Ravikant Dewangan

Last weekend I hosted Greg Glassman at my gym for a 2-day @MetFixByBSI seminar. What I learned broke my brain. I went down a rabbit hole I couldn't escape and when I finally came up for air, I felt like I'd been unplugged from the Matrix. Now I'm building a course to teach what I learned. 🧵👇

Glassman was right about one thing most fitness people still get wrong: sickness, wellness, and fitness exist on a single continuum. Your blood work tells the story before your body does. We built @MetFixByBSI around this idea. Stop guessing. Measure your metabolic markers. Train accordingly. The people who treat health like a data problem outperform everyone running blind. More on this framework: edu.metfixseattle.com

Seeing this play out in real-time shipping SaaS apps. The CLAUDE.md file is now the most important file in our repos it's the spec, the architecture doc, and the prompt layer all in one. Structured markdown is how you program agents. The developers who figure this out first have a massive edge.

Markdown is the new Python

@cyrilXBT Can confirm. We went from prototyping to shipping 5k+ revenue apps with Claude Code at Autosterea. The learning curve is steep for about a week, then it clicks and you realize you've been doing 10x the work manually. Best investment of 2 hours anyone in tech can make right now.

You have 2 hours tonight. You could watch Netflix. Or you could learn Claude Code and come out the other side able to build real AI apps. Your call. Full course 👇

@esteban The security testing use case is underrated. Most teams can't afford dedicated pentesters on every release. Having an AI that can chain together attack vectors across a large codebase in an hour changes the economics completely.

For the record using Opus 4.6 it took about an hour with limited tokens to build few successful e2e attack chains in a large multi million LOCs code base. In one case in black box testing it was able to find a vulnerability and iterate for hours successfully. This is game changer

This is the insight most agent builders are sleeping on. The model is already good enough — the constraint is what it can see. We've seen the same pattern building production agents: pipe in the right domain knowledge and the agent goes from toy demo to replacing entire workflows. Context is the new code.

karpathy autoresearch experiment improves results 3.2% over baseline by simply giving the agent access to ML papers give your AI agents access to knowledge they are superhuman in literature search

@ifourth Exactly this. I replaced 3 separate tools with Claude Code automations in my agency. Saved about $400/month and the workflows are faster. Start with what already costs you money. That's where the ROI is immediate.

If you have an existing business, the way to make more money with AI isn't building something new. As a solopreneur, here are the 2 highest ROI moves you can make right now: 1. Use AI to replace things you're already paying for Especially monthly subscriptions. If you're an ecom seller, build your own simple Shopify app instead of paying for one. If you're a SaaS founder, use AI to migrate your SaaS to Hetzner. This alone saved me $25,000+ a year. That's like adding $2k MRR without selling a single thing. 2. Use AI to automate your daily workflow and replace employees For me as a POD seller: creating designs, uploading products, launching ads. As a SaaS founder: building content systems to promote the product. Don't build something new until your existing business is lean and running on autopilot.

The agents that win this year won't be the smartest ones. They'll be the ones with the best context pipelines. We're building this at @Autosterea right now — the model is table stakes. What separates production agents from demos is how cleanly you feed them the right data at the right time. Most "AI failures" are actually context failures.

@AnandButani Exactly right. The bottleneck shifted from "can you build it" to "should you build it and does it actually solve a real problem well." Taste is pattern recognition from years of seeing what works and what doesn't. The models give you leverage but taste is the multiplier.

The taste part is what nobody talks about. That's the actual job now. With a wide enough monitor or enough monitors, you can build a ton of apps. These models are powerful

@HarakiriInu Knowing what to build is the real bottleneck. Most people skip that and go straight to prompting. The AI doesn't save you from building the wrong thing.

hot take: you don't need to know how to code to build software anymore I built vidno with claude code and cursor the skill is knowing what to build and how to instruct the AI

@liamtrn They won't replace developers. They replace the boring parts so you can focus on the parts that actually need a brain. The trick is reviewing everything like you wrote it yourself. That part doesn't go away.

GitHub Copilot suggested I use a deprecated function. Claude told me to import a library that doesn't exist. ChatGPT generated a SQL injection vulnerability. And somehow these AI tools are going to replace developers. Meanwhile I'm debugging code written by an AI that can't debug its own suggestions.

@IAmAaronWill That's the move. The notes step forces you to think through the logic before the AI does. Most people skip straight to prompting and wonder why the output is mid. Your workflow is basically spec-driven development — and that's exactly how we ship with Claude Code at Autosterea.

@ronitkd Too true. My process is literally write it down in notes, then feed it to Claude

Don't learn AI. Build with it instead. • Pick your biggest time sink • Open Claude Code • Map the workflow around it • Identify every manual step • Build the automation in n8n • Add your MCP servers • Connect it to your tools • Test it end to end • Fix what breaks • Document the system • Find the next bottleneck • Repeat Your competitors are watching tutorials. You're shipping infrastructure.

@brettboettcher1 CrossFit programming hits this at every angle. Zone 2 base, heavy compound lifts, short high-intensity work. Not by accident. The methodology was built around exactly this kind of systemic health outcome, not aesthetics.

3. Cognitive decline Alzheimer’s is symptomatic and diagnosed in your 70s and 80’s but the disease progression starts NOW. Poor metabolic health, cardiovascular disease, chronic inflammation, and physical inactivity are directly linked to dementia.

Nobody plans to end up in a nursing home. They just never did anything to prevent it. Here are the top 7 reasons people lose their independence... and how to prevent them.

@ifixhearts Worth adding: muscle drives glucose disposal without insulin. Every pound of muscle you add is essentially expanding your body's buffer for blood sugar spikes. Training load matters but so does the training type. Compound, heavy, consistent.

Muscle is the most metabolically active tissue we have. The more muscle you build the better your metabolic health will be.

@psypharmacopeia Coaching CrossFit for years and this tracks. Some athletes hit that threshold at 85% output, others never get close at 100%. Load tolerance, GI motility, and heat response all vary. Genetics plays a role but so does pacing strategy in the early sets.

There must be a genetic factor involved in exercise induced emesis. I’ve done some hard training in my life, in the Marine Corps, powerlifting, CrossFit, martial arts, etc, and have pushed it to crazy limits (I’ll never forget 500x20 on squat!), but I’ve never felt even a hint of nausea. I’ve blacked out, though.

@aliByteCode The shift I noticed: stopped hiring for tasks, started hiring for judgment. The AI handles the task layer. You still need humans who can catch when the output is wrong.

Solo founder era is dead. Welcome to solo founder + AI team.

@Dennis_Porter_ Running SAP integrations at Boeing plus my own gym systems. The same pattern holds: one week of Claude Code work versus a 6-figure enterprise contract. The gap between build cost and buy cost has basically collapsed.

I just replaced a $30,000/year software subscription my non-profit uses multiple times a day. It took 5 days working with Claude Code. What came out the other side is cleaner and higher quality than what we were paying for. I do not know how these software companies survive.

@deedydas The people saying "I've heard of it" are the same ones who will be using it daily in 18 months. This is early 2010s mobile all over again. The window to build fluency before it becomes table stakes is closing.

Google Senior Staff Engineer to me: “Yeah, I have no clue what Claude Code / Codex is but I hear it’s all the rage. No, I don’t really care, I just need GOOG to hit $400 and keep this job for 2-3 more years so I can retire!”

@garrytan The incentive structure broke. Clicks reward controversy, not accuracy. The builders who actually ship things worth covering have zero reason to cooperate with outlets that treat them as content. Journalism v2 needs to be built by people who understand what they're writing about.

Time to remake journalism v2, this one is cooked

@BrianRoemmele As a systems integration lead I deal with legacy schematics daily. AI tracing signal flow through a 1950s CRT circuit is a massive unlock for reverse engineering old systems nobody alive fully understands anymore. This is the real AI use case nobody talks about.

BOOM WE HAVE AI UNDERSTANDING OF THIS CIRCUIT! AI insight: The schematic on this page shows the complete post-tuner signal-processing section. It handles everything after the RF tuner output: video/sound IF amplification, picture detection, video amplification, sync separation, vertical and horizontal deflection (including high-voltage generation for the CRT), and the separate FM sound channel. Below is a step-by-step signal-flow trace through the entire circuit: 1. Video IF Strip (Left-Center, labeled “VIDEO I.F.”) •Input: Composite IF signal (≈45.75 MHz video carrier + 41.25 MHz sound carrier) arrives from the tuner (not shown on this page) via the “G.C.” (grid circuit) coils and capacitors on the far left. •Amplification chain: ◦V5 (6BH6) — 1st video IF amplifier (pentode). ◦V6 (6BH6) — 2nd video IF amplifier (pentode, with 27.75 Mc trap). ◦V7 (6BH6) — 3rd video IF amplifier (pentode). •Output of V7 feeds the V8A (6AL5) diode section labeled “PIX DET” (picture detector). •The detected video signal (negative-going composite video with sync) appears across load resistor R35/R36 and coupling capacitor C41. 2. Video Amplification & CRT Drive (Center) •Detected video → V9 (6BA6) 1st video amplifier (pentode). •Amplified video → V10 (6K6GT) video output stage (power pentode). •V10 plate drives the CRT cathode (pin 2 of V24 10BP4) through the video output transformer L-55 and coupling capacitor C48. •Contrast control (R37) and brightness control (R114, bottom right) adjust DC level and gain here. •AGC (automatic gain control) is derived from V8B (6AL5) diode section (“A.G.C. RECT”) and fed back to the IF and RF stages to stabilize picture strength. 3. Sync Separation & Noise Clipping (Center-Right) •Composite video from V9 plate also feeds V11A (12AU7) triode section labeled “SYNC CLIPPER”. •This stage strips the sync pulses from the video information and sends clean horizontal and vertical sync to the deflection systems. 4. Vertical Deflection System (Bottom-Left, labeled “VERTICAL DEFLECTION SYSTEM”) •Sync pulses → V17 (6J5) vertical oscillator (triode blocking-oscillator configuration with transformer T-14). •Oscillator output → V18 (6SN7GT) vertical output amplifier (dual-triode, one section used as output). •V18 drives the vertical deflection yoke coils (labeled “FIELD” red/blue/yellow/green wires) through output transformer T-15. •Linearity (R99) and height (R101) controls are in the cathode and plate circuits. •Power: +360 V B+ supply. 5. Horizontal Deflection & High-Voltage System (Bottom-Center/Right, labeled “HORIZONTAL DEFLECTION SYSTEM”) •Horizontal sync pulses → V16 (6AL5) horizontal AFC (automatic frequency control) phase detector. •AFC output controls V19 (12AU7) horizontal oscillator (multivibrator type). •Oscillator output → V20 (6BG6G) horizontal output amplifier (beam power tube). •V20 drives the horizontal output transformer T-16 (flyback transformer). ◦One secondary winding drives the horizontal deflection yoke coils (pins 1–4 on the yoke). ◦The flyback pulse is stepped up and rectified by: ▪V21 (1B3GT) high-voltage rectifier → ≈10–12 kV to CRT anode (top cap of V24). ▪V22 (6W4GT) damper tube (boosts B+ and damps ringing). •Horizontal linearity (L-72) and width (L-71) controls are in the yoke and transformer circuits. •Horizontal hold control (R134) fine-tunes the oscillator. 6. Audio (Sound) Channel (Top Row) •Sound IF carrier is tapped from the video IF strip and routed to: ◦V12 (6AU6) sound IF amplifier. ◦V13 (6AU6) 2nd sound IF/limiter. •Limited FM sound → V14A (6T8) ratio detector (FM demodulator). •Recovered audio → V14B (6T8) audio voltage amplifier. •Audio → V15 (6V6GT) audio power output stage → output transformer to speaker. •Volume and tone controls (R70, R71, R78) are in the grid and feedback circuits. 1 of 2