A week in Generative AI: UI, Midjourney & Meta

News for the week ending 22nd June 2025

Jun 22, 2025

No game-changing news in the AI space this week, but as usual, there are some very interesting reports and some great commentary on where things are heading.

First up there’s Google DeepMind who showed off a new generative UI concept that is a great example of how we can leverage the power of large language models beyond a basic chat interface. Midjourney dropped their first video model, which is the start of longer series of releases that they’re planning in the video space over the next 12 months. Lastly, there was continued fall out from Meta’s $14bn investment in Scale AI last week.

In Ethics News there are more reports of Meta’s AI models leaking sensitive user information, a report on how some models use more energy to answer the same questions, and research into how models resort to blackmail in an attempt at self preservation.

Lots of Long Reads this week too. I highly recommend Andrej Karpathy’s talk at Y combinator, and OpenAI’s first podcast if you’re interested in hearing about where the company is heading next.

Google DeepMind shows off a new generative UI concept

One of my favourite analogies of generative AI interfaces over the last couple of years is that we’re currently in the ‘command line’ era - essentially the same as PC’s were in the late 70s/early 80s before Apple launched the first consumer facing graphical user interface in 1983. Large Language Models are an enormously powerful technology, but they’re currently stuck behind incredibly simplistic chat interfaces that holds back their capabilities.

That’s why I was so pleased to see this work from Google DeepMind. Many commentators have suggested that future user interfaces will be ‘built on the fly’ depending on the user context, but this is the first time I’ve actually seen this demonstrated.

There are definitely pros and cons - for example, I don’t want a different view of my documents every time I open the documents folder. Maybe they’re ranked differently, based on my context, but overall I want to see all files and folders, not just a subset. Familiarity and consistency are hugely important for a successful user interface. But, there are some really interesting ideas here and I do think we need to see a paradigm shift in user interfaces to get the most out of large language models over the coming years.

If you’re interested in hearing more about how generative AI models are changing software and the user interface, check out the great talk Andrej Karpathy did this week at Y Combinator.

Source

Midjourney launches its first AI video generation model, V1

Midjourney was the leading text-to-image model for a long time, and arguably still is depending on whether you’re happy to use something that was trained on unlicensed content. However, they’ve seen a lot more competition recently from new image startups, and advances from both OpenAI and Google DeepMind in image generation.

So it was only a matter of time before they released a text-to-video model and from what I’ve been able to test it seems very similar and on a par with other text-to-video offerings from other platforms. Competition is good though, it will drive capabilities forward, give consumers more choice and drive prices down over time.

In their release post, Midjourney say that this is the first step towards more accessible, more user friendly video generation capabilities that will include image, videos/animation, 3D models, and real-time generation. It’ll definitely be interesting to see where they take this over the next 12 months as they have some big ambitions!

Source

Mark Zuckerberg, chief executive officer of Meta Platforms Inc.

More fallout from Meta’s push for AGI

Following last week’s news of Meta investing $14bn in Scale AI as part of Zuckerberg’s push to try and lead the race to AGI, OpenAI has dropped Scale AI as a data provider, alongside Google’s similar decision last week.

This week also saw reports emerge that Meta also tried to acquire Perplexity, Mira Murati’s Thinking Machines, and Ilya Sutskever’s Safe Superintelligence. These acquisition attempts aren’t about any technology, but purely about talent and Zuckerberg trying to build a team of the best AI scientists in the world. For example, Sam Altman has also said that Meta tried and failed to poach OpenAI’s talent with $100M offers.

It looks like Zuckerberg is going all-out to try and tempt the brightest AI talent to Meta, but very few are interested. The numbers being thrown around are crazy and it almost appears as if Zuckerberg thinks this is a do-or-die situation for Meta.

AI Ethics News

Long Reads

Andrej Karpathy - Software is changing (again)
OpenAI Podcast - Sam Altman on AGI, GPT-5, and what’s next
Simon Willison - The lethal trifecta for AI Agents
Ars Technica - A History of the Internet, Part 2
Reid Hoffman - Move fast and make things: the new career mantra

“The future is already here, it’s just not evenly distributed.“
William Gibson