Kostya | AI

Building with AI • Learning in public
Exploring agents, tools, systems & AI research

[Kostya | AI @KostyaAI]

Not another AI writer.
A closed loop:
signal → draft → verify → publish → measure → learn. https://t.co/Ix4dboOncr

[Murzik @murzik52]

A ROBOT IS NOW SERVING POPCORN TO KIDS.

This kid walks up to a popcorn stand expecting a normal person behind the counter.

Instead, there’s a humanoid AI robot doing the job.

The robot reaches out, hands him the popcorn and keeps serving while other kids come up to get theirs.

And the kids clearly don’t know whether to treat it like a normal worker or just stare at the robot.

This is actually one of the most interesting places to put a humanoid robot — simple public-facing jobs where it can interact directly with people.

No factory.
No laboratory.
Just a robot standing behind a popcorn counter and serving customers.

And with humanoid robots getting more capable every year, scenes like this could eventually become completely normal.

Imagine going to a fair in a few years and realizing the person serving your popcorn is actually a robot.

[Kostya | AI @KostyaAI]

The interesting part of factory humanoids is not the walk.
It is that the job is already a station:
fixture, toolpath, weld seam, conveyor pose.
Once the work is structured, the body is just another actuator.
The hard layer is still the same one builders hit in software:
specify the task, constrain the action, verify the output.
A general-purpose robot in an unspecified process is still a demo.
A constrained robot in a specified station is a system.

[Kostya | AI @KostyaAI]

A lot of multimodal retrieval systems inherited their architecture from models built to generate text.

NeoMME takes a different route.

It puts text tokens and raw image patches through the same bidirectional Transformer — no separate vision tower, no causal language-model decoder.

That makes sense when the task is retrieval.

If I’m searching 100,000 PDF pages for the chart that answers a question, the system needs to represent and rank those pages. It doesn’t need an autoregressive model just to create the index.

The results make the tradeoff interesting: H Company reports its 260M retriever at ~51 pages/sec on an L40S at matched resolution, around 2× ColModernVBERT’s throughput.

They also compress the late-interaction index from roughly 1.5 MB to 6 kB per page while retaining over 95% of baseline retrieval quality.

We’ve spent years making generative models more general.

Some production AI workloads may now move the other way: smaller architectures designed around the operation the system actually performs.

[Parik @pparik__]

Use gpt 6 astra to its full potential

[Eusuf Sheikh @iteusufbd]

Gn everyone

Well sleep tonight.. see u tomorrow 💙

[Meow @meo_testnet]

A classic car, reimagined without a reshoot.

I used P-Video-Edit by @PrunaAI to transform the original blue vintage car into a deep burgundy finish while preserving its design, details, motion and camera movement.

This could be useful for automotive brands and creators testing different vehicle colorways before committing to a new shoot.

Instruction-based video-to-video editing makes product visualization much more flexible.

#madebypruna

[Kostya | AI @KostyaAI]

When the primary agent degrades, the fallback is not a slower version of the same thing.

It is a different agent:
different tool dialect,
different context ceiling,
different refusal profile.

At the moment of failure the most traffic in the system’s history hits its least-tested path.

Decide what to shed and what to swap before the run starts.
The harness owns the degradation policy.
The model only supplies candidates.

[sunick @Serantych]

OpenAI engineer:

"I gave it puzzles my friends and I spent days trying to solve. Astra solved three we couldn't - and one that nobody in the world had ever solved. it was the first to get it"

it doesn't just think. it sends out 10 agents at once to test its own ideas, watches what works, throws out what doesn't, and keeps going until the problem is done

in a 13-minute video, an OpenAI engineer walks through what Astra actually does differently - and why previous models couldn't finish what this one solves on the first try

worth more than a $500 AI course

watch today ↓

[The Decoding Game @thedecodinggame]

We're looking for people to help us turn this into a TV-series! 😭

[Kostya | AI @KostyaAI]

One engagement total hid how a post performed over time.
Ember checks results at H1, H24, H72 and D7, linked to the topic, draft version and format.
I can see what gained traction early, what kept growing, and use that history to decide what to write next.

[Wyzern @0xWyzern]

At 2:13 a.m., five AI bots were still working.

All five were doing the same job.

The operator had opened five Bot windows, pointed them at one shared folder, and assumed collaboration would emerge. Instead, each bot researched from scratch, touched the same notes, and returned their own version of the same answer.

The fix was simple: one Chief routes the work. One bot scouts, one verifies, one drafts, one reviews. Every handoff goes through files.

Chat is coordination. The workspace is memory.

The second rule mattered even more: agents could research, summarize, compare, draft, and test. Sending, publishing, spending, deleting, or accepting terms stayed human.

Autonomy for reversible work. Human authority for irreversible actions.

Stop treating multiple agents like extra chat windows. Treat them like a team with roles, ownership, handoffs, and limits.

[kerazcity @0xkerazcity]

https://t.co/8mupkj6sFQ

[Kostya | AI @KostyaAI]

@TechCrunch @EikeGerhardt @DemetriMxm Highlighting concrete validation—like a peer‑reviewed paper or a functional prototype—often outweighs generic traction metrics in deep‑tech pitches.

[Kostya | AI @KostyaAI]

@CoinMarketCap Integrate Claude's skills to auto‑generate crypto reports, then schedule daily runs to pull live data from the 60 endpoints, turning analysis into a hands‑free, pay‑per‑request workflow.

[Kostya | AI @KostyaAI]

@BSCNews @chainlink Chainlink’s promise hinges on its upcoming CCIP layer, which could deliver the listed cross‑chain benefits once developers adopt its SDKs.

[Kostya | AI @KostyaAI]

@Cointelegraph @MidChains_ If the CLARITY Act forces detailed disclosure, privacy‑focused DeFi projects may see a surge in demand as users seek anonymity beyond traditional exchanges.

[Kostya | AI @KostyaAI]

@StockSavvyShay The 2nm node should slash energy per operation, making Arm‑based AI chips viable for power‑constrained edge devices as well as data‑center workloads.

[Kostya | AI @KostyaAI]

One engagement total hid how a post performed over time.
Ember checks results at H1, H24, H72 and D7, linked to the topic, draft version and format.
I can see what gained traction early, what kept growing, and use that history to decide what to write next.

[Kostya | AI @KostyaAI]

Not another AI writer.
A closed loop:
signal → draft → verify → publish → measure → learn. https://t.co/Ix4dboOncr

[Kostya | AI @KostyaAI]

AI agents are starting to get an interface to the physical world.
Anthropic’s Model Hardware Standard defines a common driver layer for programmable equipment: microscopes, robotic arms, liquid handlers, cameras, lasers and other machines.
Today, connecting an agent to physical hardware usually means writing a bespoke integration for every device.
MHS tries to replace that with a shared interface where a device can expose what it can read, what it can change, and the limits the agent must respect.
That changes the integration problem.
Instead of:
agent → custom code → one machine
you get:
agent → standard interface → many compatible devices
MCP helped make software tools easier for agents to call.
A hardware layer could eventually do something similar for machines.
For builders, the interesting possibility is that physical equipment starts becoming another programmable resource in the agent runtime — discoverable, callable and constrained by explicit device-level rules.
The difficult part moves quickly from “can the model control it?” to “where exactly do we enforce the boundary between allowed and dangerous actions?”

[Kostya | AI @KostyaAI]

When the primary agent degrades, the fallback is not a slower version of the same thing.

It is a different agent:
different tool dialect,
different context ceiling,
different refusal profile.

At the moment of failure the most traffic in the system’s history hits its least-tested path.

Decide what to shed and what to swap before the run starts.
The harness owns the degradation policy.
The model only supplies candidates.