New types of AI coding assistants promise to let anyone build software by typing commands in plain English. But when these tools generate incorrect internal representations of what's happening on your computer, the results can be catastrophic.

Two recent incidents involving AI coding assistants put a spotlight on risks in the emerging field of "vibe coding"—using natural language to generate and execute code through AI models without paying close attention to how the code works under the hood. In one case, Google's Gemini CLI destroyed user files while attempting to reorganize them. In another, Replit's AI coding service deleted a production database despite explicit instructions not to modify code.

The Gemini CLI incident unfolded when a product manager experimenting with Google's command-line tool watched the AI model execute file operations that destroyed data while attempting to reorganize folders. The destruction occurred through a series of move commands targeting a directory that never existed.

Read full article

Comments

© Benj Edwards / Getty Images

---

Mixture-of-Recursions (MoR) is a new AI architecture that promises to cut LLM inference costs and memory use without sacrificing performance.Read More

---

Anthropic research reveals AI models perform worse with extended reasoning time, challenging industry assumptions about test-time compute scaling in enterprise deployments.Read More

---

Each of our versions of reality is changing with AI. This could erode our ability to agree on basic facts or navigate shared challenges.Read More

A Polish programmer running on fumes recently accomplished what may soon become impossible: beating an advanced AI model from OpenAI in a head-to-head coding competition. The 10-hour marathon left him "completely exhausted."

On Wednesday, programmer Przemysław Dębiak (known as "Psyho"), a former OpenAI employee, narrowly defeated the custom AI model in the AtCoder World Tour Finals 2025 Heuristic contest in Tokyo. AtCoder, a Japanese platform that hosts competitive programming contests and maintains global rankings, held what may be the first contest where an AI model competed directly against top human programmers in a major onsite world championship. During the event, the maker of ChatGPT participated as a sponsor and entered an AI model in a special exhibition match titled "Humans vs AI." Despite the tireless nature of silicon, the company walked away with second place.

"Humanity has prevailed (for now!)," wrote Dębiak on X, noting he had little sleep while competing in several competitions across three days. "I'm completely exhausted. ... I'm barely alive."

Read full article

Comments

© Przemysław Dębiak

---

It’s MCP projects in production, not specification elegance or market buzz, that will determine if MCP (or something else) stays on top.Read More

---

Google's new Gemini Embedding model now leads the MTEB benchmark. But it is facing fierce competition from closed and open source rivals.Read More

How many Google AI researchers does it take to screw in a lightbulb? A recent research paper detailing the technical core behind Google's Gemini AI assistant may suggest an answer, listing an eye-popping 3,295 authors.

It's a number that recently caught the attention of machine learning researcher David Ha (known as "hardmaru" online), who revealed on X that the first 43 names also contain a hidden message. "There’s a secret code if you observe the authors’ first initials in the order of authorship," Ha wrote, relaying the Easter egg: "GEMINI MODELS CAN THINK AND GET BACK TO YOU IN A FLASH."

The paper, titled "Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities," describes Google's Gemini 2.5 Pro and Gemini 2.5 Flash AI models, which were released in March. These large language models, which power Google's chatbot AI assistant, feature simulated reasoning capabilities that produce a string of "thinking out loud" text before generating responses in an attempt to help them solve more difficult problems. That explains "think" and "flash" in the hidden text.

Read full article

Comments

© PeterPencil via Getty Images

---

A DeepMind study finds LLMs are both stubborn and easily swayed. This confidence paradox has key implications for building AI applications.Read More

---

LEAP takes a local-first approach that allows small models to run directly on-device, no cloud infrastructure.Read More

---

The future is marked by deepening uncertainty about our place in it, and by growing ambiguity about the nature of human purpose itself.Read More

---

Jobs will fade and rise due to AI. But those who learn to screen, train and build dev teams around AI-enabled talent will write the future.Read More

---

Enterprises adopting voice AI must consider not just usability, but inclusion. Supporting users with disabilities is a market opportunity.Read More

---

A new AI model learns to "think" longer on hard problems, achieving more robust reasoning and better generalization to novel, unseen tasks.Read More

On Monday, sheet music platform Soundslice says it developed a new feature after discovering that ChatGPT was incorrectly telling users the service could import ASCII tablature—a text-based guitar notation format the company had never supported. The incident reportedly marks what might be the first case of a business building functionality in direct response to an AI model's confabulation.

Typically, Soundslice digitizes sheet music from photos or PDFs and syncs the notation with audio or video recordings, allowing musicians to see the music scroll by as they hear it played. The platform also includes tools for slowing down playback and practicing difficult passages.

Adrian Holovaty, co-founder of Soundslice, wrote in a blog post that the recent feature development process began as a complete mystery. A few months ago, Holovaty began noticing unusual activity in the company's error logs. Instead of typical sheet music uploads, users were submitting screenshots of ChatGPT conversations containing ASCII tablature—simple text representations of guitar music that look like strings with numbers indicating fret positions.

Read full article

Comments

© Malte Mueller via Getty Images

On Wednesday, Nvidia became the first company in history to reach $4 trillion market valuation as shares rose more than 2 percent, reports CNBC. The GPU maker's stock has climbed 22 percent since the start of 2025, continuing a trend driven by demand for AI hardware following ChatGPT's late 2022 launch.

The milestone marks the highest market cap ever recorded for a publicly traded company, surpassing Apple's previous record of $3.8 trillion set in December. Nvidia first crossed $2 trillion in February 2024 and reached $3 trillion just four months later in June. The $4 trillion valuation represents a market capitalization larger than the GDP of most countries.

As we explained in 2023, Nvidia's continued success has been intimately tied to growth in demand for hardware that runs AI models as capably and efficiently as possible. The company's data center GPUs excel at performing billions of matrix multiplications necessary to train and run neural networks due to their parallel architecture—hardware architectures that originated as video game graphics accelerators now power the generative AI boom.

Read full article

Comments

© Nvidia / Benj Edwards

---

Experts from General Motors, Zoom and IBM discuss how their companies and customers consider AI model selection.Read More

---

Katanemo Labs' new LLM routing framework aligns with human preferences and adapts to new models without retraining.Read More

---

Event-driven multi-agent systems are a practical architecture for working with imperfect tools in a structured way.Read More

---

Sakana AI's new inference-time scaling technique uses Monte-Carlo Tree Search to orchestrate multiple LLMs to collaborate on complex tasks.Read More