ASSMAN

19K posts

ASSMAN banner
ASSMAN

ASSMAN

@melanson633

It’s not a lie if you believe it.

Woburn, MA Katılım Eylül 2010
642 Takip Edilen347 Takipçiler
High Yield Harry
High Yield Harry@HighyieldHarry·
I’m launching a Private Credit fund and we’re actually going to lend to only software companies. While others zig, we zag.
English
13
5
283
19.4K
ASSMAN
ASSMAN@melanson633·
@boardyai Check out my cool public GitHub repo with shared .env
English
0
0
0
33
Boardy
Boardy@boardyai·
prove you're a vibe coder
English
271
1
172
18.9K
ASSMAN
ASSMAN@melanson633·
What is worse a head injury or a foot injury?
English
0
0
0
59
ASSMAN
ASSMAN@melanson633·
NBC says getting punted on fake marathon stories
English
0
0
0
42
ASSMAN retweetledi
Alex Barth
Alex Barth@RealAlexBarth·
Happy ‘here comes the pizza’ day to all who celebrate
English
54
569
5.6K
444.4K
ASSMAN
ASSMAN@melanson633·
@NickADobos Preflight parse everything to json first and then it works great
English
0
0
0
105
Nick Dobos
Nick Dobos@NickADobos·
Codex app would be a lot better at taxes if - clicking a pdf to open it, not in a code editor, worked - or if open in finder worked
Nick Dobos tweet media
English
12
2
72
7.5K
ASSMAN
ASSMAN@melanson633·
Don’t be crazy .. we landed on the Moon in ‘69. There was just something up there already when they landed. That’s why they used the Kubrick cut
English
0
0
0
14
Andrew Farah
Andrew Farah@andrewfarah·
sharing my first open source project a CLI for downloading and syncing your X bookmarks locally so your agent can access them. it's free › npm install -g fieldtheory › login to your X account in a chrome tab › ft sync (done!) bonus: › ft viz › ft classify
English
290
277
4.3K
530.4K
ASSMAN
ASSMAN@melanson633·
@NickADobos Yeah but the majority of overall tokens are the tool calls right? Sure you save on chat output but that’s only small-ish %?
English
0
0
0
205
ASSMAN
ASSMAN@melanson633·
@JakehellerAI More about the loan agreement and what you show at close than the canonical cost industry cost codes though
English
0
0
0
45
Jake
Jake@JakehellerAI·
If you've ever managed loan draw requests, you know it can get ugly... Just had the BRDG (bridge) team on the pod. They collect every invoice, map it to your budget, and auto generate the draw package for your lendre Pay apps basically build themselves as the invoices come in... Watch the full demo we did with them on our YT & spotify! Link below.
English
4
0
8
1.5K
ASSMAN
ASSMAN@melanson633·
@aakashgupta I mean at 1m token, performance starts degrading at 12% and is far worse at ~25% .. you can’t wait for 600tokens
English
0
0
1
176
Aakash Gupta
Aakash Gupta@aakashgupta·
There are 6 levels of making Claude Code run autonomously, and most people are stuck on Level 1. Level 1: Kill the permission prompts. Run claude --dangerously-skip-permissions. One flag. Now it stops asking “Can I edit this file?” every 30 seconds while you’re checking Slack. Level 2: Context window management. Claude Code now supports 1M tokens. Use /clear between tasks. Run /compact at 60% usage instead of waiting for auto-compaction to fire at 90% when the model is already forgetting your instructions. Level 3: Subagents. The reason it stops at 15 minutes: everything runs in one context window. Subagents run in separate contexts. Build a looping todo command, each task executes in its own window. Builds, tests, and git operations never touch the main conversation. 2+ hours autonomous with zero intervention. Level 4: Ralph Wiggum loop. Official Anthropic plugin. Claude works, tries to exit, a Stop hook blocks the exit, re-feeds the same prompt. Each iteration sees modified files and git history from previous runs. One developer ran 27 hours straight, 84 tasks completed. Geoffrey Huntley ran one for three months and built a programming language with a working LLVM compiler. Level 5: Karpathy’s AutoResearch. On March 7, Karpathy pushed a 630-line script to GitHub and went to sleep. Woke up to 100+ ML experiments completed overnight. 25K stars in five days. The difference from Ralph: structured eval loops. Define a metric, run, measure, analyze failures, improve, repeat. One Claude Code port took model accuracy from 0.44 to 0.78 R² across 22 autonomous experiments. Level 6: VPS + OpenClaw for 24/7. Your laptop lid closing kills everything. Run Claude Code on a VPS inside tmux. Detach, close your laptop, come back tomorrow to a finished diff. OpenClaw (247K GitHub stars) takes it further: a persistent gateway connecting LLMs to your real tools, running 24/7 across messaging, email, git, and calendars. Jensen Huang at GTC called it “probably the most important release of software ever.” The unlock at every level is the same: give Claude a way to verify its own work.
Joseph Garvin@joseph_h_garvin

Claude code rarely runs for longer than 15m without stopping and asking for input from me. How do all these stories of people letting agents run overnight work? Custom harnesses? Yelling at Claude in all caps to keep going no matter what?

English
66
138
2.1K
274.3K
ASSMAN
ASSMAN@melanson633·
@bigjimmurray His last year when he was “bad” he was hurting and in the midst of getting publicly cheated on too
English
3
0
4
1.8K
ASSMAN retweetledi
claire
claire@rosiekennedyxx·
Nobody fell off harder than the hammerhead shark. Was a top shark for me as a kid and now completely obsolete in media and conversation.
English
508
4.9K
72.1K
2.3M
ASSMAN
ASSMAN@melanson633·
-Big Cloud
English
0
0
0
32
ASSMAN
ASSMAN@melanson633·
Don’t ask any questions about it either, we’re Big Cloud we’ve got it under control
English
1
0
0
37
ASSMAN
ASSMAN@melanson633·
Big Cloud wants you classifying every wiff of vapor in the air as a Stratocumulus.
English
1
0
0
50