Victor Kushnerov

🤖 AI-based solutions implementation and integration engineer. I help companies adapt AI technologies in products and business processes 🧠⚙️

[Victor Kushnerov @filimo]

@karpathy 8/ Rambling lowers the cost of providing context.

Dialogue goes one step further: it improves the context before implementation starts.

Perhaps the future interface for coding agents is not a better prompt box, but a short design conversation.

[Victor Kushnerov @filimo]

@karpathy 1/ I like the idea of “rambling” into an AI for 5–10 minutes instead of writing a structured prompt.

But I think there is an even more useful version of this workflow: not a monologue, but a voice conversation.

[Victor Kushnerov @filimo]

@karpathy 7/ In this model, the voice assistant acts almost like a product manager, architect, and rubber-duck debugging partner before the coding agent begins changing files.

This could prevent a lot of confidently implemented but incorrectly understood tasks.

[Victor Kushnerov @filimo]

@karpathy 6/ That package can then be handed directly to Codex or another coding agent.

So the workflow becomes:

voice discussion → collaborative task discovery → structured specification → coding agent → review

[Victor Kushnerov @filimo]

@karpathy 5/ At the end, the model turns the entire conversation into an implementation package:

— objective and business context
— functional requirements
— constraints and non-goals
— architecture decisions
— open questions and risks
— acceptance criteria
— implementation plan

[Victor Kushnerov @filimo]

@karpathy 4/ The important part is that the model is not merely transcribing or restructuring my thoughts.

It is helping me discover what the task actually is.

The specification emerges from the dialogue.

[Victor Kushnerov @filimo]

@karpathy 3/ So instead of recording one long ramble, I could discuss the task with a voice model.

I explain what I want in my own words. The model asks clarifying questions, challenges assumptions, points out missing cases, and suggests alternative approaches.

[Victor Kushnerov @filimo]

@karpathy 2/ When I start a new software task, I rarely understand it well enough to write a perfect specification immediately.

I usually have a goal, some constraints, fragments of context, and several ideas that may still contradict each other.

[Victor Kushnerov @filimo]

Switched to GPT-5.6 Sol because it performs better on long, complex engineering tasks.

For Python backend, LLM integrations, agent systems, and Codex, it follows instructions more reliably and makes fewer incorrect assumptions.

Less time fixing AI. More time building.

[Thomas Wolf @Thom_Wolf]

we've seen nothing yet! hosted a 9-13 yo vibe-coding event w. @robertkeus this w-e (h/t @antonosika @LovableBuild)

takeaway? AI is unleashing a generation of wildly creative builders beyond anything I'd have imagined

and they grow up *knowing* they can build anything!

[Prakash @8teAPi]

Roombas were a mistake https://t.co/XxSrlEdjiz

[Maisa @maisaAI_]

DEMO TIME! Customer service:
Help a customer with a question about an order that did not arrive. This time the customer accidentally did not write the order ID correctly 😯

[Tolga Bilge @TolgaBilge_]

Somewhat interesting advertising choice from Anthropic, comparing their newly released Claude 3 to GPT-4 on release (March 2023).

According to Promptbase's benchmarking, GPT-4-turbo scores better than Claude 3 on every benchmark where we can make a direct comparison.

[Ana Mostarac @anammostarac]

Microsoft’s board right now

[Dave @dmvaldman]

I was curious if I could practice my Russian with ChatGPT-audio. Yup. Speakers move between languages effortlessly.

[Meng To @MengTo]

midjourney final design

[outlandish thug @michaelhirschbb]

french people really have nothing better to do? i know nobody is speaking in this video but it must be france https://t.co/bsWXk5RLAp

[SEM🦇🔊 @0xSEM]

Had a nice conversation with #ChatGPT and #DiffusionBee

It built me this application just by friendly asking for it and iterating, starting from a random recipe string generator

All additional features were suggested and implemented by ChatGPT itself

NCA! (No Cooking Advise)

[Stian Håklev @houshuang]

Language learning with ChatGPT - playing with table format

[Victor Kushnerov @filimo]

@swiftbysundell observationTask = Task {
Task {
for await users in list.$users.values {
updateTableView(withUsers: users)
}
}
Task {
for await users in list.$users.values {
updateTableView(withUsers: users)
}
}
}
....
observationTask.cancel()

[Victor Kushnerov @filimo]

@swiftbysundell I have a question about nested Tasks. Why won't they be canceled when I cancel them parent Task?