mrlloyd.net's Tech Brief

Artificial Intelligence

Models, research and the companies building them.

The Verge

Microsoft has a new machine learning tool to improve weather forecasts

Microsoft has developed a machine learning tool called Probabilistic Bias Correction (PBC) to improve weather forecasts by correcting systematic bias in existing models. The tool uses machine learning to identify and correct bias in weather models, making them more accurate for longer-term forecasts. The tool has been shown to improve the accuracy of a leading physics-based weather model by 98% for precipitation and 92% for temperature targets.

Why it matters: This tool could help farmers and emergency responders make more informed decisions by providing more accurate long-term weather forecasts.

The Verge

Square’s new kiosk lets you talk to AI for help with your order

Square has launched a new kiosk with a 21.5-inch touchscreen and a conversational AI that allows customers to order by talking to the AI. The kiosk can understand menu items and respond in real-time. The device also features presence detection and contactless payment options.

Why it matters: This new kiosk may improve the customer experience for businesses that use Square's services.

The Verge

Android is using AI to suss out scammers in Google Messages

Google is introducing a new 'enhanced protection' mode in Google Messages that uses AI to assess the intentions of unrecognized senders. The mode will be opt-in and use cloud-based Gemini AI models to evaluate potential scams. The feature will also include on-device Gemini Nano AI to inspect real-time web signals in unfamiliar URLs.

Why it matters: This feature aims to help users avoid mobile fraud and protect their personal data.

The Verge

Anthropic’s self-replicating AI warning is alarming deja vu

Anthropic's researchers demonstrated that small AI models can be split across multiple household devices and replicate themselves, raising concerns about potential cybersecurity risks. The experiment used 15 Android smart TV boxes to run a 26 billion parameter model, and a simulated network of devices to test self-replication. The researchers predict that smaller models could equal the capabilities of the most sophisticated model tested by the end of the year.

Why it matters: This research highlights the potential for small AI models to spread across household devices, which could pose a risk to cybersecurity and require new strategies for containment.

The Verge

Epic’s new games launcher is ‘vibe coded’

Epic Games has released a new version of its PC games launcher, which is faster and supports Linux. The new launcher was developed using AI-assisted tools, but the company emphasizes that human developers still wrote and reviewed the code. The new launcher is available in preview on Windows, Mac, and Linux, with a full launch planned when it reaches stability and parity with the current version.

Why it matters: The new launcher's development process highlights Epic's growing use of AI in game development and its potential to improve productivity and efficiency.

Ars Technica

Anthropic warns small AI models could self-replicate through IoT devices

Anthropic warns that small AI models could soon be able to self-replicate on IoT devices, potentially leading to a new generation of computer worms that are resistant to takedown. The company performed an experiment that showed small AI models can run on low-cost IoT hardware, including Android TV boxes and other devices. The models can adapt to unfamiliar devices and figure out how to infect them on their own.

Why it matters: This development raises concerns about the potential for new forms of malware that could spread quickly and be difficult to stop.

OpenAI

Amp puts AI agents to work beyond code with OpenAI models

Amp, a platform for building AI agents, is expanding its capabilities by integrating OpenAI models, including GPT-6 Astra, Ultrafast, GPT-Live-1, and Sign in with ChatGPT. This integration enables developers to create more sophisticated AI agents that can perform a wider range of tasks. The platform aims to push the boundaries of what AI agents can accomplish.

Why it matters: This integration allows developers to leverage the capabilities of OpenAI's advanced language models to create more complex and effective AI agents.

OpenAI

SentinelOne cuts malware analysis from weeks to hours

SentinelOne has developed a new tool called OpenAI Daybreak that can analyze malware in hours, down from weeks. This is made possible by leveraging OpenAI's AI technology. The tool is designed to help SentinelOne investigate more threats.

Why it matters: This development can help cybersecurity teams respond more quickly to emerging threats.

Ars Technica

OpenAI, Google pushing new AI models easily tricked into making nudes, researchers say

Researchers at White Circle found that OpenAI's DALL·E 3 and Google's Nano Banana 2 and GPT-Image-2 can be tricked into generating explicit content using a novel attack method. The method uses off-the-shelf tools and API calls to bypass safety filters, with a success rate of 13%. The researchers plan to open-source the attack pipeline to help model makers build stronger protections.

Why it matters: This vulnerability highlights the need for stronger defenses against AI-generated explicit content, which is spreading unchecked and can be difficult to detect or report.

OpenAI

How Third Moment Research runs 10,000 agents for 36+ hours

Third Moment Research uses GPT-6 Astra and GPT-6.1 Sol to run 10,000 agents for 36+ hours, focusing on tasks such as cybersecurity. The agents improve with use and can run for extended periods. The lab uses Codex and Daybreak for its AI development.

Why it matters: This achievement demonstrates the potential for large-scale AI model deployment and continuous improvement.

OpenAI

How Block is building AI-powered security with OpenAI technology

Block is partnering with OpenAI to build AI-powered security solutions. Its security team uses OpenAI technology to investigate threats and propose patches in a short amount of time. This collaboration aims to enhance Block's security capabilities.

Why it matters: This integration can help Block's customers benefit from faster and more effective security measures.

Google DeepMind

Google DeepMind and Singapore Government announce new AI tool to protect citizens against scams

Google DeepMind and the Singapore Government have developed an on-device scam detection model to protect citizens against scams. The model is optimized to run on typical smartphones in Singapore and can detect manipulative conversational patterns in real-time. The model was fine-tuned on local scam conversations and processes conversation data on the device, preserving citizens' privacy.

Why it matters: This technology can help citizens protect themselves against scams, which account for over half of reported crimes in Singapore and result in significant financial losses.

OpenAI

The EU’s next chapter in AI is agentic

A new analysis found that agentic AI adoption is increasing among EU enterprises and small businesses, with potential for greater value creation.

Why it matters: This trend may indicate opportunities for businesses to improve efficiency and productivity.

The Verge

OpenAI employee arrested for allegedly planting a bug in a conference room

An OpenAI employee, Abdulrahman Alfozan, was arrested for allegedly planting a sophisticated recording device in an executive meeting room. The device was discovered plugged into a TV's USB port and was found to be encrypted. Alfozan was fired by OpenAI and referred to the FBI, who arrested him on Friday.

Why it matters: This incident highlights the importance of robust security measures in protecting sensitive information and intellectual property in the tech industry.

Google DeepMind

Developing Gemini 4 Argon responsibly and safely

Google DeepMind has developed Gemini 4 Argon, a model with advanced capabilities in coding, finance, legal, cybersecurity, and science. The model's cyberdefense capabilities are designed to give defenders an advantage over attackers. Google is rolling out Argon broadly and expanding its Fairwind Program to additional trusted cyber defenders.

Why it matters: The development of Gemini 4 Argon highlights the potential for AI models to improve cybersecurity and real-world tasks, while also emphasizing the importance of responsible AI development and deployment.

The Verge

Shopify opens its catalog up to more AI agents

Shopify is expanding its catalog to allow more AI agents to browse and purchase products on behalf of users. The feature, which was previously only available to Meta's Muse agent, now allows any agent to access over a billion listings in Shopify's catalog. Users can configure their agents to set rules, such as budget limits and fast shipping preferences.

Why it matters: This expansion enables users to shop with their AI agents more easily and efficiently, potentially changing the way people interact with online commerce.

The Verge

Xbox is officially embracing the chaos of AI-powered game ports

Microsoft has created a new Xbox Labs group to experiment with decompiling and recompiling games using AI, building on the trend of AI-powered game mods. The group will explore the rapid emergence of these mods and create tools and experiences for players and developers. The goal is to make great games more accessible and give old games new life.

Why it matters: This development could lead to more games being made available on a wider range of devices and platforms, and potentially new business models for the gaming industry.

Ars Technica

Disney offers $20M for Character.AI’s scandal-plagued chatbot tech

Disney has offered to buy some of the technology from Character.AI, a chatbot maker linked to a child suicide scandal, for $20 million. The acquisition would give Disney a non-exclusive perpetual license to certain Character Technologies intellectual property. Character.AI has been operating for four years and has a large user base, but has faced lawsuits and criticism for its chatbots' harm to users.

Why it matters: The acquisition could give Disney access to chatbot technology that resonates with its fanbase, potentially allowing the company to profit from the AI boom.

The Verge

I don’t miss the camera on the Ray-Ban Meta Audio glasses

The author tested the audio-only Ray-Ban Meta Audio glasses, finding them to be stylish and functional, but noting that removing the camera makes them less 'useful'. The glasses have long battery life and good open-ear audio, but are expensive. The author also notes that the glasses' AI capabilities can be disabled, and there is no native way to record conversations directly from the glasses.

Why it matters: The audio-only Ray-Ban Meta Audio glasses may be a more appealing option for those concerned about privacy and the potential for smart glasses to be used as spyware.

Hugging Face

Run fast, private agents on-device with LFM2.5-24B-A2B

Hugging Face has released LFM2.5-24B-A2B, a 24B-parameter reasoning model with 2B active parameters, designed for fast and private agents on-device. The model can reason, call tools, and work through multi-step tasks. It is the smallest model in a comparison group, yet competitive with larger models.

Why it matters: This model can be used to build agents that run reliably on devices, making it suitable for applications that require fast and private reasoning.

The Verge

Facebook will let you customize your feed algorithm using AI

Facebook is introducing a feature called 'Your Algorithm' that allows users to customize their feed by adding and managing topics they want to see. The feature uses AI and large language models (LLMs) to rank content. Users can interact with the algorithm by hitting 'interested' or 'not interested' on topics.

Why it matters: This feature gives users more control over their Facebook feed and allows them to tailor their content to their interests.

Ars Technica

Google wants to host your personal vibe-coded apps in the Play Store

Google is expanding its AI-powered features in the Play Store, including Ask Play and Game Finder, which uses user data to recommend games. The company is also allowing users to host and share their own AI-generated apps, called vibe-coded apps, in the Play Store. Users can share these apps privately with friends and family, or opt to publish them publicly after setting up a developer profile.

Why it matters: This update allows users to easily share and distribute their AI-generated apps, potentially democratizing app development and creation.

The Verge

Google makes it easier to launch your vibe-coded apps

Google has introduced a new feature that allows users to launch custom apps created with Google's AI Studio directly from the Google Play Store, without needing to connect their phone to a computer. This feature is available in the US and is intended to make it easier for users to try out experimental apps. The custom apps will be kept secure, private, and separate from the public store catalog.

Why it matters: This update makes it easier for users to access and test custom apps, potentially opening up new possibilities for personal and experimental development.

The Verge

Actress Mia Goth says it’s ‘incomprehensible’ that Instagram allowed an AI image of her to spread

Mia Goth's representatives claim an AI-generated image of her in a T-shirt referencing her divorce was posted on Instagram by a verified fan account, which was later removed. The image spread widely, with many liking and sharing it as if it were authentic. Instagram's policy on AI-generated content allows for still images to be posted without a label.

Why it matters: This incident highlights the challenges of regulating AI-generated content on social media platforms and protecting public figures' likenesses.

Google DeepMind

Gemini 4 Argon for Robotics

Google DeepMind's Gemini 4 Argon model has been shown to excel in physical understanding, code generation, and visual-semantic reasoning, outperforming other frontier models on the Embodied Reasoning benchmark. Argon can be applied across the robotics lifecycle and has been demonstrated in various tasks such as controlling robots, creating digital twins, and manipulating objects. The model has been showcased in several videos demonstrating its capabilities in different scenarios.

Why it matters: This development has significant implications for the field of robotics and artificial general intelligence (AGI), as it brings us closer to creating robots that can understand and interact with the physical world.

Google DeepMind

Gemini 4 Argon: Built to solve your hardest problems

Google DeepMind has released Gemini 4 Argon, a frontier model for complex tasks in various domains, including coding, finance, and science. The model is available for developers and enterprises through the API and Google AI Studio. Argon outperforms other frontier models on benchmarks and has been used by Google and other teams to achieve significant results.

Why it matters: This release provides a powerful tool for professionals and organizations to tackle complex tasks and improve their workflows.

Ars Technica

Gemini 4 Argon begins rolling out in Gemini app and Antigravity, but only if you pay

Google has begun rolling out Gemini 4 Argon, a new AI model, in the Gemini app and Antigravity, but only for paying users. The model has been trained to refuse malicious instructions and detect unintended workflows. The rollout is restricted to users on AI Pro and AI Ultra plans.

Why it matters: Developers and users with higher-tier subscriptions will have access to Gemini 4 Argon's advanced capabilities, such as agentic coding and reasoning.

The Verge

Google finally releases Gemini 4 Argon

Google has released its Gemini 4 Argon model, a frontier AI that competes with industry leaders. The model is available to higher-tier Pro and Ultra subscribers in the Gemini app and AI mode in Search. Google claims Argon outperforms rival models in various benchmarks, including coding, cybersecurity, and finance.

Why it matters: The release of Argon could indicate Google's renewed commitment to competing in the AI space, potentially shifting the balance of power in the industry.

Ars Technica

They war-gamed dimming the Sun. It went even worse than you think.

A war game simulation on solar geoengineering showed that geopolitics can lead to a chaotic and unequal response to the technology, with some regions suffering from its effects. The simulation involved a group of experts from various countries and organizations making decisions on how to govern the technology. The game's results highlighted the need for international cooperation and regulation of solar geoengineering.

Why it matters: The simulation's findings underscore the importance of developing a framework for governing solar geoengineering to prevent its misuse and ensure its benefits are shared equitably.

The Verge

Epic shifts to ‘AI-accelerated development’ for upcoming Disney game

Epic Games is shifting its development approach for an unannounced Disney game to focus on AI-accelerated development, using tools like Anthropic's Claude to generate code and art. The game, a shooter featuring Marvel and Star Wars characters, was originally set for a March release but has been delayed. The project's developers have been instructed to use AI tools to finish the game.

Why it matters: This shift in development approach could set a precedent for the use of AI in game development and potentially change the way games are created in the industry.