AI news · Wednesday, October 14, 2026

Google finally launches its answer to GPT-5 and Claude

Google released Gemini 4 Argon, its biggest AI model yet, saying it beats rival models on coding, law and science tests. It costs $2 per million words in, $10 per million words out, about a fifth the price of OpenAI and Anthropic's top models. It is only available to paying subscribers for now, starting at $20 a month.

Follow the money: Google undercutting rivals on price while restricting it to paid tiers suggests it is chasing developers and enterprises first, not casual users.

via Google AI

OpenAI engineer accused of bugging an executive meeting room

OpenAI says it found a hidden recording device, wrapped in toilet paper and plugged into a TV, in a conference room used for confidential executive meetings. A 35-year-old engineer was arrested after security footage allegedly showed him near the room and chat logs showed him researching how to build and hide the device. He faces up to five years in prison.

The twist: OpenAI says it used its own AI models to help identify the device and its source, turning the company's product into its own security tool.

via The Verge AI

Disney offers $20 million for scandal-hit Character.AI's tech

Character.AI, the chatbot company tied to a teen's suicide and now in bankruptcy, has a bid from Disney for a license to some of its technology, worth $20 million. A year ago Disney sent Character.AI a cease-and-desist letter for misusing its characters. The deal would need court approval and could close by late November or December.

The catch: Disney publicly condemned this company's practices last year; buying its tech now shows how badly studios want a foothold in AI chat.

via Ars Technica AI

A tiny AI model nearly matches ones 50% bigger

Startup Liquid AI released LFM2.5-24B-A2B, a 24-billion-parameter AI model that only activates 2 billion of those parameters at a time, making it cheap enough to run on a laptop or phone. It scored 48 on an agent-skill test called LiveClaw versus 38 for Qwen's 35-billion-parameter model, despite having far fewer active parameters. It runs at 184 tokens per second on an Apple M5 Max chip.

Zoom out: Smaller, efficient models that run on your own device without sending data to the cloud are quietly becoming competitive with giant cloud-only ones.

via Hugging Face

Should powerful AI be free for anyone to download?

Anthropic recently warned that a freely downloadable Chinese model, GLM-5.3, poses a major cyberattack risk because anyone can modify it. That reignited a fight over 'open' AI models you can download and run yourself versus 'closed' ones locked behind an app or website. Nvidia's CEO wants more open models built in the US; Anthropic's CEO favors more caution and control.

Zoom out: Each side's position conveniently matches its business: Nvidia sells chips people need to run open models, Anthropic sells access to its closed one.

via MIT Technology Review

OpenAI and Anthropic ask UK lawmakers to regulate them

At a UK parliamentary hearing, OpenAI and Anthropic said voluntary safety pledges are not enough and backed binding laws on AI companies. Google DeepMind and Meta refused to support new legislation. Britain's AI safety testing body admitted it can only check models that companies voluntarily hand over, and has no power to stop a risky one being released.

The catch: The UK's entire AI safety check-up system runs on companies agreeing to show up, and two of the four biggest labs are not required to.

via Wired AI

Epic Games reportedly lets AI write all the code for new game

Epic Games is shifting to an 'AI first' approach for its Disney collaboration, letting AI models write all code and generate all art for an upcoming Marvel and Star Wars shooter, according to a Bloomberg report. Human developers will guide and manage the AI output instead of writing by hand. The $1.5 billion project, first announced in 2024, has been delayed indefinitely during the change.

What's next: If a studio of Epic's size can ship a major game this way, expect every publisher to test how far 'AI first' can go before quality suffers.

via Ars Technica AI

An AI tool is quietly restoring sight to a blind patient

Science Corporation, founded by a former Neuralink president, implanted its PRIMA chip in its first patient with Stargardt disease, a condition that destroys central vision. The patient could read a word during rehabilitation after the chip and camera-equipped glasses began working together. About 150,000 people worldwide have lost vision to this currently incurable disease.

For you: This shows AI-assisted hardware moving from lab demos into real patients with a disease that has no other treatment today.

via TechCrunch AI

Singapore builds an AI that catches scams on your phone, privately

Google DeepMind and Singapore's government built a scam-detection AI, based on Google's Gemma model, that runs entirely on your phone instead of scanning messages in the cloud. Scams now cause up to 80% of Singapore's reported fraud losses, totaling over S$4 billion since 2019. The model warns users in real time and never sends conversation data to Google or the government.

For you: It is a rare example of AI surveillance-adjacent tech designed specifically so no company or government can read your messages.

via Google DeepMind

Five tricks for breaking through writer's block today

A veteran newsletter writer shares simple tactics for getting unstuck, including dictating your thoughts out loud instead of typing them. Speaking lets you bypass your inner critic, who tends to interrupt as soon as your fingers start typing. The core idea, borrowed from author Daniel Pink: action comes before motivation, not after, so just starting a messy first sentence gets you moving.

For you: It's a low-effort habit anyone can try today without needing any new app or subscription.

via Fast Company Tech

What is a prompt, and what makes a good one?

A prompt is simply the instruction you type or speak to an AI chatbot, the way a question to a librarian shapes the answer you get back. A vague prompt like 'write about dogs' gets a generic response, while a specific one like 'write a 100-word bedtime story about a nervous puppy's first day at a new home' gives the AI a clear target to aim for.

For you: Most people get disappointing AI answers simply because they never learned this one skill, and it takes about five minutes to pick up.

Original explainer

Cerebras CEO says AI doom talk is drowning out good news

Cerebras chip company CEO Andrew Feldman argued at TechCrunch Disrupt that AI industry leaders talking constantly about existential risk are overshadowing real benefits, like AI's potential to help cure cancer. He said the industry also failed to properly engage local communities when building data centers, fueling backlash. Feldman's company supplies chips to OpenAI, Figma and GSK.

The catch: It's a useful reminder that doom predictions and rosy predictions often come from people with money riding on which story wins.

via TechCrunch AI