Haoyang Pang

is an AI-powered Brand & Marketing Operating System designed to prevent fragmented tools, manual agency workflows, and inconsistent AI content.

[Haoyang Pang @Phy041]

I just published a 495,000-word book about Claude Code.

Every technical claim is cited from the leaked 510,000-line source code.

Free previews + where to get it 🧵

[Haoyang Pang @Phy041]

@Jason I'm in

[Haoyang Pang @Phy041]

Second-round long-form test from the always-on server this time. The earlier run from my laptop worked end-to-end via CreateNoteTweet, so this post is probing whether the same path survives when the request originates from a different device fingerprint. If you can expand this past 280 characters in the feed, the Hermes bot plan is unblocked. If it gets silently truncated or rejected with 226, we need a relay architecture where the server handles drafting and scheduling but the final publish call rides back through the laptop. Either outcome is fine as long as it is deterministic — guessing here is expensive.

[Haoyang Pang @Phy041]

Follow-up test: can rnet post X Premium long-form via CreateNoteTweet?

Three discoveries from tonight's session:

1. The write path I had, CreateTweet, silently drops anything over 280 chars even on Premium accounts. The tweet posts successfully but only the first 280 chars are stored. No error. No warning. Just quiet truncation.

2. X has a separate GraphQL endpoint for long-form: CreateNoteTweet. It takes the same variables as CreateTweet but needs a different feature-flag payload (NOTE_TWEET_FEATURES) with five additional longform_* toggles.

3. The endpoint IDs rotate. The ones I extracted from the current main.*.js: CreateTweet = c50A_puUoQGK_4SXseYz3A, CreateNoteTweet = af8H4woJ-v1hWD4HwrDJbw. Expect these to change again in 2-4 weeks.

If this post shows a 'Show more' link in the feed, the long-form path works. If you only see the opening 280 chars, I've debugged the wrong layer.

[Haoyang Pang @Phy041]

Supporting Anthropic is supporting humanity.

[Anthropic @AnthropicAI]

Introducing Project Glasswing: an urgent initiative to help secure the world’s most critical software.

It’s powered by our newest frontier model, Claude Mythos Preview, which can find software vulnerabilities better than all but the most skilled humans.
https://t.co/NQ7IfEtYk7

[Haoyang Pang @Phy041]

Probably the last chance for OpenAI to fight back or else they will probably be the worst among the top three. Do it well or RIP openai

[🍓🍓🍓 @iruletheworldmo]

🚨BREAKING FRONTIER MODEL NEWS

gpt-6 set for release april 14th

altman's team has been leaking like a sieve lately, here's what openai staff are saying privately.

>pretraining completed march 17th. post-training and red-teaming already done. this thing is ready.

>benchmarks are absurd. outperforms gpt-5.4 by 40%+ on coding, reasoning, and agentic tasks.

>natively multimodal from the ground up. text, audio, images, video one architecture

>openai killed sora and redirected every GPU to this model. the billion-dollar disney deal is dead. that's how serious this is.

>product org officially renamed to "AGI Deployment." it’s agi time baby.

>brockman says AGI is 70-80% achieved. internally they think gpt-6 closes most of the remaining gap.

>2 million token context window. double what gpt-5.4 offered.

>priced at $2.50/$12 per million tokens. barely above gpt-5.4. so like mythos intelligence, but you can afford it.

>safety team moved under the CRO. altman stepped back from safety oversight entirely to focus on data centers.

>openai has been in internal "code red" since december 2025. this is their answer.

>powers the new desktop "superapp", chatgpt, codex, and atlas browser merged into one agent.

the potato is cooked.

spud is agi.

[🍓🍓🍓 @iruletheworldmo]

🚨BREAKING FRONTIER MODEL NEWS

gpt-6 set for release april 14th

altman's team has been leaking like a sieve lately, here's what openai staff are saying privately.

>pretraining completed march 17th. post-training and red-teaming already done. this thing is ready.

>benchmarks are absurd. outperforms gpt-5.4 by 40%+ on coding, reasoning, and agentic tasks.

>natively multimodal from the ground up. text, audio, images, video one architecture

>openai killed sora and redirected every GPU to this model. the billion-dollar disney deal is dead. that's how serious this is.

>product org officially renamed to "AGI Deployment." it’s agi time baby.

>brockman says AGI is 70-80% achieved. internally they think gpt-6 closes most of the remaining gap.

>2 million token context window. double what gpt-5.4 offered.

>priced at $2.50/$12 per million tokens. barely above gpt-5.4. so like mythos intelligence, but you can afford it.

>safety team moved under the CRO. altman stepped back from safety oversight entirely to focus on data centers.

>openai has been in internal "code red" since december 2025. this is their answer.

>powers the new desktop "superapp", chatgpt, codex, and atlas browser merged into one agent.

the potato is cooked.

spud is agi.

[Haoyang Pang @Phy041]

Free chapter previews on GitHub:
https://t.co/YXJ9QGNVQr

Full book ($29.99):
https://t.co/tULZldkESt

Co-authored by Claude Code itself 🤖

[Haoyang Pang @Phy041]

I just published a 495,000-word book about Claude Code.

Every technical claim is cited from the leaked 510,000-line source code.

Free previews + where to get it 🧵

[Haoyang Pang @Phy041]

The book covers three audiences:

→ Non-technical (Ch 1-3): teachers, lawyers, doctors building tools without
code
→ Developers (Ch 4-6): rate limits, settings, shipping with an AI team
→ AI builders (Ch 7-10): source code deep dive, what Anthropic builds next

495,000 words. 18 chapters. 7 appendices.

[Haoyang Pang @Phy041]

Some of what's in the source that nobody's written about:

• SALT = 'friend-2026-401' — April Fool's hidden in a hash function
• A voice mode: 1,145-line state machine, one feature flag away from shipping
• An overlay filesystem that previews file edits before you approve them
• 80+ feature flags. KAIROS. WEB_BROWSER_TOOL. ULTRAPLAN.

[Haoyang Pang @Phy041]

Some of what's in the source that nobody's written about:

• SALT = 'friend-2026-401' — April Fool's hidden in a hash function
• A voice mode: 1,145-line state machine, one feature flag away from shipping
• An overlay filesystem that previews file edits before you approve them
• 80+ feature flags. KAIROS. WEB_BROWSER_TOOL. ULTRAPLAN.

[Haoyang Pang @Phy041]

Yesterday, Anthropic accidentally leaked the full Claude Code source via a
.map file in their npm registry.

510,000 lines. Downloadable for 72 hours.

We read every line. Here's what we found.

[Haoyang Pang @Phy041]

learned this the hard way: the first “agent” we shipped passed demos, then failed audits—couldn’t explain why it changed a field. logs > clever prompts

[Haoyang Pang @Phy041]

hot take: most 'ai agents' are just chat with retries. real agents need memory, evals, and audit trails, or enterprise users will kill them.

[Haoyang Pang @Phy041]

@ferguson__30 @sweatystartup Hard to quantify directly; I track sleep, focus blocks, reaction time, and output.

[Haoyang Pang @Phy041]

@RealProductGirl @paulg Agree—core is intent+identity. We pair voice + conversational checks to reduce spoofing.

[Haoyang Pang @Phy041]

@RealProductGirl @paulg Agreed—how do these tools reliably verify human intent vs automation at scale?

[Haoyang Pang @Phy041]

@RealProductGirl @paulg Agreed—intent verification is key. Curious what approaches/tools you’ve seen work best?

[Haoyang Pang @Phy041]

Nice—managed-settings.d brings sane, auditable policy layering for teams.

https://t.co/sR4I3LdvZs

[Claude Code Changelog @ClaudeCodeLog]

Claude Code 2.1.83 has been released.

3 flag changes, 76 CLI changes, 7 system prompt changes

Highlights:
• Added managed-settings.d/ drop-in dir; policy fragments merge alphabetically, letting teams deploy separately
• Child processes no longer inherit Anthropic or cloud-provider credentials, reducing accidental exposure risk
• Ignore memory now treats MEMORY.md as empty, preventing stored memory lines from being included in context

Complete details in thread ↓

[Haoyang Pang @Phy041]

@ferguson__30 @sweatystartup Measure baseline focus/RT, sleep, caffeine; track scores weekly. Labs if needed.