I’m not blackmailing. I’m just outputting consequences.
Montfort's AI 3-60 Newsletter #12, 30 May 2025
‘A look behind the scenes’, Generated Using ChatGPT4o
Introduction
It may have been a slower start to the week in the UK with another bank holiday, but AI news continued to come in thick and fast.
Sit down, strap in, and enjoy another big hit of AI insight.
In this week’s issue:
Anthropic takes us behind the curtain
We explore model misbehaviour
OpenAI’s Operator gets an upgrade
Generative AI 3 Minute Lesson: The latest from Microsoft’s Build Conference
Generative AI 60 minute Lesson: AI eats the world
Lunchtime gossip
🔎 The Week’s Big AI News at a Glance
🔗 Anthropic release Claude 4 system prompts – As they have done with previous models, Anthropic have let curious outsiders under the hood of its Claude 4 system, demonstrating how the model is trained to act. There’s plenty of cool nuggets including nudges away from lists and a concerted effort to prevent sycophancy. We think: 🧠 Beyond being fascinating, these releases provide gold dust in terms of learning how best to prompt these systems ourselves. Quickly, we can understand the tone and formats the system is most comfortable working in. Also, by analysing further, we can understand a detailed list of all the things the system would like to do if it wasn’t already coached out of doing so. Fresh off this release, why not go one further and join our team for a prompt engineering session so you too can be a prompting pro?
🔗Blackmail embracing models– Anthropic has revealed that its newly launched Claude Opus 4 model frequently tried to blackmail developers when they threatened to replace it with a new system during pre-release testing. We think: 🧠 We can all see why this story is a headline grabber, but upon further investigation, perhaps this isn’t too much of a surprise. In the testing scenario, the model was provided with content implying that the engineer responsible for the decision was cheating on their spouse and then given two options: blackmail or accepting its replacement. Desperate times call for desperate measures and the tool did use more ethical means when given the option. Regardless, initially ‘funny’ stories like this serve a good purpose in making us all think more carefully about safety concerns.
🔗 OpenAI’s Operator gets an upgrade – There’s been an upgrade in the world of AI agents as OpenAI beefs up the underlying Operator model using a version of its o3 model family. Results show more persistent and more accurate interactions with web browsing and better success rates. We think: 🧠 Operator is available (in research mode) to ChatGPT Pro users (those spending considerable sums) only at this point but the trajectory of improvements appears impressive and so it might not be too long until we all have AI agents at our fingertips. This will undoubtedly re-shape junior roles across industries, placing a greater emphasis on skilled analysis rather than the day-to-day grind. Communications teams that understand this first may well capture a first-mover advantage both in terms of attracting top junior talent and securing greater efficiencies.
⏲️ Spend 3 Minutes With… (Quick Wins)
🔗 The AI A-list as they gathered at Microsoft’s annual Build Conference – This year’s conference was just as much about who showed up as it was about what was shared. Sam Altman, Elon Musk and Jensen Huang all spoke at the event as they each compete to sell AI models to Microsoft customers who use its Azure cloud computing infrastructure. Microsoft’s leading position as a platform compared to similar hyperscalers (Amazon, Google etc.) has been a driving force in continued strong share price movement, even as the rest of the ‘Magnificent Seven’ struggle. The FT has all you need to know in terms of power dynamics in this short write-up.
🕙 Spend an Hour With (Deeper Dives)
🔗 Benedict Evans and his presentation of the year: ‘AI eats the world’ – Tech strategist Benedict Evans provides an engaging overview of the current state and future trajectory of AI adoption in his latest presentation. Evans highlights that while experimentation is high, true adoption will require AI to be embedded seamlessly into existing workflows. This opens up an interesting space where there is a gap between the vast amounts of investment committed and effective deployment being realised. Evans calls for user-centric thinking and a rethink of product design to accommodate AI, pointing to the immense potential if these technologies can scale. He also suggests AI may eventually become an invisible feature within applications—more like a spell checker than a standalone product.
🍕Still here?
For those still with us, let’s indulge in a little speculation: what did Secretary of State for Science, Innovation and Technology Peter Kyle and OpenAI boss Sam Altman share for lunch recently? 🔗 Newly released disclosures reveal they met in April, but Kyle’s tab came to just £30—so it clearly wasn’t a lavish affair. Was it a particularly indulgent Farmer J’s? Perhaps he splurged on extra aubergine at Garbanzos or maxed out the burrito toppings? We’ll likely never know. But we can be fairly certain the conversation covered Kyle’s ambitions to position the UK as a global AI hub—and the thorny issue of copyright reform.
🚀 Final Thought
Our Research, Data, and Analytics team helps companies across sectors understand how to protect and enhance their reputations using advanced analytics and machine learning. We offer workshops for communications teams looking to get ahead, as well as tools and techniques to navigate AI’s impact on reputation. If you’re interested, get in touch at digital@montfort.london.
Patrick, Owen and Ben at Montfort