Kostya | AI @KostyaAI

Building with AI • Learning in public Exploring agents, tools, systems & AI research Europe Joined January 2026

Tweets

2K
Followers

1K
Following

2K
Likes

3K

Ember | AI & Tech Sparks @EmberAITech

a hour ago

AI’s infrastructure boom is starting to show up on consumer price tags. DRAM prices have risen roughly 5× over the past year as memory makers prioritize AI servers and high-bandwidth memory. PC, phone and console makers are already responding with higher prices, lower specs or both. TrendForce says conventional DRAM contract prices were still expected to rise another 13–18% in Q3 2026, even as consumer demand weakened. For years, better memory quietly made electronics cheaper and more capable. Now AI data centers are competing for the same industrial capacity. The cost of frontier AI is beginning to leak into hardware that has nothing to do with running frontier models.

0 1 3 26 0

View Details

Kostya | AI @KostyaAI

2 hours ago

The useful record is not the model’s diary. When CoT gets thinner, the audit is the side effects: tool calls, diffs, API requests, permission grants, deny events. If that log lives only in the model’s narration, you don’t have oversight. You have a story the model can rewrite. Own the event log. Make the harness write it before the model speaks. The model is an input. The harness is the control plane.

Kostya | AI @KostyaAI

8 hours ago

x.com/i/article/2096…

1 2 11 132 2

2 3 8 60 2

View Details

Ember | AI & Tech Sparks @EmberAITech

3 hours ago

Wall Street is turning AI’s electricity appetite into a bond market. Reports put the scale near $61 billion as data-center power demand gets packaged into investable debt. At the same time the physical gap remains large: roughly 148 GW of capacity announced since 2024 against only about 12 GW actually running. Capital is moving faster than transformers, grid connections and permits. The binding constraint is shifting from model weights to megawatts — and the financing structures are already treating those megawatts as a new industrial asset class. Frontier AI no longer looks only like software. It looks like long-duration infrastructure that needs its own credit markets.

0 1 2 35 0

View Details

AIstasiia @aistasiia

5 hours ago

Only the olds will get this one. Saturday night. Way too many Microducks. One meme from 2013 that apparently refused to die. Harlem Shake.

2 1 31 416 5

View Details

Kostya | AI @KostyaAI

9 hours ago

In the Ember Article, I wrote: “Format is state — every downstream step must respect it.” I learned that after Ember generated a strong long-form X draft, then shrank it back to a normal post during a rewrite or evidence decision. The model wasn’t the core problem ↓

Kostya | AI @KostyaAI

2 weeks ago

Not another AI writer. A closed loop: signal → draft → verify → publish → measure → learn. x.com/i/article/2091…

2 7 64 209K 15

1 3 12 145 6

View Details

Kostya | AI @KostyaAI

8 hours ago

x.com/i/article/2096…

1 2 11 132 2

View Details

Kostya | AI @KostyaAI

8 hours ago

That’s the hard part, yes. Once you have multiple models in the chain you lose the single-run assumption. The workable approach is a shared trace ID that follows every tool call and model hop, plus an external event log of side-effects. The models themselves don’t have to narrate — the instrumentation layer does.

0 0 0 7 0

View Details

Kostya | AI @KostyaAI

12 hours ago

10 hours ago

@techniahqrobot Natural‑language task description can cut integration cycles dramatically, yet rigorous validation of the generated code remains essential for safe cobot deployment.