Tech Landscape #424
A faster and cheaper Veo, the rise of passive social media, and on-device intelligence.
Hello!
As I write this I’m just back from a two-day walking trip around Chichester Harbour, a designated National Landscape (or “Area of Outstanding Natural Beauty”). I like to take long walks because I use the time to think about nothing; no work, no deep thoughts, just noticing and paying attention to the things around me. The Japanese call this shinrin-yoku, which we translate as forest-bathing, but for me it can be anywhere — such as a coastal path.
When I tell people about this they tend to presume that I don't use any technology — that I go “analogue” or “off-grid”, in the modern parlance. But it's not that at all; I use technology when it enhances my walks. I don’t use social media (which is still professionally valuable but gives me less and less… well, socially), I don’t check the news, anything like that. I use Komoot for route maps, and my camera, and I take notes, but mostly I use it to know more.
I use Merlin to identify birds from their songs, and Google’s ‘Circle to Search’ to identify birds and plants I can see, and the Web to look up the history of buildings and places around me. In short, I use it to deepen my connection to my immediate environment. I don't spend a lot of time with it, but the time I spend with it is time well spent. As in all things, the value lies in balance.
Anyway, let’s get on with it. Hope you’re well!
Synthetic Audio-Visual
Google released Veo 3.1 Lite, a variant of its AI video model that’s 50% cheaper than Veo 3.1 Fast but just as fast. blog.google
It seems to have little-to-no drop in quality, from my limited tests (e.g. ⬇️). It’s available in Flow, AI Studio, and through the Gemini API.
Grok Imagine added Quality Mode for image generation, bringing improved realism, stronger text rendering, and higher world knowledge for more creative control. x.com/xai
This is only available to SuperGrok users which means I can’t try it, because I’m not paying £28.49 a month for that.
Seedance is rolling out to creative platforms including Freepik, Higgsfield, and more.
Interestingly it seems that platforms which wish to use the API have to ask their users for a contractual agreement that passes on liability for copyright infringement; this is the first model I’ve used in Freepik that has asked me to accept that. Letz AI is the first platform to take a moral stance:
The idea of gatekeeping models through complex and expensive contractual agreements will age badly. That's why, under the current conditions, we will not proceed with our Seedance 2 integration.
Alibaba released Wan 2.7 models: Wan 2.7-Video, a multimodal video generation, editing, and motion transfer model with the usual enhancements to visual fidelity, motion stability, and prompt adherence; and Wan2.7-Image, a unified image generation and editing model with realistic varied faces, complex text rendering, and colour palette support.
These models aren’t the cheapest, the highest-quality, or the fastest (the video model is downright slow)… they’re very mid-range. You can see in my tests ⬇️ of Wan 2.7-Video that the quality of the text-to-video is really variable (WTF clip 1?), which is strange as Wan 2.7-Image is pretty good; in other clips it fails to lip sync, or invents a voiceover I didn’t want. Wan models were much more interesting when they were open source and enabled the developer community to build useful things; now that they’re commercial they’re just… there.
Also of note: Netflix released its first AI video editing model. VOID (Video Object and Interaction Deletion) removes objects from a video with “physically-plausible inpainting” — that is, it won’t just remove the object but will regenerate the rest of the scene so that the removal impacts it; in a scene with two cars colliding, removing one car will also prevent the collision. It’s available to download under an open source license so that anyone can use it.
Creative Tools
Flora introduced FAUNA, an AI creative agent which automates complex workflows such as rapid ideation and content transcreation. flora.ai
This agentic creative interface is becoming popular — dare I say, even, the standard? Just in the past month we’ve seen Hedra Agent, Luma Agents, CapCut Video Studio, and now this.
Higgsfield launched Cinema Studio 3.0, an update to its video generation platform featuring “unprecedented output quality, accurate physics, seamless complex motion, and native audio”. higgsfield.ai
From what I can tell, this means that Cinema Studio now uses Seedance 2.0 as its base model.Google Flow added Voice Ingredients, enabling consistent voices across generations. instagram.com/flowbygoogle
Only 30 voices, and only for Ultra subscribers, but a step forwards nonetheless.
Pika Selves can now join Meet calls using the new real-time video chat feature. instagram.com/pika_labs
It’s not straightforward to use; it’s a Skill for AI coding agents (e.g. Claude Code, OpenClaw) which means you have to have all that in place first, then it costs $0.275 per minute, so the barrier is high and… I’m not sure what it’s for yet. Maybe there’s a long-term plan, but right now Pika just seems to be furiously pivoting to whatever the latest trend is.
Social & Messaging
All (eligible) Snapchat creators can now set up Creator Subscriptions to offer exclusive content and priority replies to their fans for a monthly fee. newsroom.snap.com
Telegram’s latest update introduced an AI Editor that can translate, rewrite, or fix grammar in multiple styles, plus new features for Polls, native support for Live and Motion Photos, and more. telegram.org
Rec Room will shut down on June 1st due to a lack of sustainable profitability and shifts in the gaming market. blog.recroom.com
Another metaverse-era product gone.
UK adults are less active on social media. Its use is “becoming more passive and circumspect” according to the latest Ofcom research; although 89% of adults use social media, only 49% now actively post, share or comment — down from 61% in 2024. That’s partly due to the increased focus on video, which makes the barrier to posting higher, and partly to concern about being held accountable for their posts.
The push to video has another effect: with fewer links in posts, fewer UK adults are discovering new websites; from 70% in 2024 to 56% today. The social Web aggregator / discovery platform Surf wants to change that; it launched its web version (in beta) this week with a new feature called Social Websites which lets creators and publishers build community destinations that aggregate content and conversations from across multiple open social networks, including RSS.
Assistants & Voice
Google released Gemma 4, a family of open multimodal AI models designed for advanced reasoning and agentic workflows, in four sizes optimised for various hardware, with an open source (Apache 2.0) license. blog.google
This is notable because a) the smaller models can be run locally on smartphones (they’ll be the default on-device model for Android), and b) the open source license means developers can use them to build their own products on, in a push back against the dominance of Chinese open source models.
Microsoft launched two new in-house AI models: MAI-Transcribe-1 for speech-to-text, and MAI-Voice-1 for natural speech generation. microsoft.ai
These join the recently-released MAI-Image-2 [TL 422] in helping Microsoft wean itself off of its dependency on OpenAI. The headliner is MAI-Transcribe-1, which claims state-of-the-art performance.Microsoft 365 Copilot added Cowork, a tool designed to plan and execute long-running, multi-step workflows across Microsoft 365 apps, that’s based on Claude Cowork. microsoft.com