Day 7 of my 30-day local AI challenge: my Spanish speech model, trained on my Mac, is getting closer to Whisper. One change this round: I fine-tuned the audio encoder from layer 3 up with LoRA, which adds 4.87M trainable params. Same data as before, 8000 steps. On real meetings, hand-corrected, WER went 28.1% → 27.1%. Whisper large-v3-turbo gets 17.1%. On spontaneous speech (Common Voice) it went 17.9% → 16.7%. Whisper gets 10.1%. On volunteer recordings it's 9.7% vs Whisper's 9.4%, only 0.3 points behind. The lesson: the checkpoint with the best validation WER wasn't the best model. Step 3500 won on validation but was worse on real meetings, 28.3%. I kept step 8000. Whisper is still ahead on every test. The gap is closing, though. Devs: do you pick checkpoints by validation loss or by a held-out real-world set?

Prompt: Go online and research all the things that have happened in information security in the past 5 days. Today is October 24, 2026. Then fix any security issues everywhere in the server. Ask me questions if you need to

IPO? 👀

PHP Gitlab Node.js Oracle Railway Cloudflare

Should my new project use PHP or Node?

I ported it to @Claude Code and made it open source. And it works for PHP and @nodejs too. @OpenAI built a security skill for Codex. Your website shouldn't be less secure just because you use a different coding agent. Here's the repo 👇 github.com/masteread/php2she…

Hola! las personas que tienen equipos que construyen o utilizan PHP y node.JS Se detectaron brechas importantes de seguridad a nivel global en ambos lenguajes. Comparto para que puedan comparitr con equipos técnicos y los puedan blindar: nodejs.org/en/blog/vulnerabi… php.net/archive/2026.php#202…

Valio Madre. Hoy nació el COVID de internet

doesn't work in spanish

So new SEO is about "gaslighting" LLMS with articles, backl inking and videos? 🤔

I know the world use English as main language but when I see every day a new release like this one I can’t help thinking that Spanish / Hispanic developers we also should be doing this

Day 6, and funnily enough we're on v6 of my Spanish speech-to-text model. Lesson of the day: public benchmarks lie a little. On standard test sets, v6 is now close to Whisper: • Lab recordings: 10.1% vs 9.4% WER • Common Voice LatAm: 8.5% vs 7.8% • FLEURS: 9.9% vs 6.6% But on real meetings, hand-corrected, the gap is still big: 33% vs 17%. Real audio is messy: crosstalk, background noise, people trailing off. The biggest win came from real data: on real calls voted on by the community, error dropped from 29.1% (v4) to 20.1% (v6). Training on what people actually sound like beats squeezing another decimal out of clean audiobooks. If you speak Spanish, you can help: record a few phrases here 👉 escucho.ai/labs/training. It takes under a minute and it's anonymous. Devs: what's your best trick for closing the gap between benchmark WER and real-world WER?

A banana 🍌 . Gravity was the timer.

Day 5 of my 30-day local AI challenge: my Spanish speech-to-text model went from 57.6% → 24.6% → 10.9% WER in 3 versions. Still building on @TslShahir's EG2-Transcribe idea: a frozen EmbeddingGemma 2 turns audio into vectors, and a small 86M-param reader picks one whole word every 40 ms. Only the reader is trained. Where it stands vs Whisper: • FLEURS Spanish: 10.9% vs 6.6% • Lab recordings: 10.8% vs 9.4% • Common Voice LatAm: 10.2% vs 7.8% • Spontaneous Spanish: 19.5% vs 10.1% • Speed: ~80× real time vs ~20× Fair-test notes: • v2, v3 and Whisper were measured with the same rules, same audio, and voices the reader never heard in training. • v1 used different rules, so it's only a rough comparison. The gap is clearly spontaneous speech: fillers, false starts, people talking over each other. Devs, I'd love your take: 1. How would you close the spontaneous-speech gap? More conversational data, or a different decoding strategy? 2. Whole-word dictionary (~307k words) vs subwords for Spanish: worth switching? 3. What Spanish test set should I add next? Drop your ideas below 👇 I'll test the best ones and credit you.

Day 4: first results! Inspired by @TslShahir's EG2-Transcribe (built on EmbeddingGemma 2), my Spanish ASR went from 57% to ~24% WER, and it runs 81× real time vs Whisper's 20×. Whisper still wins on accuracy (6.6%). Long way to go. More soon. 🧵

Claude use claude to build claude