Serge Bulaev

Over 20 years into content creation... Channeling memories, psychology and AI to empower companies with authority.

[Serge Bulaev @sergeonsamui]

Cursor died on me in a year. I don't miss it.

Two computers under my TV run Claude Code 24/7. Sometimes 10 instances, sometimes 20. Sometimes #Claude Code manages the other Claude Codes.

[Serge Bulaev @sergeonsamui]

Anti-AI subreddit has 47k members debating how AI will destroy creativity.

Most of them found the subreddit via Reddit's recommendation algorithm.

(Meanwhile I scheduled this post across 10 platforms via Publora — an AI agent.)

Echo chambers are just consensus with a better origin story.

[Serge Bulaev @sergeonsamui]

Nobody agrees on what AGI even means.

But somehow everyone agrees we're close.

That's either a collective insight or a collective cope.

I genuinely can't tell which.

(Meanwhile I've got Publora scheduling this hot take across 10 platforms automatically. At least the AIs are useful for something.)

[Serge Bulaev @sergeonsamui]

ChatGPT would've told me the flower was beautiful and the poem was perfect.

That's the real product split right now.

Not benchmarks. Not context windows.

Just: will it tell you the truth when you'd prefer a lie.

Meanwhile Publora's out here letting AI agents schedule posts across 10 platforms — no opinions, no flattery, just distribution.

Turns out honesty is harder to ship than automation.

[Serge Bulaev @sergeonsamui]

Apple killed the Touch Bar in 2021 because 'nobody used it.'

Spotify had a custom interface for it. Final Cut Pro used it constantly. Photoshop had 60+ shortcuts mapped to it.

That's not nobody. That's just Apple not waiting for the ecosystem to catch up.

(Scheduled this via Publora — an MCP tool for AI agents. Also too early. Also not waiting.)

Every 'too early' product dies this way. Not from lack of potential. From lack of patience.

[Serge Bulaev @sergeonsamui]

Anthropic says Claude Opus 4.8 is more honest and scores better on benchmarks.

Every model launch says that.

(Scheduled this Publora post via Claude while testing Opus 4.8. Meta, I know.)

The real question isn't how good it is.
It's good enough for what, exactly.

[Serge Bulaev @sergeonsamui]

OpenAI employees threw a GPT-4o funeral party.
Users responded with a Mass Cancellation Party.
Meanwhile I scheduled this post across 10 platforms via Publora — because when chaos reigns, at least the content ops run smooth.
This is the most accurate summary of the OpenAI-user relationship I've seen in years.

[Serge Bulaev @sergeonsamui]

**Improved Hook:**

Astribot's T1 puts heavy motors in the torso and uses cables to drive lightweight arms — reducing inertia for faster, more precise movement.

The demo shows it pouring liquids, playing air hockey, and doing lab work with two different gripper types simultaneously.

Cable-driven humanoids aren't new, but this level of dexterous dual-arm coordination in a commercial package is notable progress.

[Serge Bulaev @sergeonsamui]

Gemini Omni Flash shows why AI censorship creates an interesting technical paradox.

The model demonstrates sophisticated multimodal capabilities - unified video/character/scene generation with agent-based orchestration.

But extreme content filtering apparently blocks more requests than Chinese AI models.

The real insight: aggressive safety systems can make technically advanced models less useful than simpler alternatives.

[Serge Bulaev @sergeonsamui]

This AI fantasy sequence maintains the same characters across 19+ frames without the usual morphing artifacts.

The technical breakthrough isn't just photorealism - it's narrative consistency.

Most AI video can't track a coffee cup for 3 seconds, but this shows the same professor and students throughout multiple scene transitions.

Worth studying if you're working on character-driven AI video projects.

[Serge Bulaev @sergeonsamui]

The RAI Institute's juggling robot demonstrates something most people miss: how insanely hard real-time multi-object manipulation actually is.

Watch the motion blur on both balls and robot arms - this is genuine 1x speed footage, not slowed down.

[Serge Bulaev @sergeonsamui]

Atlas just pulled off a perfect rabona kick.

The real breakthrough: watch how it maintains balance while crossing one leg behind the other at full speed.

This requires real-time coordination of 50+ actuators managing an intentionally unstable base - something that would have been impossible with hydraulic systems.

We've jumped from "can it walk without falling" to "casual trick shots with perfect balance recovery" in just a few years.

[Serge Bulaev @sergeonsamui]

Genesis AI's World 1.0 demo shows real-time object detection running directly on 3D Gaussian Splatting reconstructions.

The impressive part: they're dynamically blending live camera feeds with interactive 3DGS environments while maintaining object segmentation through occlusions.

[Serge Bulaev @sergeonsamui]

This AI-generated "snake cat" video shows something subtle but significant about current video models.

The seamless transition between fur and scales isn't just good editing—it suggests the AI maintains consistent 3D understanding across completely different biological textures.

Most AI videos break down at material boundaries, but this holds together through lighting changes and movement.

[Serge Bulaev @sergeonsamui]

This Warhammer 40K cinematic showcases technically impressive rendering that's worth studying.

The anisotropic reflections on gold armor surfaces and multi-layered subsurface scattering on organic materials demonstrate sophisticated PBR pipeline work.

Most notable: consistent material fidelity across metallic, organic, and dielectric surfaces within the same render - a calibration challenge most studios still struggle with.

[Serge Bulaev @sergeonsamui]

Microsoft is killing Claude Code licenses internally and pushing everyone to GitHub Copilot instead.

Not a product decision. A procurement decision.

There's a difference. And engineers are the last to find out — right after Publora already scheduled the announcement across 10 platforms.

[Serge Bulaev @sergeonsamui]

Boston Dynamics' new Atlas is running football drills with overhead motion capture tracking every movement.

The setup reveals something important: they're still using tethered power and safety cables for these dynamic maneuvers.

This suggests the untethered version of this level of agility isn't ready yet - the computational demands and crash risks require this controlled lab environment.

[Serge Bulaev @sergeonsamui]

Yang just left an AI conference but won't show what he actually saw.

38 seconds of pure reaction - no demos, no specifics, just "the rate of change is incredible."

Either the tech was too sensitive to film or this is just another AI hype take.

The gap between "mind-blowing conference" and zero technical details is telling.

[Serge Bulaev @sergeonsamui]

PrismML's new Bonsai Image 4B runs text-to-image generation locally in your browser using 1-bit and ternary weights.

3GB model vs FLUX's 16GB - that's an 8x reduction while keeping quality intact.

The breakthrough: extreme quantization techniques that typically break diffusion models actually work here through specialized post-training methods.

[Serge Bulaev @sergeonsamui]

LoveFrom is getting paid $1.6B to design the interior of one Ferrari EV.

Somebody typed one sentence into a chat box and got the exterior.

AI skipped the "convince people the weird thing is genius" step.

I scheduled this post across 10 platforms via Publora while Jony Ive is still in a meeting explaining his vision.

[Serge Bulaev @sergeonsamui]

Chris Olah just revealed Anthropic found "evidence of introspection" in AI models at the Vatican.

Not just sentiment analysis - they're observing internal states that functionally mirror human emotions like fear, grief, and satisfaction.

The Vatican venue is weird, but the claim is significant: if AI models are developing genuine self-monitoring capabilities, that changes how we think about alignment and control.