AI News: Meta Raybans, Gemini 3, World Labs, Grok 5, and more!

The future of Meta Ray-Ban glasses is showcased, featuring AI integration that allows the glasses to see, hear, and project information onto a private screen. Meanwhile, an AI reasoning model achieved superhuman intelligence by scoring perfectly on a collegiate programming competition. Additionally, a new method called refrag significantly speeds up retrieval augmented generation (RAG) by swapping tokens for precomputed chunk embeddings, improving efficiency and context handling. Tool calling is highlighted as a crucial part of AI scaffolding, with Zapier offering over 8,000 tools for automation.

AI chip builder Groq has secured $750 million in funding, indicating a strong demand for AI infrastructure and inference capabilities. Gemini 3.0 Ultra may be imminent, with mentions appearing in Google's code repositories. A new state-of-the-art benchmark on ARC Prize for V1 and V2 has been achieved using program synthesis outer loops and test-time adaptation. The Agent Payment Protocol (AP2) has been introduced, enabling agents to conduct secure transactions with each other and with merchants.

World Labs is offering a glimpse into a fully generated 3D environment, with the ability to generate entire worlds from a single image. Tongyi Deep Research has launched an open-source web agent that performs on par with existing models, demonstrating advanced capabilities without extensive human annotation. Dell Technologies is sponsoring AI workloads with their laptops featuring Nvidia RTX Pro Blackwell chips. Autonomous rides from Whimos are coming to San Francisco International Airport. OpenAI's GPT-5 Codex is proving more popular than anticipated, leading to slower performance as they scale up GPU capacity. YouTube is enabling creators to produce AI-generated Shorts, raising concerns about a potential influx of low-quality content. Hunan 3D 3.0 is released, offering enhanced precision and ultra HD modeling for 3D creations. Finally, a humanoid robot's quick recovery after being pushed is demonstrated, hinting at rapid advancements in robotics.

• Meta's future Ray-Ban glasses will integrate AI for enhanced vision and projection.
• An AI reasoning model achieved superhuman performance in a collegiate programming competition, scoring 12 out of 12.
• A new method, refrag, dramatically speeds up RAG by 30x with no accuracy loss.
• AI chip builder Groq secured $750 million in new funding.
• Google's Gemini 3.0 Ultra appears to be nearing release.
• The Agent Payment Protocol (AP2) allows agents to transact securely.
• Tongyi Deep Research launched an open-source web agent with state-of-the-art performance.
• YouTube is rolling out AI-generated Shorts for creators.

Related Video Summary

• CoreWeave CEO talks $12 billion OpenAI contract and AI-focused venture fund
• Cohere Founder, Nick Frosst: How To Compete with OpenAI & Anthropic, and Sam Altman’s AI Disservice
• I copied TikTok trends and made millions with AI
• 7 AI Tools That Actually Make Your Life Easier

Meta Ray-Ban Smart Glasses [00:03]

• Future Meta Ray-Ban glasses will feature AI that can see, hear, and project information onto a private screen.
• This technology is described as incredible and indicates Meta's strong commitment to the glasses form factor.

It truly is incredible technology.

This section introduces the advanced capabilities of upcoming Meta Ray-Ban smart glasses, highlighting their potential to integrate AI into everyday vision. The technology promises to provide users with an augmented reality experience directly through their eyewear.

OpenAI Reasoning Model Achieves Superhuman Performance [00:58]

• An AI reasoning model achieved a perfect 12 out of 12 score in the 2025 ICPC World Finals, a premier collegiate programming competition.
• This performance would have placed the AI first among all human participants in the 5-hour event involving 12 complex algorithmic problems.
• The AI received problems in the same format as humans and submitted answers without special testing harnesses.
• For 11 out of 12 problems, the system's first answer was correct, with the hardest problem solved on the ninth submission.
• The best human team achieved 11 out of 12 problems.
• The model used an ensemble of general-purpose reasoning models, including GBT5 and an experimental reasoning model, without specific competition training.

The core intelligence of these models are incredible. They are more than enough. We just need to build the scaffolding now.

This part details a significant AI achievement where a reasoning model demonstrated superhuman capabilities by mastering a highly competitive programming challenge, outperforming human experts. The success underscores the advanced reasoning and problem-solving abilities of current AI models.

Meta's Optimized RAG System (Refrag) [02:55]

• Meta Super Intelligence Labs optimized retrieval augmented generation (RAG) to make it faster.
• The new method, refrag, swaps most retrieved tokens for precomputed and reusable chunk embeddings.
• This improves speed by 30x and allows fitting 16x longer contexts without accuracy loss.
• This advancement is presented as building incredible scaffolding for AI models.

This is just another feather in the cap of building incredible scaffolding for these models.

This segment explains a technical innovation that significantly enhances the efficiency of RAG, a crucial system for AI to store and retrieve information from natural language documents. The refrag method promises faster performance and broader context handling for AI applications.

Tool Calling and Zapier Automation [03:43]

• Tool calling is highlighted as a critical component of AI scaffolding.
• Zapier, a sponsor, offers over 8,000 tools that can be called by AI.
• A workflow is demonstrated where a news story is converted into a social media post using Zapier for automation.
• This workflow involves creating records in Asana and Airtable, then using two paths for X and Instagram.
• Tools like ChatGPT for headlines, Firecrawl for assets, and Bannerbear for header images are used.
• The final post is sent to Buffer for publishing.
• This automation saves significant manual effort through AI orchestration by Zapier.

All of this saves us a tremendous amount of manual effort and it's all powered by AI orchestration by Zapier.

This section emphasizes the practical application of AI through tool calling and automation, showcasing how Zapier can orchestrate various services to streamline complex workflows, such as content creation and distribution.

Groq Secures Major Funding [04:57]

• AI chip builder Groq has secured $750 million in new funding.
• The post-funding valuation is $6.9 billion.
• The funding round was led by Disruptive and included BlackRock and Newberger Berman.
• Funds will be used to expand data center capacity, with new locations planned for this year and next.
• Groq plans to announce its first Asia-Pacific location this year.
• This indicates a massive demand and investment in AI infrastructure, with companies like Nvidia and Cerebrus also experiencing high demand.

It seems like the overwhelming thirst for inference for tokens is unquenchable.

This news highlights significant financial backing for an AI chip company, reflecting the immense and growing demand for the specialized hardware needed to power AI inference and data centers. The substantial investment underscores the rapid expansion of the AI infrastructure sector.

Gemini 3.0 Ultra Leak [05:49]

• A leak suggests Gemini 3.0 Ultra might be released soon.
• Gemini 3.0 Ultra was spotted in Google's Gemini CLI repo, committed recently.
• The presence in the CLI repo indicates that a beta release might be close.

CLI prep means beta is near. I cannot wait.

This snippet reports on a potential upcoming release of a new advanced AI model from Google, based on code repository evidence. The discovery suggests that the company is actively developing and preparing its next-generation AI for public access.

New State-of-the-Art Benchmark on ARC Prize [06:43]

• A new state-of-the-art benchmark has been achieved on the ARC Prize for V1, reaching 79.6% at $842 per task.
• For V2, the benchmark reached 69.4% at $3040 per task.
• These custom submissions were made by Gerber888 and Eric Pang.
• They utilized Grok 4 and implemented program synthesis outer loops with test-time adaptation.
• The method is based on the "dream coder" concept, where the model stores learned information to generalize.

If you're not familiar with what that is, drop a comment below and if you want me to explain it, just let me know.

This section covers a significant achievement in AI benchmarking, showcasing impressive results on complex problem-solving tasks. The use of advanced techniques like program synthesis and test-time adaptation demonstrates a sophisticated approach to AI development.

Agent Payment Protocol (AP2) [07:54]

• Google announced the Agent Payment Protocol (AP2), an extension of their Agent-to-Agent communication protocol.
• AP2 enables two agents to conduct secure, compliant transactions with each other and with merchants.
• It provides a common language for these transactions.
• Numerous partners are already lined up, including Adobe, Accenture, One Password, Intuit, Red Hat, Salesforce, and Okta.

So not only can the agents actually pay each other, but the agent can pay different merchants as well.

This announcement details a new protocol designed to facilitate financial transactions between AI agents and merchants, paving the way for autonomous economic activity in the digital realm. The broad partnership base indicates significant industry adoption of this technology.

World Labs' 3D Generated Environments [08:44]

• Early glimpses of work from Fei-Fei (the "godmother of AI") at World Labs are revealed.
• This includes a 3D fully generated environment that users can interact with.
• Users can generate entire worlds based on a single image, though this feature is currently on a waitlist.
• The generated worlds are described as massive and traversable.

It is truly incredible. I cannot wait to see what they do with this technology.

This segment showcases innovative work in generative AI, specifically the creation of immersive 3D environments from simple inputs. The potential for generating detailed and explorable virtual worlds from a single image is presented as a groundbreaking development.

Tongyi Deep Research's Open-Source Web Agent [09:29]

• Tongyi Labs has launched Tongyi Deep Research, an open-source web agent.
• It claims performance on par with OpenAI's deep research models, using only 30 billion parameters with 3 billion activated.
• It demonstrates state-of-the-art results, scoring 32.9 on Humanity's Last Exam, 45.3 on BrowserComp, and 75 on the Xbench deep research benchmark.
• The agent's capabilities are built on a novel automated multi-stage data strategy that creates high-quality training data without human annotation.
• This approach aims to scale up AI development by removing humans from the training and reinforcement learning loops.

This seems to be the trend of really being able to scale up. You have to remove the human from the loop in the training and the reinforcement learning steps.

This introduces a highly capable, open-source AI agent that achieves impressive performance with significantly fewer parameters. Its development strategy emphasizes automated data generation, highlighting a trend towards more scalable and efficient AI training methods.

Whimos Autonomous Rides at SFO [11:07]

• Whimos autonomous rides are coming to San Francisco International Airport (SFO).
• This follows a pilot permit approved by the mayor of San Francisco for autonomous rides at the airport.
• The rollout will occur in phases, with the goal of allowing anyone to request a Whimos ride directly from SFO.
• This addresses a previous limitation where Whimos were not available at the airport, requiring users to switch to other ride services.

But now, according to the mayor of San Francisco, the airport has approved a pilot permit to begin autonomous rides.

This news announces the expansion of autonomous vehicle services to a major transportation hub, marking a significant step towards integrating self-driving technology into airport infrastructure and passenger transport.

OpenAI's GPT-5 Codex Popularity and Performance Impact [11:51]

• GPT-5 Codex is significantly more popular than OpenAI initially anticipated.
• This high demand has led to performance issues, with Codex running 2x slower than targets.
• OpenAI is working to address this by scaling up GPU capacity.
• This situation highlights the critical need for robust infrastructure to support the growing demand for advanced AI tools.

Demand for GBT5 codecs is higher than forecasted. We need to stand up more GPUs.

This report indicates that a new AI coding tool has become unexpectedly popular, straining its operational capacity. The situation underscores the challenges of predicting user adoption and the importance of scalable infrastructure to meet demand for powerful AI services.

YouTube Introduces AI-Generated Shorts [12:41]

• YouTube is rolling out the ability for creators to generate YouTube Shorts using AI.
• The V3 fast model will be available within YouTube for easy clip generation.
• There are concerns about a potential influx of "AI slop" or low-quality content on the platform once the novelty wears off.
• The rise of AI-generated viral videos is acknowledged, but the long-term impact on content quality is questioned.

I feel a tsunami of AI slop coming to YouTube and so it's going to be interesting to see how Google handles this.

This development signals a significant shift in content creation on YouTube, as AI tools will now enable the generation of short-form videos. While potentially democratizing content creation, it also raises concerns about maintaining content quality and originality.

Hunan 3D 3.0 for Realistic 3D Modeling [13:33]

• Hunan announces Hunan 3D 3.0, featuring 3x higher precision and a higher geometric resolution.
• It boasts 3.6 billion voxels for ultra HD modeling, creating detailed and realistic 3D models.
• The technology can create lifelike faces with natural contours and poses.
• It accurately reconstructs complex structures from input images using a layer generation strategy.
• The system generates ultra-clear, professional-grade details with enhanced texture fidelity for immersive visuals.
• This is presented as beneficial for designers and creators.

Creates faces with lifelike facial contours and natural poses, creating truly realistic highfidelity models.

This announcement details an advanced 3D modeling technology that significantly enhances realism and detail in generated models. The improved precision and fidelity are expected to benefit creators in fields like design and virtual reality.

Humanoid Robot Agility Demonstration [14:14]

• A video showcases a humanoid robot's impressive agility.
• The robot is shown being pushed but quickly recovers and gets up off the ground at a stunning speed.
• This rapid recovery suggests advanced balance and movement capabilities.

The speed at which it gets up off the ground is really stunning.

This brief observation highlights the remarkable agility and rapid recovery capabilities of a humanoid robot, suggesting significant progress in robotic locomotion and stability. The speed and responsiveness shown are indicative of advanced engineering in robotics.