iiinova — where your digital self keeps evolving

CoffeeVectors @coffeevectors.bsky.social · 01/06/2025

Another dialog test with #Veo3! This time in a #darkfantasy setting with one of the characters as an #undead #witch Converted to #verticals with Luma Labs Reframe and upscale with Topaz Labs #aivideo #aicinema #aifilmmaking #aifilm #fantasy #dnd #lich

060

CoffeeVectors @coffeevectors.bsky.social · 28/05/2025

Using #Veo3 to try and recreate a the footstep flowers effect from Princess Mononoke (without the dying flowers). Then used @lumalabsai Reframe to turn the 16:9 into a #vertical. Upscaled in @topazlabs #aivideo #aifilm #verticals

130

CoffeeVectors @coffeevectors.bsky.social · 26/05/2025

My first test with Veo 3! Looking into over the shoulder dialog shots and emotional expressiveness. These are two text to vid prompts I cut together. It took a couple of tries to get two clips that paired up well. Having sound come with the video is definitely a game changer! #veo3 #aivideo #aifilm

010

CoffeeVectors @coffeevectors.bsky.social · 18/04/2025

Interesting how we went from people doing the robot to people acting like NPCs and game avatars.

010

CoffeeVectors @coffeevectors.bsky.social · 10/04/2025

Really impressed by chatGPT's new extended memory! It's mentioning context and ideas now that I'd normally have to manually feed it. Almost like I was continuing a chat rather than starting a new one. #chatGPT

000

CoffeeVectors @coffeevectors.bsky.social · 08/04/2025

Making the personalization profile for Midjourney 7 cracks me up. Do you prefer this picture of a pregnant woman or a closeup of a screaming bat? #midjourney

010

CoffeeVectors @coffeevectors.bsky.social · 06/04/2025

It’s interesting how a no/low code approach leveraged UI to abstract away having to know syntax. But now with LLMs, even though the code is imperfect and it helps to know how to debug and design architecture, I'm finding I want no/low UI with apps now. Is there an API? #aicoding

000

CoffeeVectors @coffeevectors.bsky.social · 03/04/2025

Just realizing you can access MCP servers in Windsurf! So it’s pretty simple to just switch between LLMs if you’re chatting with an SQL database and you want to see how the different models see things. I was just using Claude desktop but wanted to see if Gemini 2.5 was better. #mcp

000

CoffeeVectors @coffeevectors.bsky.social · 27/03/2025

Here’s the original image I uploaded for chatGPT to use as a calligraphy reference.

000

CoffeeVectors @coffeevectors.bsky.social · 27/03/2025

Super impressed by chatGPT’s new image model! Here I’m testing how far I can push the handwriting with a fake drink recipe. For the calligraphy I uploaded a German manuscript from 1765. #aiart #aiartcommunity

120

CoffeeVectors @coffeevectors.bsky.social · 16/03/2025

Would love to hear this in voice!

010

CoffeeVectors @coffeevectors.bsky.social · 16/03/2025

Been trying to use the Blender MCP server to make some 3D scenes with Claude. Asked it to make a bowl and in trying to delete some existing geometry it ran a factory reset of the entire app. Luckily it was on a fresh Blender install. Be careful if you try it. #mcp

010

CoffeeVectors @coffeevectors.bsky.social · 12/03/2025

While there’s some assurances on security, I think it’s best to be conservative with what Claude Desktop sees. I’m not analyzing anything overly sensitive. I’ll need to try using a local LLM with MCP down the road but don’t want to overwhelm myself with too much setup. 3/3

000

CoffeeVectors @coffeevectors.bsky.social · 12/03/2025

I’m still running most of the scripts manually in DBeaver (the SQL server for MCP can’t do everything and I want to be careful about what gets executed). Having Claude read the DB can help with writing tedious SQL. Here I’m linking transactions to tags in a join table. 2/3 #mcp #sql

100

CoffeeVectors @coffeevectors.bsky.social · 12/03/2025

Learning how to use SQL and to analyze databases with Anthropic’s MCP. I’m new to this DB stuff so it’s been really eye opening. Here I had Claude Desktop make some animated charts from expenses tagged as transportation.🧵1/3 #sql #mcp #aicoding #dataanalysis

110

CoffeeVectors @coffeevectors.bsky.social · 09/03/2025

While I totally get the value in building apps without reading code, learning to read code/architecture has given me a huge new vocabulary for breaking down problems and understanding systems thinking outside of software. But it’s about as tough as learning a new language. #aicoding

000

CoffeeVectors @coffeevectors.bsky.social · 09/03/2025

Testing #HedraLabs new lip syncing Character 3 model on some #AImusic I made in #suno Ref images made in #magnifique and also upscaled for the closeups variations. Pretty incredible for just img&audio2vid! Camera zooms/editing/color -grade/grain in post. Upscaled to 4k. #aivideo #aifilmmaking

020

CoffeeVectors @coffeevectors.bsky.social · 08/03/2025

I want to see a movie like Flow, but with Godzilla. Like forget about the humans.

010

CoffeeVectors @coffeevectors.bsky.social · 07/03/2025

Haven’t been doing as much with the creative side of AI gen lately because I’ve been so focused on vibe coding and learning to code in general. Its already doing something because when new stuff comes out, using it isn’t nearly as intimidating or difficult to me as it used to be. #aicoding

000

CoffeeVectors @coffeevectors.bsky.social · 27/02/2025

Initial tests with Wan 2.1 14b text2video running locally in ComfyUI on a 3090. 1st vid was 20 steps and took 30 minutes. 2nd vid was 50 steps and took 80 minutes. Both were 1280x720 and using close to all 24GB of VRAM. Still messing with settings and different models. #aivideo #comfyui

120

CoffeeVectors @coffeevectors.bsky.social · 26/02/2025

Quick o3-mini deep research prompt asking for all AI releases in the past 7 days that can actually be used by a person or small business (with a powerful consumer GPU). Report includes release dates. Fast way to get up to speed if you’ve been working on other things. #deepresearch #chatGPT

000

CoffeeVectors @coffeevectors.bsky.social · 26/02/2025

Deep Research available in the regular chatGPT subscription! Nice!

010

CoffeeVectors @coffeevectors.bsky.social · 23/02/2025

By chat with docs I just mean loading it into context so you can ask questions about it. I could be using the term wrong. But it’s useful if you want to expand or summarize anything in the report.

100

CoffeeVectors @coffeevectors.bsky.social · 23/02/2025

Been loving this open source version of Deep Research by dzhng and reading the files in Frogmouth. Here I’m just asking what the best Gaussian Splat repos currently are. You can also manually paste the reports into an LLM and chat with it.

200

Reposted by CoffeeVectors

Ethan Mollick @emollick.bsky.social · 15/02/2025

This new paper shows people could not tell the difference between the written responses of ChatGPT-4o & expert therapists, and that they preferred ChatGPT's responses. Effectiveness is not measured. Given that people use LLMs for therapy now, this is an important (and urgent) topic for study.

712316

Reposted by CoffeeVectors

mr. TIM @timkellogg.me · 15/02/2025

Perplexity announced their own DeepResearch that includes a free tier and a generous $20/mo tier People who have tried both are finding the Perplexity version to be on par, perhaps better than OpenAI’s ($200/mo) Analysis of the Chernobyl drone strike yesterday: www.perplexity.ai/search/yeste...

perplexity.ai

yesterday a drone hit the Chernobyl nuclear power plant. Please investigate if...

The recent drone strike on the Chernobyl Nuclear Power Plant’s protective containment structure has reignited global concerns about nuclear safety amid the...

5258

CoffeeVectors @coffeevectors.bsky.social · 08/02/2025

You can see my last #Veo2 video here: bsky.app/profile/coff...

000

CoffeeVectors @coffeevectors.bsky.social · 08/02/2025

And while you can still get videos where the physics goes crazy, people morph, or heads go 180 degrees, it’s much less than other apps. You typically get coherent stuff that feels like footage, if you’re going for photorealism. 3/3 #Veo2

100

CoffeeVectors @coffeevectors.bsky.social · 08/02/2025

It’s really impressive how well it handles some of these complex setups. Liquids, fabrics, shapes showing under fabric, metal, reflections, hands, water on skin, wind. A lot is going on in these. 2/3 #Veo2

100

CoffeeVectors @coffeevectors.bsky.social · 08/02/2025

Testing how #Veo2 handles liquids & fabrics. In post I added some zooming, speed ramping, some occasional focus blur and a little color grade, but these shots are basically out of the box. The rain, lens flares, splashes and slow motion were prompted. (Music: Suno) 🧵 1/3 #aivideo #aifilm

110

CoffeeVectors @coffeevectors.bsky.social · 05/02/2025

Was chatting with R1 about simulation theory and it brought up the idea of the creators of the sim being in a sim themselves, that the base reality might be the infinite ladder of nested sims. Like a vertical multiverse. I wonder if you could make this into a game mechanic?

000

Reposted by CoffeeVectors

Ethan Mollick @emollick.bsky.social · 03/02/2025

wrote about Deep Research, which is very, very good at doing nuanced and complex research. It is also the first narrow agent that can do sophisticated and likely quite economically valuable work, which tells us something important about the future. open.substack.com/pub/oneusefu...

open.substack.com

The End of Search, The Beginning of Research

The first narrow agents are here

19716

CoffeeVectors @coffeevectors.bsky.social · 02/02/2025

That said it’s not 100% ready to go. Sometimes I’ll get errors and the Blender API docs won’t be enough even for o3-mini or R1. But I’m hoping in the future a prompt > scripting > API approach gives us more flexibility with using programs without needing perfect UX recall. 5/5

010

CoffeeVectors @coffeevectors.bsky.social · 02/02/2025

This is useful if you’re not too familiar with Blender or setting up fluid sims, or if you’re juggling a lot of different apps and interface switching is burning you out. It can be tough to navigate all the UXs from memory and using AI prompts + Python/scripting can make it easier. 4/5

110

CoffeeVectors @coffeevectors.bsky.social · 02/02/2025

I made the script that built the scene using o3-mini. The scene is made using the fluid sim tools and physics libraries in Blender, but by using Python and prompts we can set up the scene without having to go through whole UX. 3/5

100

CoffeeVectors @coffeevectors.bsky.social · 02/02/2025

Here’s the source video. Note that it’s not rendered and there’s no lighting. I used OBS to screen record the UX in solid view with all the overlays removed. Then I just zoomed into the camera view in video editing software. 2/5

100

CoffeeVectors @coffeevectors.bsky.social · 02/02/2025

Took a simple fluid sim scene that I set up in Blender using a python script made in o3-mini. And then I ran it through several styles with vid2vid in Runway ML 🧵1/5 #aivideo

111

CoffeeVectors @coffeevectors.bsky.social · 02/02/2025

Should mention that I also asked for a water material to be applied and lighting for EEVEE, but it was having a tough time making materials programmatically even tho it had access to the docs.

000

CoffeeVectors @coffeevectors.bsky.social · 02/02/2025

Took about 3-4 tries to get this to work so this isn’t the exact prompt, but it’s the basic outline:

Propose a water simulation in a tank. It should be set up with a python script to run in Blender 4.3.

Add a camera parented to an empty following a circular path over the course of 10 seconds.

Please list out the steps before writing the script, and then follow your blueprint to make sure you include all the relevant details and constraints.

100

CoffeeVectors @coffeevectors.bsky.social · 02/02/2025

Asked o3-mini to make a quick water sim with an orbiting camera in Blender 4.3 using a python script. This took a few tries and some tweaking so it’s not a one shot. I’m optimistic this will improve down the road. Prompt below. #o3mini #aicoding #blender #fluidsim

130

CoffeeVectors @coffeevectors.bsky.social · 01/02/2025

I’m also building out some custom features that better reflect how I use AI to learn stuff. It’s not so much a product for the market, but an exploration of accelerating my own process. If I can’t build my own tools, I feel like I don’t have full access to my own potential.

020

CoffeeVectors @coffeevectors.bsky.social · 01/02/2025

It also gave me an appreciation for details I wouldn’t normally think about as a user (and non-dev), like styling markdown language, what context should I send to the model, and how I want to display reasoning tokens.

120