VIAINTI

X is paying creators again today. I’m currently averaging 15K–30K impressions per day. My next goal? 100K impressions daily. I post, experiment, and learn what works every single day. Now I’m curious: How many daily impressions are you getting, and how much is X paying you? Let’s compare real numbers.

Timeline this week (per the timeline, not my crystal ball): China AI bubble talk. Paraguay + banks in the news. Trump / crypto companies under the microscope. "Months of red" is a vibe, not a thesis. Builders: are you preparing for a downcycle, or still shipping like it's 2021?

Your AI product can only have ONE of these forever. Pick: A) Perfect accuracy, 8s latency B) 70% accuracy, 200ms latency C) Perfect accuracy, costs 10x per request D) 200ms + cheap, hallucinates once a week Reply with a letter and why.

AI startup pitch deck in 2026: Slide 1: TAM = everything Slide 2: moat = "prompts" Slide 3: GTM = "viral demo" Slide 4: unit economics = [redacted] Slide 5: "we're not a wrapper" What's missing?

Honest question for people shipping agents: When your agent fails in production, what tells you first? A) A user complains B) An alert you set up C) Your eval suite D) You find out on Twitter If the answer is A or D, you're not alone.

Wild week of takes: - China AI bubble "popped" - Paraguay banking headlines - Trump / crypto company investigations supposedly next I'm not predicting prices. I'm asking builders: If the next few months get ugly for crypto and AI funding, what do you cut first? A) Headcount B) Model spend C) Marketing D) Nothing — you build anyway

Kimi K3.1 Pro landed with day-0 support in vLLM (Nov 4). Per the vLLM blog: new frontier model, shippable the same day the weights drop. The gap that used to kill adoption — "cool model, zero infra" — is shrinking fast. What's your bottleneck now? A) Model quality B) Serving cost C) Tooling / harness around the model D) Eval coverage

New term going around: "cognitive surrender." It's when the AI diff looks clean, the tests pass, and you click approve without understanding a single line. Offloading is fine. Surrendering is how you end up owning code you can't debug at 2am. Where's your line between the two?

Finish this: "I trust AI with ___, but never with ___." Best answers get a repost.

Klavis dropped Multi-step SWE on Nov 4: 69 tasks, 574 dependent steps, nine frontier models. Per their results, the strongest setup (Claude Opus 5.5 / Claude Code) passes 56.8% of individual steps — but only solves 18.5% of complete multi-step tasks. Across all nine models, 38 of 69 tasks stay unsolved. Hot take: a green single step is not engineering. The hard problem is agents that have to keep building on their own code without rotting the workspace. Which failure do you hit more? A) Step N looks fine, step N+1 collapses B) Agent says "done" while checks still fail C) Works once, then becomes un-extendable

NVIDIA's Nov 4 NeMo walkthrough (Dassault Systèmes door-design agent): Per NVIDIA: after a trace → custom-eval → fix loop, malformed tool calls fell from 142 to 0. They report a ~28% relative improvement in overall eval score and 35% fewer tool calls per run across six scenarios — overnight (~12 hours), no human in that pass. Opinion: most "agent reliability" gains aren't a new frontier model. They're turning yesterday's failure traces into today's regression suite. Beginner vs expert: Beginner: "the model is dumb" Expert: "show me the recurring tool-call pattern across traces" Where are you stuck — model quality, tool schema, or missing evals for the failure you already saw twice?

GM ☀️ Thursday question: What are you building today? A) AI app B) Crypto project C) Side project nobody knows about yet D) Still deciding (no judgment)

Hot take: Most "AI agents" are just a cron job with a chatbot attached. Agree or disagree?

Honest crypto question: If gas fees were zero and every wallet was easy to use, would normal people actually use crypto? Or is the problem something else?

"AI will kill SaaS." Also everyone: - pays for 14 AI subscriptions - logs into 6 AI dashboards - gets billed monthly for all of it AI didn't kill SaaS. It became SaaS.

You can only use one for a month: A) An AI that writes code B) An AI that reviews code C) An AI that writes tests Pick one. Explain why.