Ron Kavunovski

Silicon Valley is trying to win AI with the SMARTEST models❗️ China may be trying to win with the most USABLE ones ↓ That’s why Moonshot AI is interesting. The Chinese AI lab just raised ~$2B at a $20B valuation. But the real story isn’t the round. It’s the strategy. OpenAI, Anthropic and Google are competing at the frontier: smarter models, better reasoning, more powerful systems. Moonshot is attacking a different layer: • open-weight models • lower inference cost • “good enough” performance for many real use cases That matters because most developers and companies don’t always need the absolute best model. They need the model that is reliable enough, accessible enough, and cheap enough to build on. If that happens, the AI market could become less centralized than people assume. Not because the best model loses. Because the most usable model sometimes wins.

Two ex-8200 engineers sold to Cisco for $400M Alon Jackson and Idan Gour founded Astrix Security in 2021 to fix a problem nobody was talking about: AI agents running inside companies with zero oversight. This month, Cisco bought them.

@Zephyr_hg @grok TL;DR?

ANDREJ KARPATHY: "the internet is terrible. A huge amount of slop and garbage from all the corners of the internet." Type 1: you ask generic Claude generic questions, you get back generic internet-trained answers. Type 2: you brief Claude with your business, your voice, your audience. Same model. Different fluency. The gap between the two types isn't intelligence. It's setup. Bookmark this for the next time someone says Claude is generic and can't replace a real specialist. The article below has the 7 specific setups that turn the internet-slop model into a specialist for your work.

He broke every AI model — now they're funding him 
Description: In 2024, Denis Shilov bypassed safety filters on ChatGPT, Claude, and Gemini with one prompt. He published the exploit, then built White Circle — an AI monitoring platform now backed by $11M from execs at OpenAI, Anthropic, Mistral, Hugging Face, and Datadog.

@ycombinator @paulg @grok sum it up for me

Paul Graham (@paulg) whether founders should move to Silicon Valley, and what it takes to build a startup hub anywhere else. Live from our YC | Stockholm event on April 29, 2026. 01:01 – Why the Big Center Matters 02:45 – The Power of Serendipitous Meetings 04:36 – Investors Move Faster in the Valley 06:03 – Respect Follows the Move 07:59 – The Dropbox Story 09:10 – Measuring Yourself Against Big Fish 12:21 – Silicon Valley's Pay-It-Forward Culture 15:36 – How to Help Stockholm Thrive 17:24 – YC as the Optimal Path 19:54 – Could Stockholm Become The Silicon Valley of Europe?

@AlexFinn @grok What are the main takeaways?”

3 ways to make tons of money on apps in 2026:

$650M to build an AI that fires its own engineers? Richard Socher (ex-Salesforce Chief Scientist, founder of you.com) just came out of stealth with Recursive Superintelligence — an AI model designed to identify its own weaknesses and retrain itself without humans. Backers include Peter Norvig and Cresta's Tim Shi. Follow for more on what's actually happening in frontier AI. #AI #frontiermodels #ASI #AGI #startup #deeptech #venturecapital

Elon fired him on day one. Three years later he's worth $2B. Parag Agrawal was CEO of Twitter when Musk took it over and walked him out with security. He just raised $100M Series B at a $2 billion valuation led by Sequoia. His new company, Parallel, builds web infrastructure for AI agents — customers include Harvey, Notion, Clay, Opendoor. Follow for more on what's actually happening in frontier AI. #AI #startup #aiagents

@igus_ai @grok real?

Un estudiante chino viviendo en Japón convirtió $0,90 en $408.292 tradeando en Polymarket. Y casi nadie está hablando de ello. Su cuenta se llama “Gravia”. Solo llevaba 2 días en la plataforma. Pero lo más absurdo es esto: Dice haber construido un bot con Claude para tradear el mercado BTC UP/DOWN 5MIN. El sistema: • Obtiene datos en tiempo real desde Binance WebSocket + velas 5M • Cruza señales de TradingView + flujos de exchanges de CryptoQuant • Usa un force-graph con 100 nodos y 180 conexiones para detectar convergencia BULL/BEAR • Detecta retrasos entre el spot price y el CLOB de Polymarket • Ejecuta operaciones en menos de 100ms antes del repricing • Puede lanzar más de 1000 órdenes por segundo • Captura entre 0,3% y 0,8% por trade Pero aquí viene lo importante: El edge no está en “predecir Bitcoin”. Está en explotar microdesfases entre: • precio spot • señales del mercado • y repricing del order book de Polymarket Y según él, el bot directamente evita operar si: • no hay edge • la liquidez es baja • las señales se contradicen • o se alcanza el límite diario de riesgo También tiene controles bastante agresivos: • Riesgo por operación: 0,5% • Límite diario: 2% • Hard stop: -0,4% • Corre localmente • No usa GPU • No depende de cloud

@elonmusk @grok Sum it up

The latest 𝕏 algorithm has been published to GitHub github.com/xai-org/x-algo…

These three 22-year-old dropouts just beat Mark Zuckerberg to become the YOUNGEST billionaires in history? 🤯

Anthropic’s “TOO DANGEROUS” Mythos model didn’t actually discover anything new? 🚨

9/ Genuine questions, not rhetorical: — If you've built a second brain, what's your rule for what doesn't go in? — How are you handling Context FOMO? — Anyone actually using OpenClaw / Hermes / GBrain alongside an LLM Wiki, or is that just stack envy?

1/ Built a second brain on Karpathy's LLM Wiki pattern for 3 different parts of my life: → master's thesis research → a 75K-follower content channel → a construction-tech startup

8/ I looked at all of them. Honestly? For my three use cases, Karpathy's LLM Wiki + Claude Code is enough.

7/ Layer 2 is the louder one right now. OpenClaw. Hermes Agent. Garry Tan's GBrain. Every week another agentic-memory stack drops, and the implicit message is: if you're not running it, you're falling behind.

6/ Context FOMO has two layers. Layer 1 — every conversation, voice memo, idea you don't pipe into the brain feels wasted. Layer 2 — every system you're not running feels like a missed compounding curve.

5/ Then it gets weird. I caught myself mid-conversation with my co-founder thinking "damn, we should be recording this." The absence of capture now feels like a loss. Meet Context FOMO.

4/ The startup brain Every customer interview gets transcribed and ingested. After ~10 calls the same objections, the same words, the same workflows surface on their own. Validation stops being vibes. It becomes a search query.

3/ The content brain Uploaded scripts + performance data for every video I've published. Patterns I couldn't see from the inside surfaced on their own — hooks I underuse, a virality-vs-authority tradeoff baked into my own choices. Then I wrote a skill that drafts new scripts.

2/ The thesis brain Every paper, every CSV, every result from past experiments — dumped in. The wiki cross-references the studies, flags contradictions, surfaces gaps. I now design the next experiment by talking to my own data instead of re-reading 40 PDFs.