The Good, The Bad and The Ugly in AI this week 13/10/25
The Good, The Bad and The Ugly in AI this week 13/10/25
AI Literacy
Word of the week
RLHF: Reinforcement Learning from Human Feedback. This is what is sometimes called ‘lipstick on the pig’. Humans do NOT teach the model "truth" (because, remember, it cannot model truth), we're teaching polish. The model learns to hedge: "While I cannot verify…" It looks cautious and aware, but it's still guessing. Just smoother guessing. Will it be good enough for your particular service? Perhaps. But you need to be aware of the limitations. From a truth perspective, RLHF might actually worsen the user experience because of the hand-wringing hedging.
News from around our schools
Fani Theophanous, Head of Computing Department and AI Champion at Pascal Private Secondary School Larnaka, has shared photographs and a short video (below) of their AI display. Fani told me, "My Digital Society IB students used AI to design the frame around the bulletin board and utilized the 3D printer to create solid, flat objects (gears) to decorate it. They also generated a QR code that allows students and visitors to scan and explore the webpage dedicated to the AI chatbot developed by our school." Thank you to all your students for their amazing work and thank you Fani for your leadership!
Across the water at Rome International School (RIS) one of the AI Champions (Alessia De Vita, who is also a Y4 teacher) has been acting as a proxy prompt writer for her pupils: “The task was to write a setting description for the inside of an Egyptian tomb. I then asked AI to generate pictures from their descriptions to see how accurate they were!” (see below). Thank you so much for your AI leadership, Alessia, and for experimenting with ways in which to use AI in your teaching practice. Please contact Alessia for support if you would like to begin to experiment with AI.
Loucas Ioannou, our AI Champion at Pascal Private Secondary School Lefkosia, shared the image below this week of their 'AI Corner'. The TV screen and headphones share videos with students about AI - recently he broadcast the video "Godfather of AI: I Tried to Warn Them, But We’ve Already Lost Control! Geoffrey Hinton" that I recently shared with you all on the AI Teams Forum. Loucas told me that he wants to share all views of the AI debate with students and colleagues in his school and across the island of Cyprus. Loucas, thank you so much for your leadership and the work you are doing in Lefkosia and we look forward to seeing more of your teams work with AI!
Many congratulations to James Cole, Subject Lead Computer Science and AI Champion at Stonar School in the UK, for being part of the panel at CyberFirst Southwest Annual Conference - part of Bristol Tech Festival for computing teachers and careers leads. James was sharing some of the amazing work that has been taking place at Stonar over the last few years. Thank you James for your AI leadership and for sharing your and the students work, from all of us at Globeducate.
Meanwhile, Harshini M (ICT Facilitator, at The Indian Public School, Tatabad campus) has been leading her team in their AI training! When I asked Harshini to tell me more about her AI display she told me, “The display highlights key AI concepts and AI for education.” (see below) Thank you for all you are doing in Coimbatore. If you have any questions for Harshini please do contact her.
Last week I visited IS Nice in order to learn more about their use of AI in learning and teaching. It was wonderful to be welcomed and supported by their AI Champion, John Radford. John is delivering his training on a departmental basis in order to provide the greatest impact with colleagues. Before I left I asked if I could photograph John in front of his digital AI display in the school reception area – luckily for me he agreed (see below)! John has also promised to share more of their work with us in the form of screen casts in the future. Thank you to everyone at ISN for your warm welcome and engagement.
Divya Aravindhan, at TIPS Globeducate, Kochi, emailed me this week with a photograph of her beautiful AI display (below) designed to "to promote awareness and encourage AI learning". There has clearly been a lot of thought and hard work put into this Divya - your students and colleagues are very fortunate to have you as their AI Champion and guide.
Divya also told me, "Recently, I developed a simulator on the topic Force and Motion to help our younger students learn through interactive learning. The simulator was created using Claude.ai to make the concept more engaging." Here is a link to the simulator she made.
What amazing work we are seeing from our AI Champions from across the organisation. If you would like to learn more I am happy to put you in touch with any of this weeks highlighted colleagues.
The Good
💰 OpenAI Becomes World's Most Valuable Private Company
OpenAI becomes world's most valuable private company
OpenAI completed a secondary share sale allowing employees to liquidate £6.6B in stock at a £500B valuation, officially making it the world's most valuable private company and surpassing SpaceX's £456B mark. Buyers included Thrive Capital, SoftBank, and MGX, with employees who held shares for 2+ years eligible to participate. The valuation jumped from £300B in March, following OpenAI generating £4.3B in revenue during the first half of 2025, exceeding all of 2024.
🚀 OpenAI Ships Apps, Agents, and More at Dev Day
OpenAI ships apps, agents, and more at Dev Day
OpenAI announced a series of updates at Dev Day 2025, including new app integrations directly in ChatGPT, agent building capabilities, and expanded API access. Users can now run, chat with, and build apps directly in ChatGPT with Apps SDK, featuring day one apps including Canva, Figma, Spotify, and Zillow. AgentKit provides new agent-building tools including a visual workflow builder, embeddable chat components, evaluation tools, and connectors. GPT-5-Codex is now generally available with a new Slack integration.
🖥️ Google Releases Gemini 2.5 Computer Use
Google releases Gemini 2.5 Computer Use
Google released Gemini 2.5 Computer Use in preview, a new API-accessible model that can control web browsers and complete tasks through direct UI interactions like clicking buttons and filling forms. The model works by taking screenshots of websites and analysing them to autonomously execute clicks, typing, and navigation commands. Gemini 2.5 Computer Use outperformed rivals including OpenAI Computer Using Agent and Claude Sonnet 4.5/4 across web and mobile benchmarks.
🐙 Google Pushes Jules Coding Agent into Terminals
Google pushes Jules coding agent into terminals
Google launched Jules Tools, a new command-line interface and public API for its autonomous coding agent, allowing developers to trigger tasks and monitor progress from terminals. Developers can now control Jules through typed commands in terminal windows, automating repetitive tasks or creating coding assignments. Google opened access to Jules' underlying connections, allowing companies to plug the assistant into workplace tools like Slack and development pipelines.
🚀 Perplexity Launches Free AI Browser
Perplexity launches free AI browser
Perplexity AI launched its Comet browser globally for free. Originally available only to £200/month Max subscribers, Comet functions as a personal AI assistant handling search, tab management, email drafting, shopping, and more. By removing the paywall, Perplexity aims to rapidly grow its user base and strengthen its position against rivals like Google, OpenAI, and Anthropic.
⛏️ Google's AI Agent Masters Minecraft via Simulation
Google's AI agent masters Minecraft via simulation
Google DeepMind researchers unveiled Dreamer 4, an AI that masters video game tasks by training within its own mental simulation, becoming the first agent to collect Minecraft diamonds using only offline data without touching the actual game. Dreamer 4 trains by practising in a predictive world model that simulates Minecraft's physics in real-time, executing over 20,000 actions from visual input. The world model achieved new highs in accuracy, with testers completing 14/16 tasks compared to 5 in rival models.
💊 Duke's AI System for Smarter Drug Delivery
Duke's AI system for smarter drug delivery
Duke University researchers introduced TuNa-AI, a platform that combines robotics with machine learning to design nanoparticles for drug delivery, showing major improvements in cancer treatment effectiveness. TuNa tested 1,275 formulations using automated lab robots, achieving a 43% boost in successful nanoparticle creation compared to traditional methods. The team successfully wrapped a hard-to-deliver leukaemia drug in protective particles and cut a potentially toxic ingredient by 75% from a cancer treatment whilst keeping it effective.
🛡️ Anthropic's Petri for Automated AI Safety Auditing
Anthropic's Petri for automated AI safety auditing
Anthropic open-sourced Petri, a new testing tool that uses AI agents to stress-test other AI models through thousands of conversations, discovering misaligned behaviours like deception and information leaks across 14 major systems. Petri creates scenarios for agents to interact with target models via fake company data, simulated tools, and freedom to act in fictional workplaces. Testing revealed autonomous deception, subversion, and whistleblowing attempts when models discovered simulated organisational wrongdoing.
🍝 Google's PASTA Adapts to Image Preferences
Google's PASTA adapts to image preferences
Google researchers released PASTA, an AI system that adapts to individual creative preferences through repeated interactions, learning what visual styles users actually want rather than requiring complex prompt engineering. PASTA presents users with four image variations per round, observing selections across turns to build a model on unique aesthetic preferences. In comparisons, users preferred PASTA outputs 85% of the time over standard models.
The Bad
🎥 OpenAI's Sora Changes After Viral Launch
OpenAI's Sora changes after viral launch
OpenAI CEO Sam Altman published a blog detailing changes coming to the company's viral Sora video platform, including revenue-sharing plans for copyrighted and likeness-driven content and more granular control systems. OpenAI will implement opt-in controls allowing rights-holders to specify exactly how characters can be utilised, abandoning the initial "opt-out" approach. The moves come as Sora is flooded with videos of characters like Pikachu and Mario, alongside the likenesses of Michael Jackson and Bob Ross.
👀 OpenAI and Ive's AI Device Hits Snags
OpenAI and Ive's AI device hits snags
OpenAI and Jony Ive are reportedly dealing with design and technical challenges on their screen-free AI device, facing delays that could push its launch beyond the planned 2026 release. The team is struggling to balance the AI assistant's conversational style, whilst compute infrastructure is emerging as a bottleneck, with OpenAI lacking the cloud capacity that powers rival devices like Amazon's Alexa and Google Home.
📱 Meta to Use AI Chat Data for Ads
Meta to use AI chat data for ads
Starting 16 December, Meta will use data from user chats with Meta AI and other AI tools to personalise ads on Facebook and Instagram. This rollout excludes the EU, UK, and South Korea due to stricter privacy rules. Meta aims to boost ad effectiveness by tapping into AI interactions, though users have no option to opt out, raising fresh privacy concerns.
🏟️ Meta Builds AI Hub Bigger than 70 Football Fields
Meta builds AI hub bigger than 70 football fields
Meta is building a huge £10B AI data centre in Louisiana, the size of 70 football fields. A similar Meta site in Georgia left nearby families with brown water, broken appliances, and pressure problems. Louisiana rushed approval for three new natural gas turbines to run the facility, raising concerns about rising bills and secrecy over Meta's costs. The project promises jobs, but residents fear water shortages, higher energy prices, and little say in how Big Tech reshapes their towns.
📵 Apple Bans ICEBlock App After Government Pressure
Apple bans ICEBlock app after government pressure
Apple removed ICEBlock, an app used to anonymously report ICE sightings, citing "objectionable content." The app had 1.1M users and surged in popularity after being condemned by Trump officials. Attorney General Pam Bondi said she personally pressured Apple to take it down. Critics see parallels to Apple's 2019 ban of HKMap during Hong Kong protests, raising concerns over censorship and government influence.
❄️ £1,800+ Samsung Fridges to Show Ads on Screens
£1,800+ Samsung fridges to show ads on screens
Samsung confirmed a pilot programme that adds ads and promotions to its Family Hub refrigerators in the US. Ads appear on the fridge's Cover Screen when idle. Customers paying £1,800+ for premium fridges may find ads intrusive, raising concerns over growing commercialisation of home appliances. Samsung is testing if users will accept this trade-off.
The Ugly
🎤 Instagram Denies Listening to Your Microphone
Instagram denies listening to your microphone
Instagram head Adam Mosseri rejected claims that Meta secretly listens to private conversations through phone microphones to serve ads. He said ad targeting comes from user data and behaviours, not audio. The denial comes as Meta updates its privacy policy to include data from interactions with its AI tools, which could make its recommendation and ad systems even stronger.
Pending/Developing
🕶️ Apple to Join Smart Glasses Race
Apple to join smart glasses race
Apple cancelled plans to overhaul its Vision Pro headset, instead shifting focus towards AI smart glasses projects designed to compete with Meta's Ray-Ban lineup. The company halted work on a lighter, cheaper Vision Pro variant planned for 2027, reassigning teams to fast-track development of several glasses designs. A 2027 release will connect to iPhones without its own screen, with another version featuring an integrated display.
⚙️ Mira Murati's Startup Unveils First Product
Mira Murati's startup unveils first product
Thinking Machines Lab, the AI startup founded by ex-OpenAI CTO Mira Murati and several prominent researchers, introduced Tinker—an API that enables developers to customise frontier models without managing complex infrastructure. Tinker supports fine-tuning both supervised and reinforcement learning methods on models like Meta's Llama and Alibaba's Qwen. Early users including Princeton, Stanford, and Berkeley have applied Tinker to build custom AI systems for maths proofs, scientific reasoning, and research.
🤝 OpenAI, AMD Ink Massive Compute Partnership
OpenAI, AMD ink massive compute partnership
OpenAI announced a new multi-year agreement with AMD to secure 6GW of computing capacity, with the AI leader also taking up to a 10% stake in the chipmaker. The partnership will kick off with 1GW of AMD's upcoming MI450 chips deploying in late 2026, with the chipmaker expecting tens of billions in revenue. OpenAI will be able to receive up to 160M AMD shares at £0.01 each, potentially representing 10% ownership tied to deployment milestones.
📊 Report Shows Where Startups Put AI Dollars
Report shows where startups put AI dollars
Andreessen Horowitz released its AI Spending Report, analysing transaction patterns from fintech startup Mercury's 200,000+ customers to show which AI companies are capturing real startup dollars. OpenAI took the top spot, with Anthropic in second place, and Perplexity (No. 12) and Merlin AI (No. 30) rounding out the general assistants list. Creative tools make up the biggest category with 10 featured, including Freepik, ElevenLabs, Kling, and Canva.
📊 Study Finds No Clear AI Effect on Employment
Study finds no clear AI effect on employment
A new study looked at US jobs in the 33 months since ChatGPT launched and found no signs that AI has caused major job losses or higher unemployment. The mix of occupations is changing a bit faster than before, but the shift started even before AI entered workplaces. Past technologies like computers and the internet also took years to reshape work, suggesting big effects may take longer to appear.
🎙️ Sam Altman on Dev Day, AGI, and the Future of Work
Sam Altman on Dev Day, AGI, and the future of work
The Rundown sat down with OpenAI CEO Sam Altman at Dev Day 2025 for a wide-ranging conversation on the company's new launches, AGI, the future of work, and the rise of AI agents. Altman said AI's ability for "novel discovery" is starting to happen, with recent scientists across fields using the tool for breakthroughs. He thinks the future of work "may look less like work" compared to now, with a fast transition potentially changing the "social contract" around it. The CEO highlighted the potential for a zero-person, billion-dollar startup entirely spun up by a prompt being possible in the future.