Jason Wang

Opportunist, 80% optimistic 20% pessimistic. everything is a construct.

[Jason Wang @jasonzoid]

@bcherny @lucas_builds Hmm let’s find out

[Jason Wang @jasonzoid]

@commandenteSD Lmfao

[Robert Youssef @rryssf]

Claude Code subagents bill at Opus rates by default.

Haiku 5.5 dropped just now at $0.10 input and $0.50 output per million tokens for requests under 100K tokens, and five times that above. Opus 5.5, the default main model, costs $4 and $20. That's a 40x gap on every helper you spawn.

A subagent runs in its own context window with its own system prompt and tools, then returns only a summary to the main conversation. If you set no model, it inherits the main conversation's. The built-in Explore agent does too.

routing subagents to Haiku 5.5 👇
→ one agent: add model: haiku to the frontmatter of .claude/agents/<name>.md
→ every agent: set CLAUDE_CODE_SUBAGENT_MODEL=haiku
→ Explore included: also set CLAUDE_CODE_SUBAGENT_MODEL_FORCE=1 (v2.1.257 or later)
→ the haiku alias only means 5.5 on v2.1.293 or later, on the Anthropic API

On Bedrock, Google Cloud and Microsoft Foundry, that alias still resolves to Haiku 4.5.

Hand it summaries, compactions, database queries and classification. Keep complex agentic coding on the big model. On Terminal-Bench 4.0, Haiku 5.5 scores 39.2% and Sonnet 5.5 scores 70.6%.

Anthropic's engineering team measured multi-agent systems at about 15x the tokens of a plain chat. Up to 20 subagents can run at once by default, so the worker's rate moves the bill.

Check that the routing took. Issue 100082 on the Claude Code repo reports builds 2.1.283 to 2.1.290 ignoring model: haiku in frontmatter. The reporter found 279 of 301 subagent runs on Opus 5.5.

Open any transcript in the subagents folder under ~/.claude/projects/. The first assistant message records which model actually ran.

[Jason Wang @jasonzoid]

@brandontseng2 Imagine Xbat lands on a submersible unmanned platform

[Jason Wang @jasonzoid]

@DTrades98 Interesting, bitcoin leveraged long liquidation style

[marfin @marfinxx]

this is pure f*cking treasure

A Stanford research group found how to orchestrate Claude Opus 5.5 and GPT-6 Astra together, and it completely breaks the resampling wall

most developers think test-time compute means running 32 blind retries on one model. you burn API tokens repeating the same 2-3 error clusters in loops

Stanford's paper (arXiv:2608.05643) splits the compute between two different brains:

> breadth search: GPT-6 Astra runs high-entropy exploration (T=0.8) to discover divergent solution paths
> step audit: Claude Opus 5.5 scans each step at T=0.0 and catches the exact logic fallacy
> local surgery: Opus rewrites only the broken step instead of dumping the whole reasoning chain
> verifier-free consensus: majority voting locks the proof without relying on drift-prone PRMs

the benchmark metrics outclass standard Best-of-N sampling:

> 91.2% recovery rate on detected reasoning flaws
> under 1.8% regressions on correct branches
> MATH500 jumps from 29.6% to 58.0% (+28.4% absolute gain)
> AMC-2023 improves by +7.5% with zero reward models

Astra explores the maze. Opus repairs the logic

you stop burning tokens on blind retries. you run verified reasoning depth

[Lunar @LunarResearcher]

Andrej Karpathy just dropped a 1-hour Stanford lecture on AI engineering from scratch:

"You can basically delete everything… delete everything, keep Graph"

Here are the core ideas:

10% → LLM: "I treat GPT like a general-purpose computer that can be reprogrammed at runtime"

30% → Prompt: "I give that computer a program written in natural language"

50% → Agent: "I wrap the model with a goal, context, memory, and tools that turn prediction into action"

70% → Loop: "I separate the inner loop, where the model learns from context, from the outer loop, where training updates the weights"

100% → Graph: "I organize communication as data-dependent message passing across directed graphs"

This 1-hour Stanford lecture will teach you more about AI engineering than most 100 random YouTube tutorials combined

"Just chop everything up and throw it into the mix"

Bookmark it and watch later

Then read the full article below ↓

[Annatar.md @AnnatarXBT]

Claude Code tip that's pure f*cking treasure:

Opus 5.5 + Sonnet 5.5 + Fable 5.1 as one team
plan on high, delegate on medium, keep Fable on call

the team:
• Opus 5.5 (high) - plans and ships
• Sonnet 5.5 (medium) - explores, edits, runs tests, pulls docs
• Fable 5.1 (/advisor fable) - watches the session, speaks up when it counts

Fable wakes up 3 times:
1. before a plan - is this the right approach?
2. when an error repeats - am I digging in the wrong place?
3. before "done" - what did I miss?

Jev engineering goes one layer lower: which file, which tool, retry or stop go to Jev in under half a second, so the big models only see the real forks

paste this into Claude Code 👇

"Rebuild my Claude Code setup around one team.

1. Subagents
- reuse anything in ~/.claude/agents and .claude/agents that fits explorer, worker or researcher, draft the rest
- all on model: sonnet, effort: medium
- explorer and researcher read-only (Read, Grep, Glob, WebFetch), only worker gets Edit and Bash
- each returns under 15 lines with file:line refs

2. Main session
- in ~/.claude/settings.json set effortLevel to high and advisorModel to fable
- report anything that overrides this (CLAUDE_CODE_DISABLE_ADVISOR_TOOL, DISABLE_TELEMETRY, CLAUDE_CODE_EFFORT_LEVEL) and change nothing

3. Add to ~/.claude/CLAUDE.md
- consult the advisor before a plan touching 3+ files, when the same error shows up twice, and before calling a long task done
- never let two subagents edit the same file

Run /agents to confirm, then show everything as one diff. No edits until I say go."

↳ https://t.co/E69hNP0rfm

[Vox @Voxyz_ai]

Put a few Claude Code commands together and you have a full maintenance schedule, with minor and major services, just like a car.

Every change an agent makes looks fine on its own. A few months of them add up to duplicate code, a structure that keeps getting more tangled, and a pile of outdated rules. Like a car that skips its service for six months, it breaks down sooner or later.

𝗪𝗲𝗲𝗸𝗹𝘆 𝗺𝗶𝗻𝗼𝗿 𝘀𝗲𝗿𝘃𝗶𝗰𝗲
→ Code keeps getting more repetitive: run /simplify on the folders that changed most (for example /simplify src/api). It finds duplication, needless abstractions and inefficient code, and fixes them
→ Small bugs piling up: run /code-review on the same folders. It looks mainly for bugs and only reports, without touching the code

𝗠𝗼𝗻𝘁𝗵𝗹𝘆 𝗺𝗮𝗷𝗼𝗿 𝘀𝗲𝗿𝘃𝗶𝗰𝗲
→ Context getting crowded, hooks getting slow: /doctor finds the skills, MCP servers and plugins taking up context for nothing, flags slow hooks and trims CLAUDE.md. It asks before changing anything
→ Rules getting messy: /doctor prompt-audit finds outdated and conflicting rules in your CLAUDE.md, AGENTS.md and skills (v2.1.283+)

𝗪𝗵𝗲𝗻 𝘁𝗵𝗲𝘀𝗲 𝗰𝗼𝗺𝗲 𝘂𝗽
→ Install or launch steps changed: rerun /run-skill-generator. It saves the new steps as a skill in the project, so the built-in /run and /verify launch and check your app the new way
→ Not sure where things keep going wrong: /insights gives you a report on how you've been using Claude Code and where problems keep showing up
→ An agent went in circles in a session and you had to correct it several times: run /retro (a third-party skill from mattpocock/skills) to find the rules, checks and setup worth changing

[shobhanai @HeyShobhan]

Andrej Karpathy spent 8 years at OpenAI and Tesla

Last week, he condensed everything he knows into one free 2-hour lecture

Agents → Loops → Harness → Self-Improving Systems

People pay $14K for bootcamps that teach less than this

This lecture beats most paid AI engineering courses

You probably don't have 2 hours right now

Don't let this disappear from your feed

Watch it

Then read the article below

[ClaudeDevs @ClaudeDevs]

We're adding a new plugin to Claude Code: You should Know.

It scans Claude's output for important information you might miss to help keep you in the loop.

Enable it with:

/plugin enable cc-plugin-you-should-know@builtin

[Nandkishor @devops_nk]

Only software engineers will understand why this is actually funny.

[Jason Wang @jasonzoid]

@ShanghaoJin 如今的台灣人根本不屑大陸遊客，整體素質偏低

[Jason Wang @jasonzoid]

@commandenteSD Nostalgic 80s

[Lancelot🔱 @commandenteSD]

1982 Kitt from Knight Rider

2026 Grok with TSLA with FSD

[Jason Wang @jasonzoid]

@commandenteSD Lmfao

[Eric Daugherty @EricLDaugh]

🚨 OMG. President Trump has Xi CRACKING UP in how he presented the Biden autopen portrait

"TRUMP, and then TRUMP...and this is the autopen!"

*Xi begins laughing*

"It's BIDEN!"

This man is absolutely priceless 😭😭

Even the other members of the Chinese delegation are cracking up!

Aside from showing off the helipad, this might be the highlight of Trump's morning 🤣

Trump first put up this autopen portrait roughly 1 YEAR AGO on the "Presidential Walk Of Fame" and it's remained up ever since!

Welcome to the Trump White House 🔥

[Boris Cherny @bcherny]

I used Opus 5.5 to formally verify the Claude Agent SDK using Lean. A couple short prompts = 16 PRs fixing various bugs and race conditions. Video attached.

TLA+ also works well. I sometimes combine Lean and TLA+ to look for issues around data flow, concurrency, and state mgmt.

I don't know either language well, but Claude is excellent at both. This approach is super useful for formally modeling your code and finding bugs that a human probably wouldn't have spotted.

Is formal verification the future of coding (or at least, bug finding)?

[Sola Awodiya @Solaawodiya]

Lmao 🤣. Anthropic and OpenAI now of slowing down

[NoLimit @NoLimitGains]

A professor turned $100K into $7M rebalancing two ordinary stocks.

No forecasting. Just exploiting volatility.

Watch this video today. It will change your life.