Tech Landscape #417
A great leap forward for AI video, more realistic avatars, and simplifying creator tools,
Hello!
An extremely impressive new video model launched this week, leading to a chorus of “Hollywood is cooked!” across social media. That is, of course, nonsense, and you shouldn’t pay attention to the people who say it because they’re invariably the people who largely only watch superhero movies. There’s a vast difference between generating a bunch of 15-second action sequence and making a full, coherent, engaging and emotional movie. I just watched One Battle After Another and the idea that AI video is capable of reproducing that is ridiculous.
I say all of this as a proponent; I think clever and creative people will use AI video in clever and creative ways — but I also think that dull and unimaginative people will use it in dull and unimaginative ways. Will we see it used in movies to replace the CG in car chases and colossal superhero fights? Almost certainly. And Zac Snyder might well be worried. But the notion that this is the end of Hollywood is laughable. Please ignore those people.
Right, let’s get on with it. Hope you’re well!
p.s. The exclusive content for paid subscribers is almost done; hopefully will be sent mid-week.
Synthetic Audio-Visual
ByteDance launched Seedance 2.0
It’s a unified audio-video generation and editing model that supports text, image, and audio inputs, with exceptional motion, stability, and control, and cinematic output.
seed.bytedance.com/en/seedance2_0
You have very probably seen this blowing up across the internet this week; it is incredibly good quality, and had very few guardrails at launch around what it will generate, leading to it being denounced by the U.S. Motion Picture Association and SAG-AFTRA actors union, and a cease-and-desist letter from Disney. ByteDance has reportedly added some restrictions to human likeness generation since it was released.
I haven’t had the chance to use it yet; these ⬆️ are all examples I found being shared on social (without creator attribution, sorry). It seems to have been launched to select creators in China first, but is now available to Dreamina creative partners here, and its imminent release is being teased by creative platforms. When it officially launches here, I’ll write more about it. What I will say, though, is that you’re only seeing the very best cherry-picked examples, and it remains to be seen how it handles in reality.
ByteDance also launched Seedream 5.0 Lite, a multimodal image generation model that integrates real-time web search and multi-step reasoning to improve the accuracy and relevance of generated visuals. seed.bytedance.com
It’s a very good model with decent photorealism (as you can see in my example ⬇️) and capabilities that come close to, although don’t quite meet, those of Google’s Nano Banana Pro — although this is, to be fair, only the Lite version. It’s exclusively available in CapCut AI Studio and Dreamina, for now.
Alibaba introduced Qwen-Image-2.0, an image generation and editing model with complex prompt comprehension, a typography engine for generating long text, and improved photorealism, in up to 2K resolution. qwen.ai
This is another model that sits in the same territory as Nano Banana Pro, although it’s not at the same level; you can see in my example ⬇️ that the characters and their positions and poses are a little off, although it handles the text well. For now it’s only available through Qwen Chat, where it also powers the new Slides feature.
Creator Tools
Krea’s Prompt-to-Workflow lets users create node-based workflows from a text prompt. instagram.com/krea_ai
Krea launched an iPad app with custom brushes and near-real-time AI image generation. instagram.com/krea_ai
The iPad app looks like a quick rebadge of an app called Wand which Krea just acquired, and uses the real-time feature introduced a few weeks ago [TL 414]. Both of these two stories show a plan from Krea to give people different ways to create with little effort.
Suno’s Studio 1.2 gives musicians more AI tools, including Remove FX, Warp Markers with Quantize, and support for more time signatures. suno.com
Avatars & Voice
Hedra launched Avatar, enabling lip-sync with natural motion and full expressions at up to five minutes length. threads.com/@hedra.labs
You can see in my demo ⬇️ that the lip-syncing and character motion is very good (ignore her silently muttering companion); the voice isn’t great but I didn’t have enough credits to run it again. This is the second new talking character model from Hedra, following the launch of Omnia last week [TL 416]; Avatar is for longer videos with less movement, Omnia for shorter and more expressive ones.
Tavus announced Raven-1, a “multimodal perception system” that can better understand a user’s emotional state, context, and intent to more appropriately respond in a conversation. tavus.io
ElevenLabs introduced Expressive Mode for its ElevenAgents, which gives “unprecedented control” over a voice agent’s tone so it can “de-escalate, reassure, and guide conversations to a clear resolution”. elevenlabs.io
Both of these are technically impressive, but ethically tricky; there’s a moment in one of the ElevenLabs demo videos where the AI system detects frustration and replies with “Ugh, yeah, I hear you, and I’m so sorry about that”, and I think that if I knew I were talking to a machine that level of hollow artificial empathy and sincerity would make me furious.
Startup Xmax announced X1, a real-time AI video model that can create interactive characters in XR, transform people into AI avatars, and more. x.com/XmaxAIOfficial
It’s fun enough, still rough and maybe a little bit pointless in its current state. But an indicator of how we might remix reality in the future.
Assistants & Search
Google laid out its vision for AI-powered commerce in 2026, including ads in AI Mode in Search, agentic checkout in AI Mode and Gemini, and “instantly matching brands with the creator communities that will love their products” in YouTube. blog.google
OpenAI began testing ads in ChatGPT for users in the U.S. on the Free and Go tiers. openai.com
Claude moved some of its most popular features to the Free plan: file creation, connectors, and skills are all now available without a subscription. threads.com/@claudeai
Research
ChatGPT upgraded Deep Research, adding connection to apps, site-specific search, real-time progress tracking, and fullscreen reports, now powered by GPT 5.2. x.com/OpenAI
Google released a major upgrade to Gemini 3 Deep Think that aims to solve complex science and engineering challenges. It’s available to Google AI Ultra subscribers and via the Gemini API. blog.google
Chinese Foundation Models
I started this edition by talking about China taking the lead in video models, and they’re also leading in LLMs too — and what’s more, they’re often giving them away for free (open source).
Z.ai ’s GLM-5 is a new-generation flagship foundation model targeted at complex systems engineering and long-horizon agentic tasks. z.ai
It has a record low hallucination rate, beating even the best commercial models from the U.S.MiniMax’s M2.5, is a multimodal reasoning model optimised for real-world productivity tasks such as coding and office work, with significantly reduced costs. minimax.io
ByteDance’s Seed2.0 is a family of general-purpose AI agent models that feature enhanced multimodal reasoning and improved stability for complex, long-chain tasks. seed.bytedance.com [in Chinese]
I haven’t tried these so can’t evaluate them, and TBH I doubt most people reading this will use them much; this is more of a note to show how market power is shifting towards China — and American companies are crying foul. OpenAI has accused DeepSeek of “free-riding”, and Google says unnamed operators are running distillation attacks to copy Gemini’s reasoning abilities. Having the most popular AI models is a form of soft power.
Social
Facebook added Meta AI features to give looping motion to profile pics, restyle Stories and Memories, and generate animated backgrounds for posts. about.fb.com
Threads launched Dear Algo, which lets users customise (temporarily) their feeds by writing a post with a specific request. about.fb.com
It’s been quite funny to see people request “Dear Algo please don’t show me AI”.TikTok US introduced the Local Feed, a new home screen tab designed to show users nearby content and businesses by using their real-time location. usdsjv.tiktok.com
This was launched in other countries (as the Nearby Feed) in December last year [TL 409] but it appears that the US version of the app may have a different feature release cadence from now on.Discord rolled out its Teen-by-Default settings globally. If the age prediction system isn’t confident that a user is an adult they’ll be asked to prove their age when trying to access age-restricted servers. discord.com
MrBeast bought a bank. Well, his company Beast Industries acquired Step, a youth-focused financial services app, but I prefer the alliteration. step.com
Everything Else
The v85.0 update for Quest adds a feature that can turn any surface into a keyboard, voice control for windows, and more UI and privacy refinements. meta.com
Meta may have been cutting back on Reality Labs spending, but it’s good to see that the Quest is still a priority; these are all useful quality-of-life updates ahead of the predicted launch of Quest 4 later this year.