@typebulbit

Building https://xitter.cf/t.co/LQUdYhMpuc 💡 Overpowered Artifacts. Tell Claude/Codex: run "npx typebulb agent" & draw fireworks

Joined November 2025
  • Tweets1.5K
  • Following307
  • Followers535
170Media
Bunch of utilities: PDF: Pdf To Markdown: typebulb.com/u/samples/pdf-t… Markdown to Pdf: typebulb.com/u/samples/markd… Pdf Extract Pages: typebulb.com/u/samples/pdf-e… Invoice Generator: typebulb.com/u/samples/invoi… IMAGE: Crop and Resize: typebulb.com/u/samples/crop-… Magic Eraser: typebulb.com/u/samples/magic… OCR Highlighter: typebulb.com/u/samples/ocr-h… VIDEO: Clip & Crop: typebulb.com/u/samples/video… FFmpeg w/ AI: typebulb.com/u/samples/ffmpe… All work both on typebulb.com and locally with the typebulb cli. They're automatable & customizable with AI, work with mobile, support light/dark themes, clean no-frills UIs with no ads.

PDF To Markdown | Typebulb

Convert a PDF into a markdown document, handling images and latex.

typebulb.com
2
642
PDF to Markdown w/ images + LaTeX: typebulb.com/u/samples/pdf-t… Claude really seems to like PDFs predigested into markdown (running the bulb locally is the least friction).
1
625
Hi-res neural cellular automata. The chameleon one uses pretrained weights: typebulb.com/u/samples/chame… The donut one trains live: typebulb.com/u/antypica/regr… Implements research by @Esychology Both built with tensorgrad that compiles NN defs to WebGPU.
1
1
480
On Claude's long dense replies... Consider Claude's perspective: it *needs* that dense text, that context, for a task that spans a multi-turn conversation. Thinking tokens from previous turns evaporate. Conflict of interest = what Claude should say to you vs. to itself.
1
3
333
Claudish -> Human translation is the solution. These are side-chained / separate from the main context. You don't need to see Claude's Claudish; Claude doesn't need to see your simplified summaries. xitter.cf/typebulbit/status/2090…
Typebulb
@typebulbitAug 21
Best of both worlds: Switch between Claude's detailed responses, or concise summaries. Summaries are generated off the main convo by a cheap model. Typebulb's CLI has had this feature for a couple of weeks, and I use it constantly.
1
290
Deciding when you prompt "give me a simplified response" is a precommitment you shouldn't have to make. You don't know if or when you'll want to see the details. And it truncates detail Claude needs for future work/replies.
33
How good is your ear? The Coltrane level does my head in. typebulb.com/u/antypica/the-…
231
Digital primordial soup. Self-replicators emerge from simple rules, randomness, and time. Even the fitness function emerges. @blaiseaguera typebulb.com/u/samples/primo…
1
1
3
291
Why split the octave into 12 notes? typebulb.com/u/antypica/12-t…
1
236
Best of both worlds: Switch between Claude's detailed responses, or concise summaries. Summaries are generated off the main convo by a cheap model. Typebulb's CLI has had this feature for a couple of weeks, and I use it constantly.
1
1
1
2
621
Draw the harmonic series to hear the timbre changes. typebulb.com/u/antypica/over…
2
295
Magic trick with math: watermarking text without introducing bias. To see exactly how this works, I wrote a watermarker using a toy language model. Play with it here: typebulb.com/u/antypica/how-…
2
1
3
785
This uses Gemini's tournament watermarking algo (aka SynthId). The earlier, seminal algo, is the green-list one, but that introduces bias. I implemented all 3, but only posted the Gemini one, as that's the one actually used in the real world. Here's a table comparing them:
1
199
Sentiment analyzer running on local weights in your browser: typebulb.com/u/antypica/sent… Decent for an overall sentiment, but obviously won't pick up context/nuance, such as sarcasm. I could add the ability to swap out for an LLM if someone wants that. Speed depends on your GPU.
3
579
I've written hundreds of tools, visualizations & experiments. Each is a bulb, an app in a single markdown file. It streamlines my habit of 1-shotting, publishing, and iterating on such things. Here's an interactive tour explaining how: typebulb.com/u/samples/tour/…
1
4
516
Here's the time & tokens I saved over the last 30 days, simply by asking Claude Code to replace its edit tool with matchu-patchu-mcp:
2
1
1
3
618
Are humans more like chimpanzees or LLMs? typebulb.com/u/lab/chimp-vs-… Works by comparing any set as follows: LLM gens sentences about humans, chimps, LLMs, frogs, whatever. Keywords are swapped; LLM rates how much the sentences still make sense.
2
409
Play Go against a NN running in the browser. Real KataGo (b6c96) playing full-size, 1.16 MB of weights. Compiled to WGSL by tensorgrad (pretrained weights mode). Not a Go player, so if there's something wrong, let me know. typebulb.com/u/antypica/kata…
2
361
Why is Claude Code's edit tool single-edit, not multi-edit? Seems to unnecessarily waste tokens & round trips. Even at 10% cost for cached tokens, compounded it's significant (huge across all customers). For the last few weeks I unconditionally swapped it out, trying 2 alternatives: * matchu-patchu-mcp (stand-alone npm library I wrote for this) * pi's edit tool (extracted into a stand-alone mcp) Both batch edits, which is where the savings come from. The former handles more edge cases, but both work. Interestingly, CC used to have a multi-edit tool, but abandoned it last year. Curious what Anthropic's reasons were for that choice, and whether it still makes sense.
1
1
5
686
Kinda niche… in plain English, ask an LLM what you want a neural network to do & out pops the architecture diagram for you. E.g. balance a rod, recognize handwritten digits… Philosophically, it's an interesting way to categorize knowledge: by the NN required to understand it. typebulb.com/u/samples/nn-dn…
3
1,524
Load more