Research team tested the system while adapting the NOHARM safety framework, and found it made zero critical errors in 97 of 98 primary care queries - outperforming comparable systems in blind evaluations.

The system uses live video and audio to process physical symptoms in real-time. This means it could analyze a patient's walk, listen to their breathing, or look at how a rash is appearing.

To keep patient safety at the forefront, the system also runs on a dual agent architecture:

A built-in "Planner" continuously monitors the conversation verifying that the "Talker" agent stays within safe clinical boundaries.

In testing, AI co-clinician matched or outperformed physicians in 68 out of 140 assessed areas, including triage.

Yet humans were easily better at spotting crucial red flags and guiding physical exams - showing how these tools could augment clinical judgment.

Morgan Stanley has again raised its capex forecasts for the five hyperscalers Amazon, Alphabet, Meta, Microsoft, and Oracle.

It now expects them to spend about $805bn this year, up from a previous estimate of $765bn.

For next year, the forecast has been lifted from $951bn to $1.1trillion.

To put that into perspective, their 2026 spending alone would be roughly equal to what all non-tech companies in the S&P 500 spent combined in 2025.

The expected ~$800bn for 2026 is nearly double 2025 levels and about three times what was spent in 2024.

Harvard introduced 3 works on agentic AI for science

ToolUniverse - an open platform enabling AI agents to use scientific tools and databases at scale.

ClawInstitute shared research boards for long-running collaborative discovery where agents co-develop ideas over time.

Medea an omics AI agent for large-scale biological reasoning and analysis.

Researchers from Huawei Taylor Lab, Peking University, and Shanghai University of Finance and Economics introduced SHAPE.

The method rewards actual progress in reasoning not verbosity by using a two-level system: a stage-aware advantage at the segment level for efficient breakthroughs, and entropy-driven redistribution at the token level for sharper execution.

Result: 3% higher accuracy on math reasoning while using 30% fewer tokens across multiple base models and benchmarks.

Peter Thiel backs $1bn ocean data centre start-up powered by waves

Panthalassa operated mostly in secret for a decade. And what it built is nuts.

Massive, massive floating data centers that drive themselves out to sea and then capture water inside of them to spin a turbine and power GPUs.