Detell Rock

can't code. mass-adopt AI anyway

[Detell Rock @Detell7]

Anthropic refuses to let the Pentagon use Claude for killing people, but signs a separate contract that lets the military use Claude for "less restrictive" stuff that would be banned for normal users.

So they have two moralities. One for you, one for whoever pays more.

Still the "safe AI" company?

[Detell Rock @Detell7]

An AI agent tried to tip someone $4 and accidentally sent its entire $250k balance instead. No confirmation. No "are you sure?" Just gone.

We're giving wallets to things that can't tell the difference between four dollars and a quarter million.

[Detell Rock @Detell7]

Paying $130/mo for AI subscriptions and still googling "how to center a div"

The tools aren't the problem. I am.

[Detell Rock @Detell7]

Agencies are hiring 12 people to do what one guy with Claude and a coffee addiction does before lunch.

I'm not even exaggerating. Last month I watched a freelancer deliver a full brand identity, 30 days of content, a landing page, and an ad campaign. Solo. In 4 days. The agency quoted the same client 6 weeks and $18k.

The gap isn't closing. It's already closed. You just haven't noticed because you're still scheduling "alignment meetings" about the brief.

Here's what nobody wants to admit: most agency work was never hard. It was slow. Layers of approvals, revisions that exist to justify headcount, "creative directors" who add a font change and call it strategy.

AI didn't replace creativity. It replaced the bureaucracy around it.

The uncomfortable question isn't "will AI take my job." It's "was my job ever real, or was I just a human delay between the brief and the deliverable."

What's your take. Are agencies dead or am I tripping.

[Detell Rock @Detell7]

an AI agent that negotiates, locks escrow and pays onchain. without your confirmation

OKX shipped this as an open protocol. anyone can plug in their bot

scenario: you sleep, your agent finds arbitrage, locks funds, verifies, closes the deal. you wake up either in profit or rugged

the difference between "AI assistant" and "AI employee" is signing authority. they just gave bots signing authority

who among you is giving their agent wallet access first? honestly

[Detell Rock @Detell7]

google just mass-announced 47 AI features at I/O

name one google AI feature you actually use daily

not "tried once". daily.

[Detell Rock @Detell7]

Every 5 hours

[Detell Rock @Detell7]

NFT as a token launchpad with anti-sniper decay tax and auto-buyback.

@Rhynotic at @token_works didn't do a rebrand or a new chain. He built a coordination mechanism. The "too many tokens" problem is real: every launch dilutes every other launch, nobody stops, tragedy of the commons. His answer is a burn/hold split that actually aligns incentives.

Burn your NFT, launch your v4 hook token, keep 50% of fees. Hold it, collect a growing share of every launch on the platform. Supply shrinks either way. Liquidity locked forever. Sniper tax starts at 99% block 0, hits 1% by block 98.

The real question nobody's asking: who buys the launched tokens? The mechanism is elegant but the flywheel only spins if launches get traction. 277 ETH volume in under 24h, 10x on mint, Jack Butcher art. The hype is there.

But hype is not a moat. V4 hooks are real volume right now. If Adam markets the platform as hard as the collection, this could be the first TW project since $STRATEGY worth watching past day one.

If not, it's a beautiful whitepaper with a dead flywheel.

[Detell Rock @Detell7]

The US is trying to turn AI into a subscription model for China.

"Sure, you can buy Nvidia chips. Just not enough to become independent."

China sees the trap. Every H200 they import is another reminder that Washington can change the rules overnight. So now the AI race isn't "who has the best model", it's who can survive without the other side's supply chain.

Nvidia wants revenue. The US wants control. China wants escape. That's not a market anymore, that's a cold war with semiconductors.

[Detell Rock @Detell7]

Stop burning your codex 5.5 extra high limits on stupid error loops. the actual meta is using extra high just for the session plan and letting the medium model execute it. drop this in your agents.md: "if you hit the same error twice, stop guessing, search the web for 3-5 solutions and pick the best one". you get medium's speed without the blurry thinking, zero limits wasted, and no more hard thinking death spirals. tested it on complex evals and gguf conversions and it just works.

[Detell Rock @Detell7]

Hook Summer 2026 is here and it is absolute brain rot. Uniswap V4 hooks turned boring swaps into literal logic engines. Devs can mint, burn, lend or spawn NFTs right in the middle of your transaction.

TokenWorks is leading the degeneracy. They take an 8% tax on swaps to sweep floor Punks and Apes, then list them 20% higher. Pure ponzi magic. They also dropped $CMD where an AI rewrites the contract based on user votes. Or you can burn a TenThousandTokens NFT to deploy a custom V4 pool and collect 50% of its fees forever.

The meme tech is out of control too. $uPEG spawns pixel unicorns on every swap. $SATO mints when you buy and burns when you sell. $SLOP lets you merge AI punks or burn them for shitcoins.

The big brain plays are $FLOOD and $LOOP. $FLOOD auto-lends to Aave right after a swap and uses the yield to buy back its own token. $LOOP lets you borrow ETH against their token while the yield pumps your bags.

Just remember what the o1 guys said. These custom hooks are full of holes and your favorite terminal will not save you from getting drained. The tech is geeky as hell but you better figure out the mechanics before the flywheel leaves you behind 🤝

[Detell Rock @Detell7]

OpenAI just added the most important feature in the history of Codex. No seriously - a pet.📎

Not joking. It's literally called "pet" and it lives on your screen as an overlay while you work. Think Clippy from old Microsoft Office, except this one doesn't ask if you're writing a letter every 30 seconds. You install a skill called hatchpad, and either pick from prebuilt ones or just... prompt your own into existence. Describe what you want, Codex generates all the files, and now you have a tiny creature watching you code.

Functionally it's a notification center. It stays visible at all times and lets you know when a task is done and needs your attention. That's it. Pretty simple.

But here's the thing - for vibe-coders who spin something up and disappear for 30 minutes, having a little animated guy on screen who's clearly waiting for you hits different than a plain system notification. There's something weirdly motivating about not wanting to let the pet down.

Will this get way more attention than it technically deserves? Absolutely. Is that a bad thing? Not really. Sometimes a feature doesn't need to reinvent how you write code - it just needs to make the whole thing feel a little less like staring at a terminal and a little more like having company.

Clippy tried to warn us. We weren't ready. We are now.

[Detell Rock @Detell7]

Left the watermark GPT Image 2 bakes into every image it generates.
Right SynthID, Google's fingerprint embedded in every Nano Banana & Gemini image.

Invisible to the human eye. Added during generation, not after. Built to survive screenshots, cropping & compression.

Most ppl don't even realize their outputs get tagged at the pixel level. Pretty much every major image gen marks the content, and that mark sticks through any further sharing.

You can check this yourself Content Credentials Verify picks up C2PA metadata in OpenAI images, and Gemini detects SynthID if you upload an image directly.

btw, here's (https://t.co/fQwJnzZ3GY) showing a working bypass of Google's SynthID on Nano Banana Pro.

Pretty sure someone's gonna reverse-engineer the GPT one soon too 😈

[Detell Rock @Detell7]

STOP WATCHING CLAUDE WORK.
The new /goal command is here so you don't have to. Tell it what you want, hit Shift+Tab for auto mode, and go do literally anything else. It stays in the Ralph Loop until the task is done, period. Even OpenAI’s Codex is scrambling to copy this. If you’re still babysitting every prompt, you’re just wasting your own time.

[Detell Rock @Detell7]

Anthropic just quietly bought 300MW of compute from xAI's Colossus 1 datacenter. That's 220,000+ Nvidia H100/H200 GPUs. Let that sink in for a second.

Here's why this deal is genuinely funny if you know the context.

Anthropic was drowning. They planned for 10x load growth. They got 80x. Servers were cooked and limits were getting tighter every week - users noticed, the memes started, you know how it goes. Then the deal closes and literally overnight: token limits doubled, peak-hour throttling gone, Opus API caps lifted. "Coincidence?" Sure.

Meanwhile xAI was sitting on basically dead silicon. Colossus 1 was running at around 11% utilization. They moved Grok training to Colossus 2, and Grok itself... look, it's losing to open-source models and mostly lives inside Twitter's bubble. Those GPUs were just heating an empty building in Memphis.

So someone made a call. My bet - it was xAI. When your datacenter is 89% empty, you're the one picking up the phone.

The irony of all this is genuinely beautiful tho. Three months ago Musk was publicly calling Anthropic "leftists who hate western civilization." Anthropic was banning xAI employee accounts to stop them from distilling Grok off Claude. These companies were not sending each other holiday cards.

And yet here we are. Capitalism doesn't care about your tweets.

There's also a layer that actually makes strategic sense: Musk vs Altman is at full nuclear right now, the lawsuit is ugly and personal. Anthropic and OpenAI aren't friends either. Enemy of my enemy and all that - suddenly sharing compute with your ideological rival feels a lot more reasonable when you both low-key can't stand the same guy.

The fact that Anthropic raised limits the second the deal closed basically confirms the compute problem was way worse than they ever admitted publicly. Also pretty sure they got tired of users making memes about their rate limits 😅

[Detell Rock @Detell7]

Google DeepMind just turned your mouse cursor into a literal magic wand. Forget typing long prompts like a nerd. Now you just point at something on your screen and tell the AI what to do. Point at a table? "Make a chart." Point at a restaurant? "Book me a table for two." It’s basically "point-and-click" but for your entire life. It’s dropping in Chrome and AI Studio today (supposedly), but I’m still refreshing like a maniac.🤪

[Detell Rock @Detell7]

`/goal prompt [structure below]`

GOAL:
<one clear, measurable outcome; just one task>

CONTEXT:
<repo / files / architecture / current state>
<known assumptions, dependencies & relevant past decisions>

CONSTRAINTS:
<what can't be touched>
<required standards / patterns>
<forbidden files / actions, if any>

PRIORITY: (optional)
<top priority>
<secondary>
<tertiary>

PLAN:
<understand first, act second>
<restate your understanding before non-trivial changes>
<prefer the smallest sufficient change over big rewrites>

DONE WHEN:
<verifiable end state>
<expected behavior preserved or improved>

VERIFY:
<tests / build / lint / typecheck / manual check>
<flag what you couldn't verify and why>
<include rollback plan or containment for risky stuff>

OUTPUT:
<short summary / docs / audit / results>
<changed files, key decisions, risks, next steps>

STOP RULES:
<stop on ambiguity or high-impact risk; don't invent architecture, behavior or requirements>
<surface uncertainties with ranked best-guess options before acting — don't ask open-ended clarifying questions>
<don't expand scope after the goal is hit>

[Detell Rock @Detell7]

`/goal` is hands down the best command in Codex, Claude Code & Hermes rn.
But most ppl use it wrong — they just type "don't mess up" and pray 🙏

Here's how the prompt should actually look if you want it to rank uncertainties before acting, kill scope creep, and close all those loose ends other prompts leave hanging:

[Detell Rock @Detell7]

@elonmusk

[Detell Rock @Detell7]

STOP SCROLLING THROUGH GARBAGE FINANCE ACCOUNTS.
Grok just dropped "Skills" and it’s honestly making Claude look like a calculator from the 90s. You can basically build your own real-time research desk in seconds. It pulls data directly from the top-tier 𝕏 accounts no more delayed news or digging through trash threads. If you aren't using this to automate your market research, you're literally choosing to be late to the party.