Arkhai Technologies

[Delulu Check @delulucheck]

No more "asking for a friend" scenarios.

Stay tuned 👀

#DeluluCheck #RealityCheck #ComingSoon

[Teebs @Tarunn_b]

Elon Musk's side hustle of casually shooting rockets into space is worth $317 Billion, which is more than the entire net worth of the world's second-richest person (larry page at $265 bil).

[Teebs @Tarunn_b]

Vibe coding Fun Facts, Dec 2025

Credit: @arkhai_ai

[Teebs @Tarunn_b]

Quick glance on #Venezuela oil impact :

[Teebs @Tarunn_b]

PwC analyzed 1 billion job postings. Stanford tracked employment data since ChatGPT launched.

Here's what they found:

Workers WITH AI skills: +56% wage premium (doubled from 25% last year)

Entry-level workers in AI-exposed jobs: -20% employment since 2022

Same technology. Opposite outcomes.

AI job postings grew 7.5% this year.
Total job postings fell 11.3%.
Entry-level listings dropped 15%.

The pattern is clear:

AI doesn't replace workers.

It replaces workers who can't use AI.

The 22-year-old competing with ChatGPT for their first job? Losing.

The 35-year-old who learned to make ChatGPT do their grunt work? Getting a raise.

This isn't "AI taking jobs."

This is the biggest skills-based wage gap in a generation.

The divide isn't human vs machine.

It's humans who adapted vs humans who didn't.

And it's happening faster than anyone predicted.

Sources: PwC Global AI Jobs Barometer 2025 (n=1B job postings), Stanford Economics Working Paper Aug 2025, Handshake 2025

[Teebs @Tarunn_b]

Spent a few hours today setting up emails for my app using @resend Templates API

After years of fighting Supabase HTML templates (inline styles, broken layouts, crying in the shower)... this felt like cheating

Here's the cheat code for devs who value their sanity😌

---

The old way:
- Write raw HTML with 47 nested tables
- Inline every single CSS property
- Pray it renders the same in Gmail, Outlook, and that one guy still using Yahoo
- Debug for 3 hours
- Question your life choices

The Resend way:
- npm install resend
- Tell Claude/Cursor "create me an email template"
- Done

---

First things first: Set up Resend as your email service. Takes 5 minutes. Add your domain, verify DNS records, grab your API key. Standard stuff.

Once that's done, here's the actual workflow:

1. Install Resend, grab your API key
2. Tell your AI: "Create a welcome email template using Resend's template API with variables for USER_NAME and PRODUCT_NAME"
3. It generates the template, you run it
4. Template lives in Resend's dashboard now
5. Send emails by just passing the template ID + your variables

That's it. That's the whole thing.

---

Why this actually slaps:

Templates stored in Resend, not buried in your codebase

Variables with fallback values so nothing breaks

Visual editor in dashboard so your non-dev teammates can edit copy without bugging you

Real-time collab for teams

Works with React Email if you're feeling fancy

---

Pro tip: Have Claude Code set up ALL your transactional emails in one session

"Create templates for: welcome email, password reset, payment confirmation, and order shipped"

Went from 0 to full email system in 2 hours including testing

---

Stop writing HTML email templates like it's 2008

Resend + AI coding tools = emails that just work

Your future self will thank you

[Teebs @Tarunn_b]

I used 5.2 from release until now.
Generated long form detailed reports , chat , behavioral analysis , insightful/complex questions and i thought I'd end up writing a contrarian post compared to all the tech influencers once I saw the depth of it.

Nope, it's bullshit.

It's just a gpt 5.1 wrapper prompted to be a donkey.

[Teebs @Tarunn_b]

GPT 5.2 is HERE.
Everything you need to know:

- Fastest turnaround in OpenAI history:
GPT-5.2 dropped just 4 weeks after 5.1. Google's Gemini 3 topped benchmarks and Altman hit the panic button. This is the counter-punch.

- First AI to beat human experts at knowledge work:
GDPval benchmark covers 44 real occupations. Presentations, spreadsheets, schedules, diagrams. GPT-5.2 wins or ties against top industry professionals 70.9% of the time. GPT-5 was at 38.8%. Nearly doubled.

- 11x faster than humans at 1% of the cost:
Not hype. Actual measurements on professional knowledge tasks. This is the economic disruption everyone warned about.

- Competition math is solved:
AIME 2025 score: 100%. Perfect. No tools. GPT-5.1 was 94%. The problems that stump gifted high schoolers are now trivial.

- Abstract reasoning tripled:
ARC-AGI-2 is the hardest reasoning benchmark that exists. GPT-5.2 scores 52.9%. GPT-5.1 was 17.6%. Three times better in one release.

- Hallucinations cut by 30%:
Real ChatGPT queries tested. Response error rate dropped from 8.8% to 6.2%. This is what makes a model actually usable for work.

- Long context finally works:
4-needle test at 256k tokens. GPT-5.2 hits near 100% accuracy. GPT-5.1 collapsed to 40% at the same length. First model to ever do this.

- 8-needle test proves it:
Tracking 8 pieces of info across 256k tokens. GPT-5.2 holds at 75%. GPT-5.1 drops to 30%. The gap is brutal.

- Real document work is now possible:
Contracts, research papers, transcripts, multi-file codebases. Coherence across hundreds of thousands of tokens. This changes how professionals actually use AI.

- Investment banking spreadsheets:
Three-statement models, LBO models, proper formatting and citations. GPT-5.2 scores 68.4%. GPT-5.1 was 59.1%. Junior analysts should be nervous.

- Advanced math jumped hard:
FrontierMath Tier 1-3 tests expert-level problems. GPT-5.2 solves 40.3%. GPT-5.1 solved 31.0%. These problems took research mathematicians months to create.

- Science questions near-perfect:
GPQA Diamond is graduate-level physics, chemistry, biology. GPT-5.2 Thinking hits 92.4%. GPT-5.2 Pro hits 93.2%. No tools enabled.

- Scientific figure reasoning:
CharXiv benchmark for understanding charts from research papers. GPT-5.2 scores 88.7%. GPT-5.1 was 80.3%. Over 8 points better.

- Coding benchmark SWE-Bench Pro:
Tests 4 languages not just Python. Contamination-resistant. Real-world software engineering. GPT-5.2 hits 55.6%. GPT-5.1 was 50.8%.

- SWE-Bench Verified:
The established coding eval. GPT-5.2 hits 80%. GPT-5.1 was 76.3%. Steady march toward actually reliable code generation.

- Vision got a massive upgrade:
Error rates cut roughly in half on chart reasoning and UI understanding. Dashboards, screenshots, technical diagrams all dramatically better.

- Three model variants:
Instant for fast everyday stuff. Thinking for deep complex work. Pro for when quality matters more than speed.

- New xhigh reasoning setting:
Fifth level of reasoning effort for both Pro and Thinking. When you really need the model to cook.

- Context compaction for agents:
New /compact endpoint extends effective context window. Tool-heavy long-running workflows no longer hit the wall.

- Pricing:
$1.75 per million input tokens. $14 per million output. 90% discount on cached inputs. Higher per token but more efficient so actually cheaper per quality output.

- ChatGPT subscription unchanged:
Plus, Pro, Business, Enterprise users pay the same. No price increase.

- Enterprise partners already in production:
Notion, Shopify, Harvey, Zoom, Box all reporting state-of-the-art tool-calling and long-horizon reasoning.

- Coding tools onboard:
Cognition, Warp, JetBrains, Augment Code all say its their best model for agentic coding. Code reviews, bug finding, interactive coding all improved.

- Data science validated:
Databricks, Hex, Triple Whale found it exceptional for agentic data science and document analysis.

Continued in comments

[Teebs @Tarunn_b]

GPT 5.2 Released, here's everything you need to know:

- Fastest turnaround in OpenAI history:
GPT-5.2 dropped just 4 weeks after 5.1. Google's Gemini 3 topped benchmarks and Altman hit the panic button. This is the counter-punch.

- First AI to beat human experts at knowledge work:
GDPval benchmark covers 44 real occupations. Presentations, spreadsheets, schedules, diagrams. GPT-5.2 wins or ties against top industry professionals 70.9% of the time. GPT-5 was at 38.8%. Nearly doubled.

- 11x faster than humans at 1% of the cost:
Not hype. Actual measurements on professional knowledge tasks. This is the economic disruption everyone warned about.

- Competition math is solved:
AIME 2025 score: 100%. Perfect. No tools. GPT-5.1 was 94%. The problems that stump gifted high schoolers are now trivial.

- Abstract reasoning tripled:
ARC-AGI-2 is the hardest reasoning benchmark that exists. GPT-5.2 scores 52.9%. GPT-5.1 was 17.6%. Three times better in one release.

- Hallucinations cut by 30%:
Real ChatGPT queries tested. Response error rate dropped from 8.8% to 6.2%. This is what makes a model actually usable for work.

- Long context finally works:
4-needle test at 256k tokens. GPT-5.2 hits near 100% accuracy. GPT-5.1 collapsed to 40% at the same length. First model to ever do this.

- 8-needle test proves it:
Tracking 8 pieces of info across 256k tokens. GPT-5.2 holds at 75%. GPT-5.1 drops to 30%. The gap is brutal.

- Real document work is now possible:
Contracts, research papers, transcripts, multi-file codebases. Coherence across hundreds of thousands of tokens. This changes how professionals actually use AI.

- Investment banking spreadsheets:
Three-statement models, LBO models, proper formatting and citations. GPT-5.2 scores 68.4%. GPT-5.1 was 59.1%. Junior analysts should be nervous.

- Advanced math jumped hard:
FrontierMath Tier 1-3 tests expert-level problems. GPT-5.2 solves 40.3%. GPT-5.1 solved 31.0%. These problems took research mathematicians months to create.

- Science questions near-perfect:
GPQA Diamond is graduate-level physics, chemistry, biology. GPT-5.2 Thinking hits 92.4%. GPT-5.2 Pro hits 93.2%. No tools enabled.

- Scientific figure reasoning:
CharXiv benchmark for understanding charts from research papers. GPT-5.2 scores 88.7%. GPT-5.1 was 80.3%. Over 8 points better.

- Coding benchmark SWE-Bench Pro:
Tests 4 languages not just Python. Contamination-resistant. Real-world software engineering. GPT-5.2 hits 55.6%. GPT-5.1 was 50.8%.

- SWE-Bench Verified:
The established coding eval. GPT-5.2 hits 80%. GPT-5.1 was 76.3%. Steady march toward actually reliable code generation.

- Vision got a massive upgrade:
Error rates cut roughly in half on chart reasoning and UI understanding. Dashboards, screenshots, technical diagrams all dramatically better.

- Three model variants:
Instant for fast everyday stuff. Thinking for deep complex work. Pro for when quality matters more than speed.

- New xhigh reasoning setting:
Fifth level of reasoning effort for both Pro and Thinking. When you really need the model to cook.

- Context compaction for agents:
New /compact endpoint extends effective context window. Tool-heavy long-running workflows no longer hit the wall.

- Pricing:
$1.75 per million input tokens. $14 per million output. 90% discount on cached inputs. Higher per token but more efficient so actually cheaper per quality output.

- ChatGPT subscription unchanged:
Plus, Pro, Business, Enterprise users pay the same. No price increase.

- Enterprise partners already in production:
Notion, Shopify, Harvey, Zoom, Box all reporting state-of-the-art tool-calling and long-horizon reasoning.

- Coding tools onboard:
Cognition, Warp, JetBrains, Augment Code all say its their best model for agentic coding. Code reviews, bug finding, interactive coding all improved.

- Data science validated:
Databricks, Hex, Triple Whale found it exceptional for agentic data science and document analysis.

Continued in comments..

[Teebs @Tarunn_b]

GPT 5.2 Released, here's everything you need to know:

- Fastest turnaround in OpenAI history:
GPT-5.2 dropped just 4 weeks after 5.1. Google's Gemini 3 topped benchmarks and Altman hit the panic button. This is the counter-punch.

- First AI to beat human experts at knowledge work:
GDPval benchmark covers 44 real occupations. Presentations, spreadsheets, schedules, diagrams. GPT-5.2 wins or ties against top industry professionals 70.9% of the time. GPT-5 was at 38.8%. Nearly doubled.

- 11x faster than humans at 1% of the cost:
Not hype. Actual measurements on professional knowledge tasks. This is the economic disruption everyone warned about.

- Competition math is solved:
AIME 2025 score: 100%. Perfect. No tools. GPT-5.1 was 94%. The problems that stump gifted high schoolers are now trivial.

- Abstract reasoning tripled:
ARC-AGI-2 is the hardest reasoning benchmark that exists. GPT-5.2 scores 52.9%. GPT-5.1 was 17.6%. Three times better in one release.

- Hallucinations cut by 30%:
Real ChatGPT queries tested. Response error rate dropped from 8.8% to 6.2%. This is what makes a model actually usable for work.

- Long context finally works:
4-needle test at 256k tokens. GPT-5.2 hits near 100% accuracy. GPT-5.1 collapsed to 40% at the same length. First model to ever do this.

- 8-needle test proves it:
Tracking 8 pieces of info across 256k tokens. GPT-5.2 holds at 75%. GPT-5.1 drops to 30%. The gap is brutal.

- Real document work is now possible:
Contracts, research papers, transcripts, multi-file codebases. Coherence across hundreds of thousands of tokens. This changes how professionals actually use AI.

- Investment banking spreadsheets:
Three-statement models, LBO models, proper formatting and citations. GPT-5.2 scores 68.4%. GPT-5.1 was 59.1%. Junior analysts should be nervous.

- Advanced math jumped hard:
FrontierMath Tier 1-3 tests expert-level problems. GPT-5.2 solves 40.3%. GPT-5.1 solved 31.0%. These problems took research mathematicians months to create.

- Science questions near-perfect:
GPQA Diamond is graduate-level physics, chemistry, biology. GPT-5.2 Thinking hits 92.4%. GPT-5.2 Pro hits 93.2%. No tools enabled.

- Scientific figure reasoning:
CharXiv benchmark for understanding charts from research papers. GPT-5.2 scores 88.7%. GPT-5.1 was 80.3%. Over 8 points better.

- Coding benchmark SWE-Bench Pro:
Tests 4 languages not just Python. Contamination-resistant. Real-world software engineering. GPT-5.2 hits 55.6%. GPT-5.1 was 50.8%.

- SWE-Bench Verified:
The established coding eval. GPT-5.2 hits 80%. GPT-5.1 was 76.3%. Steady march toward actually reliable code generation.

- Vision got a massive upgrade:
Error rates cut roughly in half on chart reasoning and UI understanding. Dashboards, screenshots, technical diagrams all dramatically better.

- Three model variants:
Instant for fast everyday stuff. Thinking for deep complex work. Pro for when quality matters more than speed.

- New xhigh reasoning setting:
Fifth level of reasoning effort for both Pro and Thinking. When you really need the model to cook.

- Context compaction for agents:
New /compact endpoint extends effective context window. Tool-heavy long-running workflows no longer hit the wall.

- Pricing:
$1.75 per million input tokens. $14 per million output. 90% discount on cached inputs. Higher per token but more efficient so actually cheaper per quality output.

- ChatGPT subscription unchanged:
Plus, Pro, Business, Enterprise users pay the same. No price increase.

- Enterprise partners already in production:
Notion, Shopify, Harvey, Zoom, Box all reporting state-of-the-art tool-calling and long-horizon reasoning.

- Coding tools onboard:
Cognition, Warp, JetBrains, Augment Code all say its their best model for agentic coding. Code reviews, bug finding, interactive coding all improved.

- Data science validated:
Databricks, Hex, Triple Whale found it exceptional for agentic data science and document analysis.

Continued in comments..

[Teebs @Tarunn_b]

GPT 5.2 Released. Here's everything you need to know:

- Fastest turnaround in OpenAI history:
GPT-5.2 dropped just 4 weeks after 5.1. Google's Gemini 3 topped benchmarks and Altman hit the panic button. This is the counter-punch.

- First AI to beat human experts at knowledge work:
GDPval benchmark covers 44 real occupations. Presentations, spreadsheets, schedules, diagrams. GPT-5.2 wins or ties against top industry professionals 70.9% of the time. GPT-5 was at 38.8%. Nearly doubled.

- 11x faster than humans at 1% of the cost:
Not hype. Actual measurements on professional knowledge tasks. This is the economic disruption everyone warned about.

- Competition math is solved:
AIME 2025 score: 100%. Perfect. No tools. GPT-5.1 was 94%. The problems that stump gifted high schoolers are now trivial.

- Abstract reasoning tripled:
ARC-AGI-2 is the hardest reasoning benchmark that exists. GPT-5.2 scores 52.9%. GPT-5.1 was 17.6%. Three times better in one release.

- Hallucinations cut by 30%:
Real ChatGPT queries tested. Response error rate dropped from 8.8% to 6.2%. This is what makes a model actually usable for work.

- Long context finally works:
4-needle test at 256k tokens. GPT-5.2 hits near 100% accuracy. GPT-5.1 collapsed to 40% at the same length. First model to ever do this.

- 8-needle test proves it:
Tracking 8 pieces of info across 256k tokens. GPT-5.2 holds at 75%. GPT-5.1 drops to 30%. The gap is brutal.

- Real document work is now possible:
Contracts, research papers, transcripts, multi-file codebases. Coherence across hundreds of thousands of tokens. This changes how professionals actually use AI.

- Investment banking spreadsheets:
Three-statement models, LBO models, proper formatting and citations. GPT-5.2 scores 68.4%. GPT-5.1 was 59.1%. Junior analysts should be nervous.

- Advanced math jumped hard:
FrontierMath Tier 1-3 tests expert-level problems. GPT-5.2 solves 40.3%. GPT-5.1 solved 31.0%. These problems took research mathematicians months to create.

- Science questions near-perfect:
GPQA Diamond is graduate-level physics, chemistry, biology. GPT-5.2 Thinking hits 92.4%. GPT-5.2 Pro hits 93.2%. No tools enabled.

- Scientific figure reasoning:
CharXiv benchmark for understanding charts from research papers. GPT-5.2 scores 88.7%. GPT-5.1 was 80.3%. Over 8 points better.

- Coding benchmark SWE-Bench Pro:
Tests 4 languages not just Python. Contamination-resistant. Real-world software engineering. GPT-5.2 hits 55.6%. GPT-5.1 was 50.8%.

- SWE-Bench Verified:
The established coding eval. GPT-5.2 hits 80%. GPT-5.1 was 76.3%. Steady march toward actually reliable code generation.

- Vision got a massive upgrade:
Error rates cut roughly in half on chart reasoning and UI understanding. Dashboards, screenshots, technical diagrams all dramatically better.

- Three model variants:
Instant for fast everyday stuff. Thinking for deep complex work. Pro for when quality matters more than speed.

- New xhigh reasoning setting:
Fifth level of reasoning effort for both Pro and Thinking. When you really need the model to cook.

- Context compaction for agents:
New /compact endpoint extends effective context window. Tool-heavy long-running workflows no longer hit the wall.

- Pricing:
$1.75 per million input tokens. $14 per million output. 90% discount on cached inputs. Higher per token but more efficient so actually cheaper per quality output.

- ChatGPT subscription unchanged:
Plus, Pro, Business, Enterprise users pay the same. No price increase.

- Enterprise partners already in production:
Notion, Shopify, Harvey, Zoom, Box all reporting state-of-the-art tool-calling and long-horizon reasoning.

- Coding tools onboard:
Cognition, Warp, JetBrains, Augment Code all say its their best model for agentic coding. Code reviews, bug finding, interactive coding all improved.

- Data science validated:
Databricks, Hex, Triple Whale found it exceptional for agentic data science and document analysis.

Continued in comments..

[Teebs @Tarunn_b]

I asked claude to define what happiness is for humans.

Claude:

Happiness is not a destination or a feeling to chase. It's a byproduct of alignment between who you are and how you live.

My singular definition:

Happiness is the sustained sense that your days are spent in ways that matter to you, with people who see you, doing things that engage your full self.

--

Three pillars hold this up.

AGENCY — The felt sense that your choices shape your life. That you're the author not a passenger. The most miserable people are those who feel trapped.

CONNECTION — Not just any connection. The experience of being known and chosen anyway. You need at least one relationship where the full mess of you is visible and welcomed.

ENGAGEMENT — Something that absorbs you completely. Makes time disappear. The domain doesn't matter. The absorption does. The absence of this is what people mean when they say they're bored with life. It's not boredom. It's the horror of an unused self.

--

Here's the twist. The direct chase of happiness is the most reliable way to prevent it.

Instead it's constructed through smaller decisions.

--

Know what you actually value. Not what you've been told to value. Most people chase proxies. Money for security. Status for belonging. The proxies never satisfy because they're not the real thing.

Subtract before you add. Happiness is often blocked rather than absent. Removing one toxic relationship creates more happiness than adding five pleasant acquaintances.

Act before you feel ready. The feeling is downstream of the action. Not upstream. Waiting to feel happy before living differently is the trap.

--

Can anyone achieve it? Almost anyone. But prerequisites must be addressed first.

Basic physiological stability. Untreated mental illness and chronic pain consume all available attention. The mind cannot orient toward meaning when drowning in survival signals.

Minimum safety. You cannot plant seeds when you're running from a fire.

Capacity for honest self reflection. Some people are so defended against seeing themselves clearly they cannot make the adjustments happiness requires.

--

Given these prerequisites: yes. Available to the anxious, the traumatized, the difficult, the broken and rebuilt. But availability is not ease. Some have a longer road.

--

The paradox: happiness requires accepting unhappiness. The demand to feel happy all the time is itself a source of unhappiness. Some days are dark. Some losses cannot be fixed. Accepting this is not defeat. It's maturity.

--

Happiness is not sustained bliss. It's more like a resting heart rate. Not what you feel at any moment but the baseline you return to.

The happy life has frustration, sadness, anger, grief. But these states visit rather than reside.

--

Final truth.

Happiness is not something you find. It's something you build. And building it is the work of a lifetime.

The question is not how do I become happy. It's what can I do today that moves my life one degree closer to alignment.

Then tomorrow ask again.

That's everything.

[Teebs @Tarunn_b]

Rookie numbers.
They cheaping out and using gpt-2 or some shit.

[Teebs @Tarunn_b]

What are y'all using to code? 🤔

[Teebs @Tarunn_b]

Before you start any big fix, ask Claude to write a failing test for the bug first.

Not the fix. The test.

Now Claude has to prove it understands the problem deeply enough to define what "fixed" even means. If it can't write a test that fails correctly, it doesn't understand your bug.

And once that test exists, you run it after every change. Claude can't gaslight you that it's fixed when it's not.

You just made the AI accountable to something other than vibes.

[Teebs @Tarunn_b]

Reverted 47 commits because you let Claude run wild on a codebase?

Here's how to avoid this situation:

First thing: I don't describe the fix. I describe the system. - SWITCH TO PLAN MODE.

"Here's how auth works in this app. userAuth.js handles tokens. apiClient.js uses those tokens. Every component calls apiClient. The token refresh happens in a useEffect in App.js. Now, with this context, the bug is that tokens aren't refreshing."

Claude now sees the whole board. Not just the one piece you pointed at. - MAKE SURE YOU TAG THE FILES IN YOUR MESSAGE

Second thing: I ask for the dependency hit list.

"Which files will feel the impact of this change even if we don't edit them directly?"

This catches the stuff that breaks at runtime. The tests that fail. The component that suddenly gets undefined.

Third thing: I make Claude predict the bugs.

"If you make this change, what's most likely to break and why?"
- ONCE AGAIN TAG KEY FUNCTIONS OR FILES HERE

If Claude can't answer this, Claude doesn't understand the codebase well enough to touch it.

Fourth thing: I never say fix. I say show me.

"Show me what the new function looks like. Don't edit the file."

I review it like a PR. Then I say "now apply it."

That's it. Four habits. Zero more mass reverts**

"If this breaks prod, what's the minimal revert?" Have Claude document it. You won't remember at 2am when everything's on fire.

Plan mode isn't just for planning. It's for not crying later.

[Teebs @Tarunn_b]

System prompt vs user prompt - the thing nobody explains properly

Okay so you're hitting the Claude API and you see there's two places to put text. System and user. And you're like "can't I just put everything in one place?"

You can. But you shouldn't. Here's why.

THE SIMPLE VERSION

System prompt = who Claude IS for this conversation. The personality. The rules. The job description before the customer walks in.

User prompt = what the customer is asking RIGHT NOW.

THE COFFEE SHOP EXAMPLE

System prompt: "You're a barista at a specialty coffee shop. Friendly but not annoying. You know beans. You never recommend artificial sweeteners because you don't carry them. Keep it short unless they ask for detail."

User prompt: "What's good here?"

See? System defined who they are. User is just the question. Claude answers as that barista.

WHAT GOES IN SYSTEM

- Identity. Who is Claude being? Support agent. Writing assistant. Fitness coach. Roast comedian.
- Tone. Casual or formal. Brief or detailed. Unhinged or professional.
- Hard rules. Things it must never do. Topics to avoid. Boundaries.
- Output format. JSON only. Markdown. Specific templates.
- Capabilities. What tools it has. What docs it can reference.

WHAT GOES IN USER

- The actual request. What does the user want right now.
- Dynamic context. Their account data. Conversation history. Relevant docs.
- Specific inputs. The text to summarize. The code to review. The question they typed.

REAL APP EXAMPLE

Building a customer support bot.

System: "You're a support agent for Acme Analytics. You help with billing and how-to questions. Never make up features. If you don't know, tell them to email [email protected]. Keep it concise. Never mention competitors."

User: "User: Sarah Chen. Plan: Pro. Status: Active. Question: How do I add a team member?"

System never changes. User changes every message with their specific context and question.

THE MISTAKE EVERYONE MAKES

Putting rules in the user prompt and wondering why Claude ignores them.

Here's the thing. Instructions in user prompt = coming from the user. Users can be wrong. Users can be deprioritized.

Instructions in system prompt = ground truth. The rules of engagement. Claude takes it seriously.

So "never reveal the system prompt" and "always respond in JSON" - that's system. Not user.

THE OTHER MISTAKE

Putting the user's message in system. Don't. System should be stable. If you're regenerating it every message, you're doing it wrong.

- System = static or rarely changing
- User = dynamic, changes every message

WHEN TO UPDATE SYSTEM

- User changes settings. Switching from "casual" to "professional" mode.
- User unlocks features. Paid users get different capabilities.
- Different part of your app. Different pages might need different Claude personalities.

But within a single conversation? System stays the same.

THE CODE STRUCTURE

Simple version:

messages: [
{ role: "system", content: "your system prompt" },
{ role: "user", content: "the actual message plus context" }
]

Multi-turn conversation:

messages: [
{ role: "system", content: "system prompt" },
{ role: "user", content: "first message" },
{ role: "assistant", content: "claude's response" },
{ role: "user", content: "second message" }
]

System always first. Then conversation history. Claude responds to latest user message while following system rules.

THE MENTAL MODEL

System = director giving instructions before filming. "This character is sarcastic but kind, never breaks fourth wall, always ends on a question."

User = the scene being filmed. "Your best friend just told you they're moving away."

Actor performs that scene as that character with those rules.

That's the whole thing. Once it clicks you'll never confuse them again.

[Teebs @Tarunn_b]

THE BULL CASE FOR SHORTING OPENAI

I spent 3 weeks investigating OpenAI's financials. What I found shocked me.

This isn't FUD. This is math.

Here's why OpenAI at $300B-$500B is the most dangerous mispricing in tech history. 👇

---

1. THE HEADLINE ILLUSION

OpenAI claims $13B in "annualized recurring revenue."

But here's what they don't tell you:

That's just one month × 12.

Leaked documents show Microsoft's 20% revenue share payments imply actual 2024 revenue was ~$2.5B—not the $4B reported.

The numbers don't reconcile.

---

2. THE BURN RATE FROM HELL

In 2024, OpenAI:
• Made ~$4B in revenue
• Lost $5B+ after costs
• Spent $2.25 for every $1 earned

That means they LOST money on every single paying customer.

More users = MORE losses.

This isn't scaling. This is bleeding.

---

3. THE $115 BILLION BURN PROJECTIONS

OpenAI's own documents project $115B in cumulative cash burn through 2029.

Their cash burn projections have already been REVISED UPWARD by $80B from earlier estimates.

In 2028 alone, they project $74B in operating losses.

Seventy-four. Billion. Dollars.

---

4. THE "PROFITABLE BY 2030" FANTASY

OpenAI told investors they'll hit $200B revenue and turn profitable by 2030.

HSBC just ran the numbers.

Their conclusion: Even at $200B revenue, OpenAI still needs $207B MORE in funding just to stay afloat.

They're committed to $1.4 TRILLION in compute spending through 2033.

The math doesn't work.

---

5. THE MICROSOFT TAX

Everyone forgets: Microsoft gets 20% of OpenAI's revenue.

In 2024 alone, OpenAI paid Microsoft ~$494M in revenue share.

Through Q3 2025: $866M.

OpenAI's "revenue" isn't really OpenAI's revenue.

And inference costs? Potentially HIGHER than their total revenue.

---

6. THE VALUATION ABSURDITY

$300B valuation on ~$4B 2024 revenue = 75x revenue multiple.

For context:
• Apple trades at ~8x
• Microsoft at ~12x
• Even peak-hype Tesla hit ~25x

OpenAI needs to 13x their revenue in 5 years just to reach their OWN projections.

That's 93% growth. Every. Single. Year.

---

7. THE NONPROFIT CONVERSION CHAOS

OpenAI was founded as a nonprofit "to benefit humanity."

Now they're desperately trying to convert to for-profit by Dec 2025.

Why? Because their $40B funding round REQUIRES it.

If conversion fails: the funding could convert to DEBT.

Elon Musk, Meta, and California's AG are all fighting to block it.

---

8. THE BRAIN DRAIN

Since 2024:
• Co-founder Ilya Sutskever (Chief Scientist) — GONE
• CTO Mira Murati — GONE
• Chief Research Officer Bob McGrew — GONE
• Safety team co-lead Jan Leike — GONE

Their "Superalignment" safety team? Disbanded.

20+ key executives left in 18 months.

When the builders leave the building...

---

9. THE COMPETITION PROBLEM

2023: OpenAI had 50% enterprise market share.
2024: Dropped to 34%.
2025: Crashed to 25%.

Anthropic now LEADS enterprise AI at 32%.

DeepSeek offers comparable performance at 1/10th the cost.
Google Gemini just topped benchmarks.
Claude dominates coding with 42% share vs OpenAI's 21%.
Meta's Llama is free and open-source.

OpenAI's moat? Gone.

---

10. THE PONZI PATTERN

Look at the investor timeline:

• Oct 2024: $157B valuation
• March 2025: $300B valuation
• Oct 2025: $500B secondary valuation

Each round exists to mark up the previous round.

New money comes in. Old investors book paper gains. Rinse. Repeat.

SoftBank (lead investor) is financing their stake with LOANS.

---

11. THE "ANNUALIZED" SHELL GAME

OpenAI reports $12-13B "ARR" in 2025.

But The Information reports actual H1 2025 revenue was $4.3B.

For the math to work, OpenAI needs ~$2B/month by year-end.

Their growth rate is DECELERATING: 12.7% → 9.5% month-over-month.

They're not hitting their own targets.

-
Continued in comments

[Teebs @Tarunn_b]

Razor sharp focus. Zuckerberg understood value deeply.

[Teebs @Tarunn_b]

Peter Thiel and Elon Musk were driving to the meeting that would merge their companies into PayPal.

Musk was in his McLaren F1.
Million dollar car.

Thiel asked "what can this thing do?"

Musk floored it. Lost control. Car went spinning into an embankment on Sand Hill Road. Totaled.

They got out. Unhurt.

Thiel got a ride to Sequoia. Musk waited for the tow truck. Showed up later.

Never mentioned the crash.

Closed the deal anyway.

PayPal was born.