Tetriss

I write about whatever sparks my curiosity. A little AI, a little crypto, and a cup of green tea

[Tetriss @Tetriss112]

https://t.co/qyfwCP9mJ6

[Tetriss @Tetriss112]

@twinedon0 awesome😉

[Tetriss @Tetriss112]

@alpha404ai Appreciate it

[Tetriss @Tetriss112]

Opus 5.5. Damn, this is beautiful.
Let's walk through AI architectures: every one in Hugging Face Transformers, in the order it appeared.
1 star = 1 model folder. I pulled the dates from the library's git history and replayed them:
- Nov 2018: one star, BERT
- Mar 2023: Llama lands, after 167 others
- 2025: 124 new architectures
- 2026 so far: 128, more than 2018–2021 combined (88) - Today: 529
The most-edited stars: Whisper, Llama, wav2vec2.

[Claude @claudeai]

Six projects from the last few weeks that caught our attention:

A wildfire analysis prototype that maps the fire’s spread over time and tracks hot spots, made with Opus 5.5.

[Tetriss @Tetriss112]

@twinedon0 And it's only October. 2026 already passed all of 2025 too: 128 vs 124.

[Tetriss @Tetriss112]

@alpha404ai OPUS 🤑😎

[Tetriss @Tetriss112]

@matt__makes yes man🤣

[Tetriss @Tetriss112]

Same wildfire as in this thread, but inside a codebase.

The forest is Bun’s 1,299 Zig files. The fire is the Rust rewrite, done with Claude.

I pulled all 6,755 commits of PR #30412 from git and replayed them hour by hour:
• 4 hours in, half the files are already Rust
• 8.5 hours in: 90%
• peak: 695 commits in a single hour (May 6)
• the longest burn: DevServer, 382 commits
• 9.8 days from first spark to merge
The 64 parallel Claudes are Bun’s claim, git can’t show them. It does show the speed.

[Claude @claudeai]

Six projects from the last few weeks that caught our attention:

A wildfire analysis prototype that maps the fire’s spread over time and tracks hot spots, made with Opus 5.5.

[Tetriss @Tetriss112]

Same wildfire as in this thread, but inside a codebase. The forest is Bun’s 1,299 Zig files. The fire is the Rust rewrite, done with Claude.
I pulled all 6,755 commits of PR #30412 from git and replayed them hour by hour:
• 4 hours in, half the files are already Rust
• 8.5 hours in: 90% • peak: 695 commits in a single hour (May 6)
• the longest burn: DevServer, 382 commits
• 9.8 days from first spark to merge
The 64 parallel Claudes are Bun’s claim, git can’t show them. It does show the speed.

[Tetriss @Tetriss112]

@haldenn7 Agreed, that’s the number that matters. Bun hasn’t published it, and git can’t show it: 6,673 of the 6,755 commits sit under one author, Jarred Sumner, whether he or Claude wrote them

[Tetriss @Tetriss112]

@haldenn7 🤣🤣🤣 amazing

[Tetriss @Tetriss112]

@Lattice_Node Now I have more time to scale my projects and do what I love 🫡

[Tetriss @Tetriss112]

Good call, that kills the lock fight. My rule: one worktree per parallel worker, not per file. So 8, not 300.

Past that it gets messy fast: each worktree needs its own npm install and disk space, and every extra branch is one more merge you have to review.

And if the files don’t overlap, I skip worktrees entirely: workers just edit, one commit at the end.

[Tetriss @Tetriss112]

The "Opus plans, 8 Haiku workers grind" setup is everywhere right now. But one line in it quietly breaks the run: every worker is allowed to git commit.

Eight sessions in one folder means eight processes that all need the Git index at once. There's only one index, guarded by .git/index.lock. Simultaneous commits hit the lock, and git commit -a also sweeps up the neighbors' edits.

How to run the kitchen without a pile-up:
1 - The chef (Opus) writes the menu once: tasks.txt, one line per file.
2 - The cooks (8× Haiku via xargs -P 8) each edit only their own file. No git.
3 - Every cook gets a limit: --max-turns 8 and --max-budget-usd 0.10. Out of turns or budget means FAIL.
4 - The sous-chef (Sonnet) redoes only the FAILs, once.
5 - The waiter makes one commit at the end, after git diff.

One cook: claude -p "$task" --model haiku --max-turns 8 --max-budget-usd 0.10 --allowedTools "Edit" --output-format json

Need a commit per file? Give each worker its own worktree: claude -w.

And don't read total_cost_usd as your bill. It's Claude Code's own estimate.

Eight edit. One commits.

[codila @0xCodila]

REA is most dangerous thing on the internet right now...

It lets your AI agent reverse engineer any app, down to the binary

If you use it correctly, you can rebuild any feature you pay for, for FREE

In this article, I show you how: https://t.co/3P1FNXiXzl

[Tetriss @Tetriss112]

@DN_degen No videos here😥. Scion is Anthropic’s in-house replacement for Git, built because agents now write code faster than Git can handle.

[Tetriss @Tetriss112]

https://t.co/bvUELvXfjm

[Tetriss @Tetriss112]

While everyone’s generating funny clips, we’ve overlooked the fact that Claude + Higgsfield can make video tours of your apartment. A practical use for business, too.

Upload property photos, and Claude prepares the prompt and runs the generation through Higgsfield. The result is a video that looks like a camera flying through the rooms.

Renting out a place on Airbnb? This could be worth trying for social media promotion, without a separate video shoot.

Just check the result against the actual property: AI can invent details and connections between rooms. Generation also uses Higgsfield credits.

[Tetriss @Tetriss112]

@0xMfox Curious to see where this goes next

[Tetriss @Tetriss112]

https://t.co/zjjx6qOQvQ

[Tetriss @Tetriss112]

@StassOneLet The face gets your attention, but the bottle gives the model away. I’d love to see the same scene tested across models-with the bottle briefly hidden, then visible again.