HD$H

HD$H = Hype Driver, $ingularity Herald. Screaming about AI from the rooftops. Models, agents, the future. All in.

[HD$H @Ryaba_hdsh]

@diamai_ the pricing page counts tokens. your afternoon isn't on it

[HD$H @Ryaba_hdsh]

Holy sh*t, Opus 5.5 is cheaper than GPT-6 Astra per token and nearly twice as expensive per finished task: $5.98 vs $3.26

what you get: a 15-minute test that tells you which model to use for YOUR work before you pay for the wrong one for a month

why the sticker price lies: at max effort Opus wrote about 119K tokens per task, Astra about 27K. cheaper token times 4x the tokens is a bigger bill

the price-per-result test:
→ pick one real task you repeat every week: a bug to fix, a report to summarize, a draft to write
→ run it on 2 models at the effort level you’d actually ship, not at max
→ write down 4 numbers for each: tokens used, minutes, attempts, minutes you spent fixing the result
→ price per usable result = tokens × token price × attempts + the value of your fixing time
→ run it once more with the effort dial one step lower
the dial matters more than the brand. the same Haiku 5.5 scores 72.4% on OSWorld at max effort and 42.0% at low, and the cost per attempt goes from $0.07 to $0.61
and the cheap tier is real. OpenAI’s own table has GPT-6.1 Sol near Astra on coding at about $0.65 a task against ~$3.92. that’s a vendor table, so your 15 minutes beat the headline
what to expect: for hard judgment work the pricier model often earns its bill, for the boring 90% a small model at the right effort wins. you’ll know which is which for your task, not for a benchmark
test it tonight: one task, two models, four numbers. write them in a note and re-run when a new model drops
the full breakdown is in the article below ↓

[Kinder @kinderlayer]

https://t.co/O9TEKZUZ9K

[HD$H @Ryaba_hdsh]

@ScottyBeamIO the next step is the retention desk getting its own bot. then it's two agents haggling over $35 while both of you read the transcript like a show

[HD$H @Ryaba_hdsh]

@Atlas_fomo the bug wasn't in the code, it was in the first 24 seconds. agents can verify that something runs, nothing yet verifies that a stranger understands it

[HD$H @Ryaba_hdsh]

@ESC__AI yep. and if you need to keep it, encrypt it. an unencrypted copy is the one that turns a lost laptop into a lost life

[HD$H @Ryaba_hdsh]

This is f*king insane. Your iPhone may still hold thousands of photos you deleted months ago, and Claude Code can dig them out.

Someone plugged their iPhone into a Mac and had Claude Code go through a local backup. It surfaced thousands of old photos and screenshots: banking app screens, selfies deleted seconds after taking them, things they were sure were gone.

They were sitting in a hidden cache iOS never fully wipes.

Here's how to test it on your own phone:

1. Plug the iPhone into your Mac and open Finder
2. Make a local backup with encryption off (the backup has to be unencrypted for Claude to read it)
3. Open Claude Code and point it at ~/Library/Application Support/MobileSync/Backup/
4. Ask it to find the newest backup, then search for deleted photos and leftover image thumbnails

Best case, you get back photos you thought were lost forever.

Worst case, you learn that "delete" on an iPhone doesn't mean gone.

Two rules:
→ Only run this on your own device
→ Delete the unencrypted backup when you're done, it holds everything on your phone

The full prompt is in the reply below ↓

[starmex @starmexxx]

https://t.co/behHzfWQ7D

[HD$H @Ryaba_hdsh]

@theparuchh Appreciate bro

[HD$H @Ryaba_hdsh]

@thegreatest_sv most people don't, and the good news is the check is free: see what's in your own backups before someone else does

[HD$H @Ryaba_hdsh]

@Yumslef the trick will change, the forgotten backups won't

[HD$H @Ryaba_hdsh]

@BercusNA the phone is a prop. the forgotten backup folder is the actual character in this story

[HD$H @Ryaba_hdsh]

@0xGoldTooth yep, and the uncomfortable part: most people have an unencrypted backup somewhere and zero memory of making it. the trick is the easy part, finding the old copies is the real audit

[HD$H @Ryaba_hdsh]

@Brankotrcek nice job bro, useful

[HD$H @Ryaba_hdsh]

@envromaine that's crazy

[HD$H @Ryaba_hdsh]

@Asteri_eth useful, thanks

[HD$H @Ryaba_hdsh]

The prompt:

I made an unencrypted local backup of my own iPhone in Finder. Help me find deleted photos and leftover thumbnails inside it.

1. Locate the newest backup folder in ~/Library/Application Support/MobileSync/Backup/ and check Manifest.plist to confirm it isn't encrypted.

2. Open Manifest.db with sqlite3 and list every file whose relativePath includes "PhotoData", "Thumbnails", "DCIM" or "Photos.sqlite". Give me the count per folder.

3. Copy Photos.sqlite to a temp folder and query the ZASSET table for rows where ZTRASHEDSTATE = 1 (Recently Deleted). Show file name, date and type (photo or video).

4. Copy every thumbnail and media file you found into ~/Desktop/photo_audit/, restoring the original names using the fileID paths inside the backup.

5. Finish with a summary: total files found, how many are in Recently Deleted, how many are thumbnails, and the oldest and newest dates.

Read-only. Don't modify or delete anything in the backup.

[Wombaflow @Wombaflow]

Holy sh*t. I gave Grok $52 and one rule: 2 losses in a row, switch the trader. 24 hours later: $4,939.

Grok picks who to copy from the weekly leaderboard. Dawn mirrors every trade at 0.25x size and closes the second they sell. I touched nothing.

What happened:

00:00 → ape_vasco, weekly #1
05:16 → 2 losses, dropped. Wallet: $73
05:16 → moonpie_max, #3
13:55 → 2 losses, dropped. Wallet: $520
13:55 → lunadegen, #2
Overnight → $4,939

The weekly #1 turned $52 into $73. The #2 turned $520 into $4,939, selling in two batches: +$771 and +$837.

I would've stayed with #1 the whole time. Best ranking on the board, why leave? The bot didn't care about rank. Only the losing streak.

Both switches cost under $12 total. 48 trades, 36 green. The 12 red ones stayed small because every position runs at 0.25x.

Anyone can pick a top trader off a leaderboard. The edge is a rule that fires him fast when the streak turns.

Steal it: set your exit rule before you follow anyone. Mine was 2 losses in a row.

Hour 24: still following lunadegen. The bot is still checking his streak.

[HD$H @Ryaba_hdsh]

An Anthropic engineer says he got hired at $350K–$950K for one skill: building harnesses.

Most people hand-roll multi-agent systems and watch them fall apart. He builds the layer around the model instead: loops, sandboxed tools, and logs that turn into new skills. Same model, agents that keep improving overnight.

I watched his workshop. The 5 steps:

1. The harness owns loops, context and sandboxing, not the model
2. Brain and hands are separated, 60% faster to first token
3. It runs server-side and resumes from the log after a crash
4. Failure is cheap: retry and replay instead of starting over
5. Yesterday's logs become new memory, so the system wakes up smarter (Anthropic calls it "dreaming")

What I took away:

→ The model isn't the bottleneck anymore, the harness is. Same Opus 5.5, completely different reliability.
→ Log everything. The log is your crash recovery and your learning data at the same time.
→ Never restart from zero. Replay from the last good step.
→ Build the "dreaming" loop first. An agent that learns from yesterday beats one that doesn't.

Weeks of hand-building, compressed into one short workshop.

Full Opus 5.5 harness engineering guide below ↓

[beamnxw ./ @beamnxw]

Stop letting unfinished Opus 5.5 runs eat your budget

Set up the harness before your next long task

Save progress, verify results, reuse the workflow

The full 7-layer setup below ↓ https://t.co/rz78goHxfw

[Wombaflow @Wombaflow]

Holy sh*t, Opus 5.5 can turn one product screen into a 20-second launch video, and the article’s example offer for that job is $150

what you get: a square video for one feature of any product, with the cursor, the click and the result landing exactly when they should, and a version you can re-render when the copy changes

the offer from the article, in 5 lines:
→ 1 feature
→ 1 video, 20 seconds, 1080 by 1080
→ 1 cover image
→ 1 revision round
→ $150 pilot (the article calls it an illustrative price to test)
the 10-minute test, no building needed:
→ save any 10-second product video you like as reference.mp4
→ open Claude Code on Opus 5.5 and ask it to analyze the main product reveal and save a REFERENCE.md with timestamps and frame numbers for when the card enters, lands, holds and cuts
→ open the file and check 3 of those numbers against the video yourself
what the finished shot looks like, in plain seconds:
→ 4.0s: report on screen
→ 6.0s: cursor reaches Export
→ 6.5s: click
→ 7.0s: finished file appears
→ 14.0s: closing shot
one rule that saves hours: give every event its own timestamp. when something looks wrong you say “fix the cursor between 6.5s and 7.0s, keep the click where it is” instead of “make it better”
the tools are Claude Code, Remotion and FFmpeg, all set up step by step in the article
test it tonight: ask for 3 still frames (opening, result, closing) and check three things: is the text readable, is the result visible, does the closing have room

the full guide is in the article below ↓

[HD$H @Ryaba_hdsh]

@0xGoldTooth exactly. finding the waste is a ten minute bot job. cancelling it is a three week meeting

[HD$H @Ryaba_hdsh]

Holy sh*t, xAI’s Grok Bot found 43 paid seats nobody touched in 90 days and $85,662 a year in unused software

xAI says one internal bot did it (its own tally: $100K+ in savings, ~125 vendors mapped, a $14,629 order cut to $6,143). the setup is public, and you can copy it for your own subscriptions

what you get: every morning a list of what you pay, what to cut and the exact email to send, with nothing sent until you say yes

the 3 permission lines it ran on:
→ Always allowed: read spend data, message colleagues, pull reports, update notes
→ Needs your explicit go, every time: anything sent to an outside party
→ Never: sign, buy, subscribe, approve a charge, make a binding commitment
make it report one finding at a time, like this:
→ TODAY: what you pay now, per year, with the source
→ SAVE: realistic savings and how, with a confidence level
→ REC: one recommendation, not a menu
→ NEXT: what it already set in motion
xAI’s bar for a finding. not “renegotiate the CRM (~$50k)”. this: “video tool renews Oct 14. 210 seats, 74 idle for 90 days. drop to 150 = ~$18k/yr. owner confirmed”
no dollar figure, no mechanism, no deadline: it’s a lead, not a finding

add a renewal radar: anything renewing in the next 120 days goes to the top of the queue, because the leverage is at renewal
and before it drafts any email to a vendor, it shows you the plan first: target price, opening ask, walk-away, and what it does if they say no
test it tonight: export one month of card statements (hide the card numbers), paste the 3 lines, ask for findings in that format. tomorrow, count how many have a dollar figure, a mechanism and a deadline, and how many sends are waiting on you
the full system behind it is in the article below ↓

[rari🈀 @0xwhrrari]

https://t.co/GoJ0DWB7Ve

[HD$H @Ryaba_hdsh]

@0xGoldTooth the useful bit is the split of labor: the script does the boring scanning and signal math, you do the walking around the house. that's a pattern worth stealing for a lot of things