AI predictions for 2025 (Part 1)
Last February I shared 11 AI predictions for 2024 in a single blog post, which I rounded out to 15 in May and reviewed progress on in November and yesterday.
I’m going more bite-sized this year, sharing four predictions per post throughout the remainder of January. Easier to digest and less likely to be superseded by events in the gap between drafting and publication. 🙂
So, without further ado, here are my first four AI predictions for 2025:
1.) Google will consolidate its newfound lead in AI video generation and lead the pack on AI world building
After 9 months of patiently waiting for Sora’s public release, it was Google’s AI video generator, Veo 2, that knocked everyone’s socks off in December. It wasn’t just the aesthetic quality of its output, its prompt adherence, or the promise of longer duration videos, but also its better (but still imperfect) adherence to real-world physics.
Whilst other models will undoubtedly improve and, in time, match and then surpass Veo 2’s performance, I’m betting on Google to maintain its newfound lead in video generation - and lead the pack on AI world building - for two reasons:
It’s sitting on the richest and most diverse source of video training material in the world, which it is in a unique position to exploit on an ongoing basis. The one major AI player missing from YouTube’s recently-announced AI training opt-in setting? Google, because it already has your permission thanks to its Terms of Service.
The likely fruits of the new DeepMind World Modelling Team, led by Tim Brooks (former co-load of OpenAI’s Sora) in further improving visual AI models’ adherence to real-world physics.
The case against: Google is juggling a lot of AI balls at the moment and a company more single-mindedly focussed on image and video generation may alight on the next big leap-forward in visual verisimilitude.
2.) Meta will bring AI image and video generation/manipulation to the masses
Whilst I believe Google will maintain its lead in quality of AI-generated video output, I don’t believe it’s the company that will bring AI image and video generation/manipulation to the masses in 2025.
Standalone AI image and video generation/manipulation tools aren’t currently mainstream, with the leading products (e.g. Midjourney, Adobe Firefly, Runway) attracting just 10s of millions of users (small beer in internet terms).
OpenAI doesn’t disclose how many of ChatGPT’s 300 million weekly active users generate images using DALL-E but even if it were all of them, 300 million is dwarfed by Meta’s product footprint, which welcomes 3.3 billion *daily* active users.
With Mark Zuckerberg seemingly re-emboldened to do whatever it takes to drive engagement and Instagram head Adam Mosseri teasing a Movie Gen Instagram integration, it seems inevitable that AI image and video generation and manipulation will be moved increasingly front and centre within Meta’s core product portfolio (Facebook, Instagram, WhatsApp & Threads).
TikTok will also undoubtedly play a role in taking AI image and video generation more mainstream, although it’s worth remembering that the majority of its ~1.7 billion monthly users are consumers not creators.
Apple (with its ~2.2 billion active devices) will also play a role, although its impact will be limited by Apple Intelligence only being available on newer devices. Apple is playing it safer than Meta by bifurcating its AI image functionality into enhancing existing images (where it says it will be “prioritising accuracy over fantasy”) and generating ‘fun’ non-photographic images via Image Playground.
The case against: Meta could bodge the integration of AI image and video generation/manipulation capabilities, misjudging what users want and/or failing to clear the quality bar (Meta has form here). This would leave the door open for another tech giant to lead the mainstream adoption of AI image and video generation (or keeping it in the early adopter camp for another year).
3.) OpenAI will rebrand ChatGPT as Chat
It’s got incredible brand awareness (especially for a 2 year old) but ChatGPT is a terrible product name, that people frequently get wrong and OpenAI’s applications to trademark ‘ChatGPT’ and ‘GPT’ have both been unsuccessful.
Whilst rebranding to a new name that can be trademarked might seem like the obvious way to go, I think they’re more likely to segue to ‘Chat’, retaining some of the brand equity whilst ditching the ungainly GPT.
There are a few clues hiding in plain sight. In September, OpenAI announced o1 - its first ChatGPT model without ‘GPT’ in its name. In November, OpenAI acquired chat.com, In December, those demoing advanced voice mode in its 12 days of Open AI videos addressed ChatGPT as ‘Chat’ (contrasting with its Spring Update, where ‘ChatGPT’ was used).
The case against: Chat is just too generic to be ownable and, despite being cumbersome, ChatGPT has become category defining, joining the likes of Hoover, Sellotape, Velcro, Post-It, Photoshop and Google.
4.) Amazon’s sleeping AI army will wake up and become the in-home default for many
Alexa celebrated its 10th birthday last November without much in the way fanfare (a 14-second TikTok video does not a celebration make). The voice assistant promised so much at launch but consumers quickly discovered there were only a small number of simple tasks it was worth asking it to attempt: setting alarms/timers/reminders, playing music and radio, reading news headlines and weather forecasts and - for early adopters - controlling smart home appliances.
The advent of LLMs fundamentally changes the nature of what a voice assistant can do. It no longer need rely on engineers having added a rule to cater for that specific class of request. LLMs will have a crack at helping with anything they’ve not been explicitly told not to.
In May 2023, Amazon stated that over half a billion Alexa-enabled devices had been sold. Whilst a proportion of those devices will now be in draws / landfill, most will still be plugged in somewhere, waiting to be dragged into the LLM era.
2024 was littered with the corpses of failed AI hardware launches (Humane’s AI Pin, Rabbit r1, Friend), in part because they were all predicated on the assumption that consumers can be persuaded to carry another device with them wherever they go.
I think the smart money’s on AI upgrades to devices which already have widespread adoption (smart watches, earbuds) and, in time, smart glasses. In the home, Amazon Echos and other Alexa-enabled devices are a sleeping AI army, just waiting to be dragged into the LLM era.
With Claude reportedly bringing the LLM smarts, Amazon is well placed to get the jump on Apple and Google in becoming the AI assistant you turn to around the home when you can’t be bothered to extract your phone from your pocket.
The case against: Amazon is reportedly considering charging $5-$10 per month for its upgraded Alexa, which would seriously limit uptake. Making it another perk for Amazon Prime subscribers feels like the obvious happy medium.
That’s all for this week. Four more predictions to follow next week.
Dan’s Media & AI Sandwich is free of advertising and free to access. If you value my writing and want to help me dedicate more time to it, please consider becoming a paid subscriber (huge thanks to those who already have). Alternatively, you can spread the word.
I work with a wide range of organisations to help them make sense of AI. Whether it's delivering keynotes, running training sessions, drafting policy or designing experiments, I’d love to help. Drop me a line at mail@dantaylorwatt.com.