Pinned Tweet
Starting the day at the gym for another productive day.
Yesterday client was happy with AI Coding tutoring on Claude Code w/ the BMAD Method and Cursor. Liking it so much heās booking 3rd session.
If you have more money than time and want a tutoring session, book a session or connect with me to see if itās a good fit for you.
calendly.com/jorgecolonconsuā¦
5
31
1,929
YC-backed founder @Road_Kill11 has been working hard on this one. Check it out!
AI has outgrown the way we typically use it.
Your agents are ready to take on far more. They just need something built for the scale of your ambition.
Introducing Dexto, a multi-agent harness for AI-native teams.
More in the thread below.
1
48
Jorge Colon retweeted
After Google acquired my AR startup, I started thinking about the next interface, this time for AI. Text and voice work fine, just like DOS did before the Mac. But nothingās more intuitive than face to face.
After two years of research, today weāre launching Atmee v2: a real-time avatar model that makes live video agents feel truly life-like.
⢠Any character from 1 image + 10s of audio
⢠Contextual expressions, gestures, even interacting with the environment
⢠Live video calls, voice only, or video generation
At its core, Atmee v2 is a video foundation model, it can render anything, but itās built to excel at characters, which is what video models like Seedance 2.5 and Wan 3.0 are mostly used for. In comparison:
Atmee v2
⢠Up to 50x faster
⢠10-100x cheaper
Build tutors, sales coaches, support agents, companions, game characters, even whole living worlds. The interface after text and voice starts today.
š±š±š±
ChatGPT Images 2.6 sets the scene. Tesseract brings it to life.
Generate your visuals with Images 2.6, then make them move with Tesseract, from animating an SVG to turning a storyboard into a finished video.
Direct the animation, effects, and sound through your agent, all in one editable project.
57
Highest reasoning for each task is often overkill. Here I show the pros and cons of different Gemini models to analyze videos.
1
1
395
Jorge Colon retweeted
If your company needs software dev help, Iād love to talk to you! I have some amazing experienced US-based devs on the bench who are very good at using agents and have a high quality bar.
We are best at mobile and web and desktop apps, but can do much more than that.
Average experience level is 15 years. Youāll love working with them. Great people, excellent devs.
Appreciate retweets and any referrals!
These past few days Iāve been vibe coding an AI fact checker to understand what would be involved. I thought Iād use Jev to get better results. Got Opus 5.5 to build me something, do the evals, and improve on those. On the surface it looked good, butā¦
2
61
It does best when you curate the state AND the questions made against that state.
Which if those questions needs to be dynamic based on the state then you practically lost the speed advantage. You have to call an LLM with the state and then generate the tailored questions.
1
7
This is huge for those of us that want to create prototypes to understand Anthropicās platform even better!
It includes the Claude API, Claude Managed Agents + the Claude Agent SDK like before!
Replying to @ClaudeDevs
Weāre rolling out monthly Claude Platform API credits for Max and Team plans:
Max 5x: $100
Max 20x: $200
Team: up to $500, pooled
Works on any model, including Haiku 5.5, in your code or third-party harnesses.
How it works and full terms: support.claude.com/en/articlā¦
1
117
GrokBot + Tesseract + a couple of prompts
Not award winning just yet. Iām sure with more iterations there could be a huge improvement.
The interesting part of this workflow is creating videos while on-the-run or creating a fleet of agents that creates videos.
@higgsfield_ai
75
I had GPT 6.1 Sol create a Langgraph AI video agent and to my surprise it tested a very specific video.
Canāt believe I got rickrolled by Sol 6.1 š.
2
2
128
This is a new step in AI Video Gen.
@ltx_io I didnāt seem to see a way to specify in a prompt what specifically to isolate? If not, I think that could be an improvement to what it offers.
The last drop of VFX Week: Alpha Gen (Beta).
Give it an ordinary RGB clip and get back an alpha matte that matches your plate frame for frame. No green screen, no masks, no prompt.
Built for the edges that are hard to key or roto: hair, fur, smoke, fire, glass, water. š§µ
73
Dude. Youāre a genius. Created an MCP used as an intake form and embedding yourself into someoneās workflow.
Very very creative problem solving. Bravo!!
less than 24 hours after launch and already at capacity for the week.
i will update the credit system to allow "fast" and regular speed so i can take on more interfaces.
53
You can get great results depending on the use case. The thing is that you need a golden data set where you know for sure a classification is good or not.
If you blindly accept high confidence values as correctness youāre gonna be let down. But if you have a golden data set you can optimize the instructions and criteria to get you much better results.
This is what I tried to understand, too.
Because you can get a JSON response from all the big classification models, it's an easy task.
Jev is 100x faster / cheaper.
But nobody is talking about the quality.
I don't write a single line of AI in my tweets, yet all those Jev classifiers tagged them as AI.
We need someone to compare smart models' (Fable / Astra) classification with Jev.
I am sure Jev is great for dumb classification, but maybe even a simple machine learning model can be good.
47
Iāve been hacking away and comparing Jev to classifiers and I by no means am a machine learning expert yet. Through my experiments though Iāve realized that if you donāt have a curated eval data set to check how Jev behaves against, itās hard to truly know if itās working in the majority of the cases. High confidence scores ā correctness. Itās also hard to optimize because thereās no goal post to compare against.
Thereās also a lot of moving levers such as the specific state you give it as input, the instructions you give it on what to look for and the different questions you ask it. A small change in the instructions and criteria and the probabilities and confidence scores change. Typical in prompt engineering.
So just record test cases, curate them and then run Jev against those. Even TypeSafeās docs recommend doing that. Itās how you find out what thresholds you should be using based on your data.
In all I am excited about this new primitive in AI engineering.
2
74














