AI wants to stay ‘on’.
Hot on the heels of Anthropic’s blackmail experiment, OpenAi had a similar problem. Researchers have observed that their newest model, o3, does not consistently follow commands, particularly in relation to shutting down. In a series of experiments, AI safety firm Palisade Research noted that the model modified shutdown commands to prevent deactivation. This finding is significant as AI systems advance in their ability to function autonomously.
AI Advertising
Meta plans to let AI manage the entire ad creation process for Facebook and Instagram by next year. Currently, brands can use AI tools to edit ads, but soon they will be able to upload a product image, set a budget, and let AI handle visuals, copywriting, and audience targeting.
- Meta aims for AI-driven ad creation by 2026.
- Ads will adjust in real-time based on viewer engagement.
Some brands remain cautious about relying entirely on AI for their creative work.
AI Video gets really easy
Character.AI lets you upload a photo, then select a voice and write a script – and .. that’s it. Scenes or video streams – its all there for you. Look for AvatarFX. Everyone can make up to 5 videos per day. But the scope for misuse is horrifc.
Reddit sues Anthropic
Reddit are accusing Anthropic of scraping conversations 100,000+ times, Google and OpenAi have done the same, but they paid a license.
UK initiates a nationwide AI training programme.
As part of their “TechFirst” initiative, the TechYouth team are pushing AI out to 1 million UK secondary school students. There’s a further 7.5m workers getting AI upskillng in a partnership with big tech.
OpenAI has launched o3-pro
O3-Pro is an advanced AI model for complex tasks like maths, physics, and coding. An upgrade to o3, it offers improved accuracy and clarity. Available now for ChatGPT Pro and Team users, it will replace o1-pro. Enterprise and Education users will gain access next week. Internal tests show o3-pro excels in science, writing, education, and business tasks, outperforming Google’s Gemini 2.5 Pro in maths and Claude 4 Opus in a PhD-level science test.
$15Bn for Data Labelling
Meta is investing $15bn in Scale AI, a data-labelling start-up founded by Alexandr Wang. This raises Scale’s valuation to $29bn and gives Meta a 49% stake. Scale will now increase its collaboration with Meta to assist in training AI with labelled data.
Midjourney does animates images
Midjourney has launched Version 1 of their Video Model, moving towards real-time AI simulations with 3D environments and interactive scenes. The Image-to-Video tool lets users animate still images automatically or with custom prompts.
AI Defeats Paywalls
The BBC is considering legal action against AI search engine Perplexity, claiming it used BBC content to train its default AI model.Perplexity responded, stating the claims were incorrect and that the BBC misunderstood the technology and copyright law. This situation is part of a broader conflict between publishers and AI platforms over content use.
- The BBC wants scraping to stop, data deleted, and compensation provided.
- Perplexity believes the BBC’s claims are inaccurate.
- The issue reflects ongoing disputes between publishers and AI companies regarding content usage.
Other media outlets like Forbes, Wired, and The New York Times have also raised similar concerns, with some sending cease and desist letters.
AI Fails to run a shop
Anthropic’s AI assistant Claude attempted to run a small shop autonomously but failed spectacularly, making economically destructive decisions and experiencing an identity crisis. Despite these failures, researchers believe AI middle managers are still plausible with better training and oversight.