Another Veo 3.1 experiment - I showed it physics textbooks & prompted "Continue the scene in a funny and surprising way." (I also asked for a Scottish accent audio narrator, which I think explains the Scottie dogs 🐕 🤷) Physics comprehension ❌⁉️ Trolling 💯✅

I like to poke at where gen AI models make interesting mistakes (interesting to me at least). ... but today, I was impressed at how Veo 3.1 accurately read text in a photo I took (last year) and quoted it in audio narration. Other examples confirm text reading & comprehension.

With Veo 3.1, I noticed that rowing boats travel the wrong way. People moving in the direction they face is so universal that a strong prior forms that's hard to unlearn in this exceptional case. I animated more photos I took of other cities to verify 🧐 xitter.cf/AnthonyPPhysics/status…

I tried out Veo 3.1 to animate a photo I took of Cambridge earlier this year. Locals can tell it's AI because of several problems with the boat 🚣 Veo has learnt such a strong rule that people move in the direction of facing that it hasn't also learnt this one weird exception 🤔

I saw a claim that AI video models can't generate finger clicking. I tried Veo3 - it sticks middle fingers up😱😅. The sound of clicking is correct, so it's within the training set 🤔. My guess is that it's akin to the bias against left-handed writing - unbalanced dataset effect?

You've had "pizza quattro formaggi". You may even have had an 8-cheese pizza - but have you had a 16-cheese pizza (with bacon and mushrooms)? The world record is 1001 cheeses, but I say we can beat that with a 1024-cheese pizza if we invest more in fundamental R&D 🚀🌖.

Many UK X users say that political posts are censored now that the Online Safety Bill has taken effect. I'm not political, so the censoring I see is bizarre 🥴. Now I will never know why I shouldn't incorporate in Delaware 😭 Reminder of how we got here: windowscentral.com/british-g…

I was challenged to make a cool AI video about my PhD (years ago), because I'm uncool without AI 😭 Script by Gemini, video by Veo3 🪄 Equations in the video are garbage, the script needs work- but back then, this video was inconceivable. AI moved like lightning this ~10 years🏇.

First attempt at an AI-generated "documentary" 😅. "Standard Model Particle Discoveries" Script by Grok, video by Veo3 (audio narrated). *Minimum* editing by me, e.g. where Veo3 refuses to accept names in prompts (and "Cambridge" - 🤷). Limited compute budget, so "imperfect" 🤣

Let's see if Veo3 can generate reversed length scale hierarchies. Ordinarily: Bacterium < raindrop < chest < castle < forest < ocean < planet < star < galaxy. How much nested scale reversion can we get? Examples here. With good prompts, it can do a little, but doesn't enjoy it 😅

Utopian visions from Grok, rendered by Veo3 Today: habitats high in Earth's atmosphere. 1. Scientific Research Outpost 2. Cultural and Educational Retreat 3. Disaster-Resilient Community 4. Eco-Tourism and Wellness Haven I'm not convinced, but it's a hard challenge. Better ideas?

Next up in *positive* future worlds made by Veo3 from Grok descriptions is... VR. 1. Skills training 2. Remote collaboration 3. Technical assist HUD 4. Emergency simulation Whatever happened to the Metaverse? What is an optimistic alternative to Snowcrash, Ready Player One?

After getting Grok & Veo3 to create positive visual depictions of life underwater and on Mars, I turned to life underground. What's a positive vision of humans prospering in underground cities? What would make people positively choose it instead of the status quo?

Suppose we're going to Mars 🚀. Suppose we have to 😱. What's your vision of how we're going to live 🌱? I asked Grok to suggest some habitats, and passed the descriptions to Veo3. 1. Lava tube 2. Inflatable 3. Ice cavern 4. Crater rim Any better ideas? 🧐

Oceans cover ~71% of the Earth's surface, but most people rarely think of them. We cultivate the land, but act as hunter-gatherers at sea. Humanity adapts to new environments, and is looking to new planets. What is your vision for humanity's future technology in the ocean?

There's a short supply of welders - the lack of young people becoming welders has forced the older welders out of retirement to stop the world ending 🧓. I asked Gemini to suggest some robots for general purpose mobile welding 🦾. "How hard can it be?" 😅🚀🏁 (very, of course)

I've been feeding "fake word" prompts to Veo3 xitter.cf/AnthonyPPhysics/status…. - Try "ej republast ish pluncidelst diffechindic" 4x. The two speaking ones are interesting (audio) 1) Error corrects to "we republicans pledge to defend" 2) Says the fake words Varied but not unconstrained 🤔

I'm puzzled when I see people on X imagine that LLM models are self-knowledgeable enough to answer questions about themselves not in public documentation 😕. Humans aren't born with innate knowledge of their neurobiology, and AI models only "know" what is in their training data.

Previously xitter.cf/AnthonyPPhysics/status…, I gave "fake words" to Veo3 and got random videos back. What about the same fake words again? The original three in order, plus the 1st again as a bonus. - No similarity on re-runs? - Implies no consistent meaning inferred? - How much freedom?

A common AI flex is "faster than Moore's law" growth🦾 Often, this refers to - Model size - Compute These mean that effort spent outpaces the underlying technology. This is a looming problem, not a flex. It requires Moore's law to speed up or ML methods to get smarter to sustain.