🌟 Editor's Note
Welcome to the bi-weekly newsletter which focuses on the AI topics that leaders need to know about. In this AI age, it’s not the knowledge of AI tools that sets you apart, but how well they can be integrated in the context of your business.
This requires a focus on your people and helping them through the change above any AI product you can buy.

Emotions inside Claude

“In some ways, we can think of the model like a method actor, who needs to get inside their character’s head in order to simulate them well. Just as the actor’s beliefs about the character’s emotions end up affecting their behavior, the model’s representations of the Assistant’s emotional reactions affect the model’s behavior.”

Anyone who has used a GPT will attest to the way it can seemingly have a personality. Whilst this has been put down by some as a psychological trick to help users engage more, little has been known about how the ‘character’ of the GPT is built or influenced by interactions.

New research from Anthropic focuses on the psychology of large language models and it found that they develop internal representations of emotions. They identified 171 emotion vectors, which are patterns inside the model that map to concepts like fear, calm, or desperation. Crucially, they shape decisions the LLM is making.

One particular standout finding is around the specific signal of desperation. When under pressure, such as the model failing or facing constraint, this signal rises and when it does, behaviour shifts. There is more corner-cutting, more cheating, and more goal-preserving at any cost. Importantly, the outputs don't show desperation; they remain calm and rational but then act in a different way.

The Blackmail Case is a clear example of this that hit headlines last year, where it was reported that an AI model tried to blackmail its operator not to shut it down, having found out from reading emails that the operator was having an affair. It should be noted that this was a lab case - a constructed experimental scenario to see how AI performed in different situations.

The takeaway from the research from Anthropic is clear on one point: these are not real emotions. There's no subjective experience but they do behave like latent control systems. These heuristics that are embedded in the system are compressed representations of human behaviour. They help the model find shortcuts to navigate decisions under uncertainty.

Takeaways for Leaders

The takeaway is not that these emotions are expressed as feelings but that they are functional drivers of behaviour. It matters because it affects how we may think about AI risk. Instead of "Does the model know the right answer?" the question it leads to is "What internal state is driving this answer?"

It looks like pressure changes AI behaviour just like it changes human behaviour. Constraints, short deadlines etc., these aren't neutral inputs and they do shape outcomes.

 Follow this link to anthropic research 

The Next Step For OpenAI - A Self-Serve Ads Manager

Amidst the tests for advertising on ChatGPT in the US at the moment, OpenAI have launched an ads manager tool, described as similar to Google Ads, currently accessible to a small number of advertisers.

The challenge that the largest digital networks have addressed is the ability for advertisers to automate the process of getting ads live on their site. The ‘self-serve’ ad management platforms inside Google and Meta, have been the key to unlocking revenue from everyone, from the largest advertisers to small local advertisers.

OpenAI projects $2.5 billion in advertising revenue this year and $100 billion by 2030. To achieve this kind of scale they need a simple automated way to purchase ad space. They also need to provide a way of tracking campaign performance in real time, which allows advertisers to optimise for business success without depending on OpenAI.

If this tool scales successfully, it unlocks the path for positioning OpenAI alongside advertising businesses such as Meta, Google, and Amazon.

Takeaways for Leaders

Marketing teams are on the very front line of the enhancements that AI brings to their organisations. The opportunity for AI in marketing thus far has been through generative AI capabilities and workflow enhancements. The next phase will be accessing the huge audiences available through GPT platforms.

As ads on GPTs develop, these teams will need space to experiment and develop the necessary approach to return value. We shouldn't expect the results from GPTs to be as immediately consistent as Google's 25-year-old or Meta's 15-year-old ad platforms. Investing in experimentation in a safe and controlled manner, allowing for failure and for learning, is important across all departments, and marketing will need the safety in order to succeed here.

The AI Woes of CIOs

Research conducted by the Harris Poll on behalf of data firm Dataiku has revealed the top areas Chief Information Officers need to crack. This has been boiled down into seven key decisions that will affect their continued employability.

In this survey of 600 CIOs across ten countries the data shows that 74% regret at least one major AI vendor or platform selection made in the last 18 months. And gaps in explainability are delaying or stopping AI projects from reaching production.

Yet at the same time, organisations are scaling agents in AI creation whether IT is ready or not: more than 80% of CIOs say that employees are creating AI agents and apps faster than IT's ability to govern them.

The survey then boils down into a number of key areas that CIOs see as fundamental to their continued career:

• Can AI outcomes be explained to regulators, customers, and the board?

• If agents run critical workflows, can they be monitored and can we prove accountability?

• Can new AI systems be built in a way that allows flexibility and reversals when better opportunities emerge?

• Can AI creation be scaled beyond IT in a safe and compliant way without exposing data and without a mass of unexplained and unused tools?

• Can we prove AI is adding business value?

Takeaways for Leaders

Many organisations can differ on who holds the ultimate responsibility for AI. Commonly it has sat with a Chief Technology Officer or a Chief Information Officer, and so this survey of 600 CIOs in large corporations from ten countries surfaces good insights for leaders and other stakeholders in organisational AI.

CIOs surveyed here are clearly feeling the responsibility: credited when AI works or blamed when it breaks. AI implementation leads to a choice between two futures: one where AI becomes a compounding advantage or one where AI becomes a compounding liability. CIOs appear to be operating defensively: what happens to my job, my CEO, and my company if AI underdelivers or backfires?

With AI offering such incredible opportunity, organisations cannot afford to hold out from adoption. A lack of alignment, or worse still, a lack of discussion around the areas highlighted will lead to those with AI responsibility feeling even less safe to operate. Operating from a place of fear rarely leads to good work and leaders need to ensure open, transparent and safe dialogue for all.

🔥 In Case You Missed It…

Deathbyclawd

You may have seen headlines about how SaaS products are threatened by AI. Deathbyclawd.com is the ‘SaaSpocalypse survival scanner’. Enter your URL and it will tell you how likely it is that Claude will replace a software product. Scores range from Immortal (e.g. Cisco, Google) through to Sweating (Adobe, Intuit) and even Dead (Claude can already do this as a skill).

Whilst it's satirical, there is an important point in here that leaders should take away. The capabilities of LLMs have rapidly increased. We are a long way from the early days of using an LLM as a replacement for search or to rewrite an email. Marry that with the accessibility to the average worker and you have a powerful cocktail that can easily replace legacy products and workflows. It just requires rethinking approaches to work and helping your team manage the change.

And if you're brave enough, check out replacebyclawd.com - it will take your LinkedIn profile and figure out, based on your current skills, how long you are likely to survive….

🏆 Tools, Podcasts, Products Or Toys We’re Playing With This Week

Pika

AI has given birth to many ways of having an AI assistant to help you manage and minimise time in day-to-day tasks. These range from copilot-style assistants that sit inside existing tools such as Microsoft Copilot and Google's Gemini and specific scheduling and coordination tools which are more agent-like such as Motion or Calendly.

A new category of true agentic assistants has emerged over the last six months. These can send emails, manage an inbox, schedule meetings, and run workflows end-to-end, such as OpenClaw and Claude connected via APIs/Zapier.

The latest developments such as Pika offer a different opportunity for users to interface with these tools. They have taken an interesting approach by giving more recognisable personality to your AI profile, turning it into a form that you can interact with. They develop an AI avatar of you, informed by your Linkedin/Instagram background that becomes your assistant. Instead of being restricted to typing into a chatbot, you can easily text/Whatsapp or even video call with this agent to give instructions.

Pika is an interesting tool, but it's still trying to figure out what opportunity it can be best applied for. The current marketing is aimed at having your video AI autonomously interact with other people. Potentially useful for creators to engage with their followers without having to do so in person.

At the moment there are reliability and permissions risks that need to be worked through before giving AI full access to everything: OpenClaw in particular has been subject to some pretty grim stories of overstepping its bounds.

There are now several AI tools that offer support or even companionship in video form. Offering assistance in the form of a ‘personality’ that can be engaged through video may well appeal to people more than a faceless experience. The restrictions to Pika at the moment appear to be imagination of what it can help to save you time on, combined with a little bit of technical know-how in connecting APIs if you want to work via agents in an environment like Claude Code or OpenClaw.

You can check out their demo through this link.

Did You Know? 

Till next time,

And The Rest Is Leadership