Researchers found that models make up facts because training test scoring gives full points for lucky guesses but zero for saying "I don't know."

The paper shows this creates a conflict: models trained to maximize accuracy learn to always guess, even when completely uncertain about answers.

OpenAI tested this theory by asking models for specific birthdays and dissertation titles, finding they confidently produced different wrong answers each time.

Researchers proposed redesigning evaluation metrics to penalize confident errors more than when they express uncertainty explicitly.

Prime Minister Edi Rama unveiled Diella during a cabinet announcement this week, calling her the first member “virtually created by artificial intelligence”.

The AI avatar will evaluate and award all public tenders where the government contracts private firms.

Diella already serves citizens through Albania's digital services portal, processing bureaucratic requests via voice commands.

Rama claims the AI will eliminate bribes and threats from decision-making, though the government hasn't detailed what human oversight will exist.

Authors sued after discovering Anthropic downloaded over 7M pirated books from shadow libraries like LibGen to build its training dataset for Claude.

A federal judge ruled in June that training on legally purchased books constitutes fair use, but downloading pirated copies violates copyright law.

The settlement covers approximately. 500,000 books at $3,000 per work, with additional payments if more pirated materials are found in training data.

Anthropic must also destroy all pirated files and copies as part of the agreement, which doesn’t grant future training permissions.

The team hopes to complete the feature in 9 months with a sub-$30M budget, compared to typical animated films requiring 3 years and much higher budgets.

OpenAI plans to provide its ‘tools and computing resources,’ including GPT-5 and image generation models, targeting a Cannes Film Festival debut in 2026.

Human actors will voice characters, with artists creating initial sketches that GPT-5 and image models transform into the final animations.

OpenAI creative specialist Chad Nelson developed the characters over three years, initially using DALL-E to create the first iteration of the film.

Claude gains a private computing environment where it can write and run code to create files, turning data into reports with charts, analysis, and formulas.

Users can also convert between formats, turning PDF reports into slide decks or meeting notes into formatted documents.

The feature is rolling out to Max, Team, and Enterprise subscribers, with Pro users gaining access in the coming weeks.

Anthropic recommended starting with “straightforward” tasks, also warning of potential data risks with giving Claude access to the internet.

Alibaba introduced ‘Qwen3-Max’, a 1T+ model that surpasses other Qwen3 variants, Kimi K2, Deepseek V3.1, and Claude Opus 4 (non-reasoning) across benchmarks.

AI startup Runway is expanding beyond creative AI tools into robotics, using its video-generating world models to help train robots and self-driving cars.

Higgsfield AI unveiled ‘Ads 2.0’ — a new tool it says can replace entire production, marketing, and creative teams with a Mini App.

Waymo is expanding its robotaxi service to Denver and Seattle, planning to test autonomous driving in harsher weather conditions.

Google introduced learning tools to ‘NotebookLM’, including customizable flashcards and quizzes, redesigned reports, and interactive notebooks for academic textbooks.

OpenAI will begin mass production of its own custom AI chips next year through a partnership with Broadcom, according to a report from the Financial Times — joining other tech giants racing to reduce dependence on Nvidia's hardware.

Microsoft is set to reach a deal to integrate Anthropic's AI models into Office 365 alongside OpenAI's tech, according to The Information, marking the software giant's first major AI diversification away from its OpenAI partnership.

ByteDance Seed introduced ‘Seedream 4.0’, a new image generation and editing model that competes with Nano Banana with 4K outputs and multimodal capabilities.

Google added new capabilities to its Veo 3 and Veo 3 Fast models, including vertical video outputs, 1080p HD resolution, and a 50% decrease in price.

Mira Murati’s Thinking Machines Lab published research on making AI responses more consistent and predictable, taking its initial step toward sharing work openly.

ElevenLabs introduced ‘Voice Remixing’, a new feature allowing users to change aspects of both designed and user voices for more customizable outputs.

Math Inc. just introduced ‘Gauss’, an AI system that solved a complex mathematical theorem called the Strong Prime Number Theorem in just three weeks — after top mathematicians struggled with it for 18 months.