yaros

Research AI tools and sharing what works.

[yaros @0xYaros]

https://t.co/PQqzgSssRw

[yaros @0xYaros]

Visa, in a joint report with Artemis, published this week:
"Cards and stablecoins are not in competition. They are becoming different components of a single system."

cards handle macro-commerce — bookings, subscriptions, anything human-sized. stablecoins handle micro-commerce — sub-$1 payments between AI agents, where a card's fixed fee would cost more than the transaction itself.

one protocol alone, x402, has already moved $15 million across 109.6 million payments since May 2025 — an average transaction under 14 cents.

if you run a paid API or MCP server, you can already accept x402 payments directly. no merchant account, live today.
most people still think "AI agents paying each other" is hypothetical.

it's already running millions of real transactions, most too small for a card network to bother with.

Full setup breakdown, below.

[yaros @0xYaros]

https://t.co/PQqzgSssRw

[yaros @0xYaros]

@Raytar you should learn from the best. this is a very helpful video

[yaros @0xYaros]

A Moonshot AI executive, explaining why 2.8 trillion parameters actually matters:

"more knowledge and patterns in its brain, think deeper, answer more accurately."

the numbers back it up: on GDPval-AA v2, a benchmark of real-world tasks across 44 professions, K3 scored 1,687 — beating Claude Opus 4.8's 1,600.

the pricing: $0.30 per million cached input tokens, $15 per million output. a fraction of what closed frontier models charge for similar performance.

full open weights drop July 27 — free to self-host, no API bill at all if you have the hardware.

most people are still paying premium prices for a model this one already beats on real-world tasks.

Full breakdown of the benchmarks and how to access it today, below.

[yaros @0xYaros]

https://t.co/PQqzgSssRw

[yaros @0xYaros]

Apple, the most valuable company on the planet, just admitted it couldn't build this alone.

$1 billion a year — up to $5 billion total — to license a custom 1.2-trillion-parameter Gemini model, just to make Siri actually work.

here's the part most coverage skipped: it's not a marriage, it's a bridge. Apple already has a date to walk away from it — 2027, once its own models catch up. most people think Apple paid Google to fix Siri for good. it didn't. it paid Google to buy itself two years.

Full breakdown of the deal — and what it means for anyone building on Siri, Gemini, or both — below.

[yaros @0xYaros]

https://t.co/PQqzgSssRw

[yaros @0xYaros]

STOP PAYING PER TOKEN FOR EVERYDAY CODING.

- install Ollama (2 min)
- pull a small local model like Ornith-1.0 9B
- runs on almost anything
- point your editor at localhost instead of an API key
- run your usual tasks and compare the output

one developer switched their daily workflow to this setup. $11,000 saved. 93% cost reduction. performance stayed competitive with GPT-5.5 on regular coding work. it won't replace frontier models on large multi-file refactors.

but for day-to-day coding, the gap is often smaller than the price difference. most people are still sending every prompt to a cloud API. the ones who aren't have already covered their hardware in bills they stopped paying.

Full breakdown of what actually runs well on 8GB, 16GB, and 32GB+ — below.

[yaros @0xYaros]

https://t.co/PQqzgSssRw

[yaros @0xYaros]

@karthikponna19 xd

[yaros @0xYaros]

@0xF1ction he is 1 of the smartest man for real

[yaros @0xYaros]

@AnatoliKopadze It's true that many people think that, over time, AI will replace many professions, but in reality, they will be replaced by people who can integrate AI into their work.

[yaros @0xYaros]

Google just leaked Gemini 3.5 Pro pricing — and it completely destroys the biggest myth in AI right now.Everyone assumed “2M context window = cheapest way to handle huge documents & RAG.”
Reality: $12–15 input / $36–60 output per million tokens.

More expensive than GPT-5.6 Sol and Claude Opus https://t.co/kTbdj8w2Co e-commerce team tested “replace your vector DB with raw context” for three weeks:
• Latency ×3
• Cost per query ×8
• Quality dropped 12 points

The 2M window is real and powerful — but only for whole-codebase audits, long legal reviews, and complex agent sessions where losing context is unacceptable.

It is not a free upgrade for everything you do today.Rule that already saved me thousands:
Default to Flash. Escalate to Pro only when the task actually needs the context.

Full cost breakdown + comparison table dropping as soon as official pricing lands.

[yaros @0xYaros]

https://t.co/PQqzgSssRw

[yaros @0xYaros]

Andrej Karpathy, on X, describing his coding agents:
"They don't manage their confusion, don't seek clarifications, don't push back when they should."

one day later, a developer named Forrest Chang turned that complaint into a single markdown file — four rules, zero code.

it's now sitting at over 220,000 combined stars across two repos.

most people read a viral complaint and scroll past it.
the ones who don't turn it into the fix before the outrage even cools down.

Save the file, then read how to write your own skill in the article below.

[yaros @0xYaros]

https://t.co/PQqzgSssRw

[yaros @0xYaros]

Kevin Gabeci ran, priced, or watched a friend run all ten of the "AI agent" pitches flooding your timeline right now:
"I deployed this AI agent and made $47,000 last month while I slept."

he priced all ten, honestly, no best-case fantasy math. most pay less than a part-time shift. three pay nothing at all.

Most people think the $19 agent gets them the $47k result.
It doesn't.

The agent works. That part's real. What doesn't work is the sentence right after it.

Read the honest breakdown, then build your own for free below.

Ten fake screenshots. Zero real receipts. That's the whole story.

[yaros @0xYaros]

https://t.co/PQqzgSssRw

[yaros @0xYaros]

Andrej Karpathy joined Sarah Guo for a 65-minute conversation on the future of coding agents:

06:15 – What mastery of coding agents actually looks like
15:51 – Why he built AutoResearch, an agent loop that runs experiments overnight
22:45 – Which skills stay relevant once agents write most of the code
28:25 – Model speciation — why different models are starting to diverge
37:28 – The actual jobs market data, not the hot takes about it

This one interview replaces a dozen threads explaining what "agentic engineering" even means.

Watch it today, then save the loop framework below.

[yaros @0xYaros]

https://t.co/PQqzgSssRw

[yaros @0xYaros]

Boris Cherny, head of Claude Code at Anthropic:
"I don't prompt Claude anymore. My job is to write loops."
he said it on stage in June, and it's been circulating through developer communities since. no more turn-by-turn prompting — he builds systems that discover work, assign it, verify results, and know when to hand off.
most people are still typing one prompt, reading the output, typing the next. the ones who aren't are already building loops.
Watch the talk, then save the framework below

[yaros @0xYaros]

Most "AI workflows" I see are still just a chatbot step bolted onto a process built for humans.

I rebuilt one of my own workflows this way.

The workflow is the agent.

Steps are just what it calls.

Cut two manual handoffs in the process.