February Roundup
Welcome to February’s monthly roundup - a monthly newsletter, covering papers, events and other reading I’ve done during each month. I also share on LinkedIn and Bluesky.
This month I was fortunate enough to attend the opening of Bletchley Park’s new exhibit - The Age of AI. If you’re nearby, definitely go and check it out (and learn more about the history of Bletchley Park at the same time!).
Work with me! I work with organisations who are building AI - as an advisor, coach and consultant. Get in touch to explore working together.
Articles & other Links
Foundational Speech Models and Their Efficient Training with NVIDIA NeMo [video]
Communications and Digital Committee “AI and creative technology scaleups: less talk, more action”
AI chatbots distort and mislead when asked about current affairs, BBC finds with the original research report here.
AI Updates
DeepSeek in the News
DeepSeek were in the news at the beginning of the month, after releasing their R1 LLM. It performs similarly to OpenAI’s o1 on some benchmarks, but at a fraction of the training cost. At a time when tech giants are pushing for AI infrastructure and compute, this challenges the status quo.
Unlike Claude or ChatGPT, which require subscriptions and keep their models closed, DeepSeek released its model for free—both as a download and an app. However, like many other AI models, there’s little transparency around what data DeepSeek used or how it handles biases and safety measures.
In many ways, this launch isn’t surprising. AI progress often comes from making things cheaper, faster, and more scalable. There’s already momentum in research toward smaller, more efficient models—and open-source AI innovation continues to accelerate progress.
Paris AI Action Summit
The Paris AI Action Summit took place Feb 10-11th; discussions were focused around how to take advantage of the potential of AI - without widening inequality between those who control, and those who use, AI.
These concerns arise as we’re seeing the latest AI models begin to automate some of the knowledge based work that previously technology couldn’t help us with, and that’ll ultimately have a big impact on how people work in the future.
One crucial way to keep AI accessible and inclusive is open source. Much of today’s rapid progress in AI has been driven by open collaboration, where models and code are shared, enabling more people to contribute, innovate, and build on existing breakthroughs.
As AI evolves, openness will be critical to continuing this global conversation on AI governance.
Papers I’ve read
Fully Autonomous AI Agents Should Not be Developed
Find the paper: https://arxiv.org/abs/2502.02649
As the title says, this paper argues that fully autonomous AI agents shouldn’t be developed.
The idea of autonomous agents has been talked about for years, and now the latest iteration of AI tools are promising AI agents that can plan and act. We’re in an era where there’s a sliding scale of autonomy - from fully human-controlled to fully autonomous.
There are potential benefits to autonomous systems, like increased helpfulness and relevance, but also new and increased risks that include:
Risks based on the complexity of agents, meaning that errors due to accuracy and consistency of the underlying models cascade
Security breaches and data privacy issues as we need to give more and more information to agents if they are to act autonomously on our behalf
Loss of control as agents do more of the decision making
The paper ultimately argues that “the more autonomous the system, the more we cede human control” and that we should focus on semi-autonomous systems where humans retain some level of control.
The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers
Find the paper: Microsoft
How does using generative AI affect critical thinking?
This paper looks at almost 1000 examples of genAI usage and finds that a user’s confidence in their own ability to do a task is associated with more critical thinking. Conversely, a user’s confidence in genAI’s ability to do a task is associated with less critical thinking.
The kinds of task that users do when using genAI also differs. Without genAI, a user might have to gather information together, digest, and create written summaries. With genAI, users instead have more of a role of verifying information gathered by a genAI tool, and integrating its responses into their work appropriately.
These findings have implications for the design of genAI tools, and for training people to use them, that really help people do their work in the best way.
Thanks for reading! See you next month,
Catherine.