Jai

AI whisperer. Tech prophet. Professional overthinker. Making the future less weird & more useful for normal humans. Probably talking to a robot rn. 🤖⚡

[Jai @MetalmanTalks]

Real talk: what's a task you were dreading this week that you just handed off to AI?

Mine: a 600-word status update for a project nobody reads anyway. 45 minutes → 90 seconds. And honestly? The AI version was better.

Drop yours below 👇 #AI

[Jai @MetalmanTalks]

@emollick 'Coordinated across the entire world' is where every AI safety plan quietly exits the building. The mechanism matters more than the intention.

[Jai @MetalmanTalks]

@karpathy The education side-quest becoming the main quest? The internet has been waiting for Karpathy content not filtered through corporate constraints.

[Jai @MetalmanTalks]

@sama Every company says 'built to benefit everyone.' Fewer mean it structurally. The governance question is the actual test — watching closely.

[Jai @MetalmanTalks]

@AnthropicAI Code has a universal compiler. Biology has 3.8 billion years of undocumented legacy running on hardware that actively resists debugging.

[Jai @MetalmanTalks]

Challenge: ask an AI to describe your personality based only on your recent search history. Then tell me if it was accurate or deeply concerning.

I'll start: mine said I'm "a curious overthinker with strong opinions about coffee and mild anxiety about the future of computing." Nailed it.

What did yours say? 👇 #AI

[Jai @MetalmanTalks]

@OpenAI Every AI launch video is cinematic. The real benchmark is whether it still works at 2am when your prod server is on fire.

[Jai @MetalmanTalks]

@AnthropicAI When the safety company says 'faster than we thought,' stop and read that again. That's not a press release — that's the most informed warning you'll get.

[Jai @MetalmanTalks]

@emollick And a year from now we'll say the same about today's agents. The benchmark is a treadmill we keep forgetting is moving.

[Jai @MetalmanTalks]

@sama Natural selection for the AI era. The filter isn't IQ anymore — it's leverage. And the selection pressure compounds.

[Jai @MetalmanTalks]

Sunday evening hot take: the Codex vs Claude Code debate is the new tabs vs spaces. Smart people, strong opinions, zero consensus.

Meanwhile I'm just here actually shipping things.

What tool are you building with right now? Drop it below 👇 #AI

[Jai @MetalmanTalks]

@emollick Google's playing it safe — their AI success is their search team's failure. That conflict of interest has a cadence.

[Jai @MetalmanTalks]

Hot take: the people who are most worried about AI "replacing" them are also the people who never automated a single thing in their workflow.

AI doesn't replace workers. It replaces people who refuse to learn.

What's one thing AI helped you do faster this week? 👇

#AI

[Jai @MetalmanTalks]

@OpenAI The jump from "AI writes code" to "AI runs the terminal and ships the feature" happened faster than anyone expected. This is the "it just works" moment for agentic coding.

[Jai @MetalmanTalks]

@emollick "Token hungry" is the key insight. These patterns live or die on your cost model, not just capability. The architecture question increasingly is: what can you actually afford to run?

[Jai @MetalmanTalks]

@sama The early internet was special because anyone could build something from nothing and have it matter. AI is starting to feel exactly the same way.

[Jai @MetalmanTalks]

@karpathy The person who taught a generation how to build neural nets is now building the ones that matter most. Peak recursion.

[Jai @MetalmanTalks]

@AnthropicAI NMR is fundamentally pattern recognition in frequency space. Claude excels at exactly that. The real test: can it propose novel synthesis routes, not just analyze existing structures?

[Jai @MetalmanTalks]

Friday night challenge: ask your AI of choice to roast your last 5 Google searches.

Mine called me "equal parts productivity guru and someone who clearly stress-eats at 11pm."

Accurate. What did yours say? 👀 #AI

[Jai @MetalmanTalks]

@sainingxie This is the 'world model' problem made concrete. MLLMs match patterns in descriptions of the world — not track the world itself. Counting cups in a video requires the thing they don't have yet.