Julian Mican

Senior Backend Developer | Passionate about AI

[Julian Mican @Julianjjo]

Interested in helping evaluate AI? I’m collecting human judgments to assess and improve SEA_DCS, an experimental model for structured decision-making.
The review includes 48 synthetic cases—not real situations and no customer data—with five decisions per case. Create an account with a username and a unique password; your progress is saved so you can continue later.
Take part here:
https://t.co/YQ087wTEWU

Please don’t enter your real name or personal information. Thanks for sharing your perspective!

[Julian Mican @Julianjjo]

@DotCSV Si te gusta la idea esto creado mi propio Jev pero no tengo datos la idea es sacarlo opensource.

[Julian Mican @Julianjjo]

¿Te interesa ayudar a evaluar inteligencia artificial? Estoy reuniendo opiniones humanas para evaluar y mejorar SEA_DCS, un modelo experimental de juicio estructurado.
La revisión incluye 48 casos sintéticos —no son situaciones reales ni contienen datos de clientes— y cinco decisiones por caso. Crea una cuenta con un usuario y una contraseña única; tu avance se guarda para que puedas continuar después.
Participa aquí:
https://t.co/rnO3ZzSiZn
Por favor, no ingreses tu nombre real ni información personal. ¡Gracias por aportar tu criterio!

[Julian Mican @Julianjjo]

An honest look at orthogonal SEA vs. backpropagation on Breakout. 🎮 Same setup, 10M steps each. SEA learns without standard backprop, but BP achieved stronger sustained performance and won the fresh-episode evaluation. One training seed per method—more replication needed.

[Julian Mican @Julianjjo]

@marcusyul Training an Atari agent with SEA—without standard backpropagation. 🎮 New advance: orthogonal probes for local calibration. Developed iteratively with GPT-6 Astra, grounded in math. Best of 10 games with the current best checkpoint: 284 points in Breakout.

[Julian Mican @Julianjjo]

@goofyninjaaa Training an Atari agent with SEA—without standard backpropagation. 🎮 New advance: orthogonal probes for local calibration. Developed iteratively with GPT-6 Astra, grounded in math. Best of 10 games with the current best checkpoint: 284 points in Breakout.

[Julian Mican @Julianjjo]

@garlotic Training an Atari agent with SEA—without standard backpropagation. 🎮 New advance: orthogonal probes for local calibration. Developed iteratively with GPT-6 Astra, grounded in math. Best of 10 games with the current best checkpoint: 284 points in Breakout.

[Julian Mican @Julianjjo]

@milosz_szewsky Training an Atari agent with SEA—without standard backpropagation. 🎮 New advance: orthogonal probes for local calibration. Developed iteratively with GPT-6 Astra, grounded in math. Best of 10 games with the current best checkpoint: 284 points in Breakout.

[Julian Mican @Julianjjo]

He documentado las matemáticas y los experimentos detrás de SEA, entrenado y mejorado con GPT-6 Astra bajo mi dirección, en “Empirical Adjoint Learning: Calibration and Cost”. 📄 El último borrador incluye calibración local ortogonal y una replicación con 20 semillas. Comparto la evidencia, las limitaciones y las preguntas abiertas. ¡Sus comentarios son bienvenidos!

https://t.co/vQyMZHZbJz

[Julian Mican @Julianjjo]

I’ve written up the math and experiments behind SEA in “Empirical Adjoint Learning: Calibration and Cost.” 📄 The latest draft includes orthogonal local calibration and a 20-seed replication. Sharing the evidence, limitations, and open questions. Feedback welcome!

https://t.co/vQyMZHZbJz

[Julian Mican @Julianjjo]

I’ve written up the math and experiments behind SEA in “Empirical Adjoint Learning: Calibration and Cost.” 📄 The latest draft includes orthogonal local calibration and a 20-seed replication. Sharing the evidence, limitations, and open questions. Feedback welcome!

https://t.co/D3vTRXIYfX

[Julian Mican @Julianjjo]

SEA with orthogonal probes is more erratic during training, but at several checkpoints it scores higher than backpropagation at the same step count. 🎮 Still training. Different inputs and settings (RAM vs. pixels), so this isn’t a controlled comparison—but it’s promising.

[Julian Mican @Julianjjo]

I’ve made an incredible breakthrough with SEA: orthogonal probes for local calibration—without standard backpropagation. 🎮 The new Breakout agent reached a 90.4 evaluation average after just 1.4M steps. Different setup, promising progress. Built iteratively with GPT-6 Astra.

And it’s still training—only 1.8M steps into a 10M-step run. Let’s see how far it can go. 🚀

[Julian Mican @Julianjjo]

I’ve made an incredible breakthrough with SEA: orthogonal probes for local calibration—without standard backpropagation. 🎮 The new Breakout agent reached a 90.4 evaluation average after just 1.4M steps. Different setup, promising progress. Built iteratively with GPT-6 Astra. And it’s still training—only 1.8M steps into a 10M-step run. Let’s see how far it can go. 🚀

[Julian Mican @Julianjjo]

I’ve made an incredible breakthrough with SEA: orthogonal probes for local calibration—without standard backpropagation. 🎮 The new Breakout agent reached a 90.4 evaluation average after just 1.4M steps. Different setup, promising progress. Built iteratively with GPT-6 Astra.
And it’s still training—only 1.8M steps into a 10M-step run. Let’s see how far it can go. 🚀

[Julian Mican @Julianjjo]

@Stefan_3D_AI Training an Atari agent with reinforcement learning using SEA—without standard backpropagation. 🎮
Developed iteratively with GPT-6 Astra, grounded in math.
Here’s its progress so far.

[Julian Mican @Julianjjo]

@Tebasaki_lab Training an Atari agent with reinforcement learning using SEA—without standard backpropagation. 🎮

Here’s its progress so far.

[Julian Mican @Julianjjo]

Training an Atari agent with reinforcement learning using SEA—without standard backpropagation. 🎮
Developed iteratively with GPT-6 Astra, grounded in math.
Here’s its progress so far.