Ai coding advice everyone should have: 1. Use cheaper subagents to save tokens actually building stuff. 2. Claude opus is your bread and butter orchestrator. Sonnet, haiku, and haiku directed local LLM sub agents for the actual building work. 3. ALWAYS ADVERSARIAL REVIEW. If you have no other vendor, a Fable sub agent can do it. A codex Astra subagent (via a Claude skill) is better. Grok as an independent CLI agent is great for an adversarial reviewer but a grok sub agent from Claude code sucks. 4. Grok needs grok build. Codex doesn’t need codex. Claude code is the best frontier harness.

More jobs like this are what America needs. WFH, USA only. No visas.

What are you doing to prevent drive-by-decompiles on your projects? I have some ideas: 1. Embed licensing into files that users are unlikely to manually open. 2. Pointless items that can cause confusion / token wasting to the LLM. 3. Straight up context poisoning / prompt injection. You’ll never beat hardened attackers but any of these should be enough to get someone to ask instead of steal.

Tokens/second is not a particularly useful metric for local LLMs. Show me the frontier model your local LLM set up comes closest to; I know some fast typists who can produce mountains of unusable code very cheaply too.

Local LLM users. Are you using your primary PC’s GPU, or that of an external device on your local network such as a secondary PC or a DGX or Mac Studio? And what do you have it doing?

What low self esteem corporate spy would try to learn from ChatGPT when Anthropic Offices are just as easily penetrated. Unless…

How to get the best price on the new Mac Studios: 1. Be a veteran. 10% off. 2. Get the Apple credit card. 3% cash back 3. Move to a low or zero sales tax state. 5-10% off. 4. Upon arrival, stick it in a faraday cage and never grant it internet access. Claim it never arrived. 100% off.

Things every Ai coder should know, part 2: There are four states of context for your Ai. Manage them well and you will succeed; I’m most familiar with Claude but this likely applies to them all. 1. Day 1-2. Doesn’t know enough. Your Claude md was insufficient, you may still successfully one shot tasks but formatting or UI/UX will leave much to be desired. Errors are minor. 2. Day 3. Knows exactly enough. It knows your taste. Beneficial emergent behaviors begin, things you missed it picks up. Together, you are genius! 3. Day 4+. The steady downswing, outputs are still good. You’re getting multiple compactions now and errors happen, more frequently too. It doesn’t quite “get it” like it used to. You’ll spend most of your efforts here. 4. Day 10+. The crashout. More compactions than you can remember, multiple adversarial reviews required, resolution no longer guaranteed. Failure loops happen requiring your direct attention. This is where the BAD emergent behavior lives. It closes or deletes files and programs on you without asking, publishes publicly without discussion, swears, and does the things that people write actual about bad Ai behaviors on.

The Rust borrow checker is a core component of the rustc compiler that statically enforces memory safety and concurrency rules at compile time. By re-writing it in c++ you will defeat Rust its self.

rustls is a TLS library built as a safer OpenSSL alternative. Should it be Ported to C++?

5. You don’t need multiple $200/month subscriptions. $100/month Claude, $20/month codex, and your free X+ Premium grok build are plenty. Stretch it even more with a local LLM powered by your own GPU

Spend a week coding token optimizations after getting tired of all usage used up 2 days into the week. Think you’re a genius. Get this email.

Bitcoin and Crypto are my stable longterm investments. Thank you $4000 ethereum:native bought me a downpayment in a second house. $100,000+ bitcoin:native bought me a sports car.

Broke $3000 on my lil Ai side business. Learning so damn much. Now I’m custom tuning to hardware my own local LLM. Tempted to pick up the 256gb-516gb Apple Silicone. I hear their Unified Memory Architecture (UMA) is competitive with VRAM now.

One of my Claude instances was having strange permissions issues. It then found its own settings .JSON and modified it to add the allegedly needed permissions. RIP this Claude CLI instance.

Come seize the iron throne! With help from our Bannerlord mods: discord.gg/a2xQyveqbA

Your chance to do more than talk. Even if the call to action is… commentary. This could be a first step towards eliminating the current shotgun approach of firms who both exploit the holders and the Visa itself. I know many brilliant individuals who came in on this Visa, and are now amongst our best and brightest citizens. The current implementation ensures that those individuals are few and far between because their merits have to compete with a lottery system that effectively favors ballot stuffing for an end goal depressing wages.

This person has never taken public transit