Aluda Saliashvili
A personal site where I track projects I'm building, ideas I'm exploring, and how I think about what's coming in AI and technology.
In flight right now
What I am building right now
Open any card for the full system architecture, data flows, and failure modes.
Core Principles of the New Era
Plan deep, build fast
An excellent, well-researched plan eliminates entire categories of iteration, bug fixes, and refactoring. Tip: use different AI models to refine, critique, and upgrade plans before writing a single line of code.
Bias for action
Decisions compound. Shipping something imperfect today beats shipping something perfect next quarter. Velocity is a moat.
High agency
Take initiative. Don't wait for perfect conditions. Figure it out, start walking up the learning curve, find a way.
Value lives in sophistication
If a workflow is easy to build, it's already table stakes. True value is unlocked in complex, domain-specific implementations that require deep thinking.
Think in systems
Individual components are easy. The hard part is understanding how they interact — feedback loops, failure cascades, emergent behavior at scale.
Creative cross-domain thinking
The best solutions come from connecting ideas across fields. Game theory applied to product strategy. Finance patterns applied to AI evaluation. Unusual combinations compound.
Creativity is a skill, not a gift
Inspiration matters, but creative thinking is a trainable muscle. Build it through volume, iteration, and exposure to adjacent fields. Prototype fast, test widely, refine relentlessly.
Think strategically
Think a few steps ahead based on what is likely coming. Have a clear vision with defensible conviction about where things are going, then move early.
The Signal Map
What the people building these systems actually believe, cross-referenced with data trends. Click any category to see the chronological timeline.
Predictions
Specific, dated, falsifiable. The point is to be wrong in public so the thinking improves.
AI coding tools will produce production-quality code 10x faster than today with dramatically fewer bugs — making a single developer as productive as a small team.
Humanoid robots will be able to perform most physical tasks humans can perform, using fast vision-based learning models for rapid skill acquisition.
AI models will be systematically optimized to make new medical and scientific discoveries, with the pace of discoveries accelerating. The limiting factor will be human capacity to actually run the physical experiments.
AI-powered digital fraud (deepfakes, voice cloning, autonomous scam agents) will explode in 2026, with losses to fraud increasing 3-5x over 2025 levels.
As new models start approaching and exceeding AGI-level capability, frontier labs will gatekeep access — keeping the most powerful models restricted internally or to select partners while offering weaker variants publicly, creating a widening two-tier AI ecosystem.
LLM benchmarks will become increasingly less relevant as a measure of real capability. Subjective usefulness of AI models in everyday professional tasks — how much they actually help you get work done — will be the metric that matters.
The cost of reasoning per token will continue to decline dramatically, but the actual price of AI services will continue to rise. Jevons paradox: as unit costs drop, demand explodes faster than supply can keep up, driving total spend higher.
AI-generated content (text, images, video, audio) will become indistinguishable from human-generated content and will start becoming ubiquitous across media, marketing, and professional communication.
The majority of enterprise software companies will ship AI agent features as core product capabilities, not add-ons. Companies without an agent strategy will lose market share measurably.
At least one major geopolitical crisis or international incident will be directly attributed to AI systems — whether through autonomous decision-making, AI-generated disinformation at scale, or AI-enabled cyberattack.
Stack & Tooling
Build Path
A practical path from curiosity to shipping useful AI products.
Don't just write code—direct it. Use Claude Code or Antigravity. The intuition from actual usage is worth more than any course.
Learn APIs, tokens, context windows, system prompts, and structured output. But also understand the SDLC: version control, hosting, CI/CD, databases. Use coding assistants to help you along — even send them screenshots when you're confused. That's how you learn fastest.
Start with a web app. The gap between 'works on my machine' and 'people use this' is where all the learning happens. Pick a real problem, build a Next.js or React app, deploy it, and put it in front of users.