Serhii

Software Engineer

[Serhii @salnkvs]

Skip an hour of Netflix.
Watch this 2-hour Stanford lecture instead.
It will teach you more about how LLMs like ChatGPT and Claude are actually built than most people at top AI companies learn in their entire careers.

[Serhii @salnkvs]

Anthropic Quant Andrej Karpathy:
"Most people prompt loops they don't understand.

The ones who stop prompting and start designing the loops end up faster than everyone else."

Shopify CEO tried the Karpathy Loop overnight and woke up to a 19% better model that was half the size.

He released 3 files totaling 630 lines. It got 66,000 stars in a month.

The agent ran 700 experiments while he slept and found 20 improvements he had missed in 20 years of hand-tuning.

Then researchers put a loop on top of the loop and got 5x better results.

"Humans get tired after experiment 12. The agent doesn't get tired at all."

[Serhii @salnkvs]

Amazon AI engineer explains how to build agentic loops from scratch in under 10 minutes.

This is one of the clearest breakdowns you’ll find.

Loop + model + agent + hooks + MCP equals an autonomous, self-improving system.

People pay $500 for courses that teach less.

Watch it

[Serhii @salnkvs]

@mfishbein these are solid prompts to squeeze maximum value out of Fable 5 while it's still included

[Serhii @salnkvs]

@DataChaz Claude CEO when?

[Serhii @salnkvs]

@k2sbhai it makes it trivial to benchmark them against each other without juggling multiple accounts or billing
good move

[Serhii @salnkvs]

@chesny great example of how cheap inference + good prompting can compress costs dramatically in knowledge work

[Serhii @salnkvs]

@PawelHuryn next time they ask for ID verification

[Serhii @salnkvs]

@eng_khairallah1 structured loops change everything about how you work with these models

[Serhii @salnkvs]

@nateherk pricing pressure is real once the free window closes

[Serhii @salnkvs]

Head of Codex:
“90% of people at OpenAI use Codex.

Not 90% of engineers - 90% of the entire company.

We build an agent → create a loop for it → that agentic loop does 40% of the coding at Codex.”

In this 1-hour podcast, the Head of Codex reveals the full workflow used by every OpenAI engineer.

Codex plus loops plus harness plus LLM ops - the secret stack.

Watch it today 👇

[Serhii @salnkvs]

@milesdeutscher most people are still prompting Fable 5 like it's Sonnet and leaving a lot of performance on the table

[Serhii @salnkvs]

Anthropic engineer:
"Most people will use Sonnet 5 and Fable 5 wrong.

You can set them up right in one afternoon and stop overpaying every single day."

In this 29-minute session he breaks down every Claude model and explains why running them all the same way is a mistake.

Watch it.

[Serhii @salnkvs]

@undefinedKi paying people to get hands-on with Claude while doing meaningful nonprofit work is a strong way to build both talent and goodwill

[Serhii @salnkvs]

@atomic_chat_hq Fable 5 clearly wins on output quality here, but the 6x cost premium over Opus 4.8 is brutal for most use cases

[Serhii @salnkvs]

@bcherny this is a big quality-of-life upgrade
Claude Code is moving fast

[Serhii @salnkvs]

@Bencera pricing pressure from strong open models is real and will force cuts at the frontier labs

[Serhii @salnkvs]

Spotify Chief Architect Niklas Gustavsson:

"Once we implemented loops in our workflow, our agent success rate went from 20-30% to 80%."

In this 26-minute conversation with Claude Code creator Boris Cherny, he explains how Spotify runs AI agents across 20 million lines of code.

Today 73% of their code is written by AI, most of it merged without a human ever seeing it.

The model matters less than the loop you build around it.

Watch it.

[Serhii @salnkvs]

My friend makes $1.2 million a year as an Anthropic engineer.

I asked him how he learned prompting so well.

He sent me a video that was never supposed to get out - their core team's prompting playbook.

You won’t find anything better on prompting than this.

I watched it last night and realized I’ve been using Claude completely wrong for two years.

Watch it.

[Serhii @salnkvs]

Anthropic CEO:
"I have engineers within Anthropic who don’t write any code. They just let Claude write the code and they edit it and look it over."

"At Anthropic writing code means designing the next version of Claude itself, so we essentially have Claude designing the next version of Claude. Not completely, but most of it."

In the last 52 days, the Claude team dropped 50+ major feature launches.