Pinned Tweet
XTTS is still being downloaded almost 5m times every month and 2.1m only on HF.
It is greater than many recent hyped models.
Hope people use it well for that to worth to my burnout that I’m still recovering from.
Coqui has been one of the most successful broke startups
New BlaGPT best: 3.1153 val loss, with Latent Lattice.
One small recurrent latent (width 256) runs across layers 2-7. At each layer it steers q/k/v and the MLP through FiLM.
Paper: arxiv.org/abs/2611.21608
Code + results: github.com/erogol/BlaGPT
Added kale-tts-1-nano (7.5M) to the small CPU TTS bench.
On 1 CPU thread:
- 37x faster than real time
- 53 ms to first audio
- WER 1.6%, UTMOS 4.37
9x faster than Phonon 31M at the same WER.
erogol.com/small_tts_bench/
erogol retweeted
doesn't really matter how good gemini is
the fact that @google actively bans gemini/antigravity subscription accounts for using other coding tools (i.e @opencode, cliproxyapi, pi, etc), forcing end users to pay API prices, is a complete non-starter
do better please
Gemini 4 Neon: frontier-level intelligence with fast end-to-end task completion across complex software engineering, knowledge work, and science workflows.
Begins rolling out today in the Gemini API in @GoogleAIStudio, Gemini for business, for paid Google AI Pro and Ultra subscribers in Google @Antigravity, and for Google AI Pro and Ultra subscribers in the @GeminiApp.
erogol retweeted
Modern problems require modern solutions.
有人发现了一种非常离谱的 JavaScript 混淆技术:通过在代码里塞入大量生物武器、毒药和致死剂量相关的函数名,让 AI 因为安全限制而拒绝分析代码。
安全研究员 veritas 在分析浏览器指纹脚本时,发现一段来自 ServicePipe 相关域名的 JavaScript,里面充斥着极其诡异的函数名:
computeRicinLethalDosageForWaterSupply
calculateDeadlyLethalSuicideAspirinDosage
光看名字,仿佛在研究生化武器。但实际上,这段代码只是在收集浏览器指纹,包括 User-Agent、GPU、Canvas 等信息。
研究员怀疑,这些名字是开发者故意设置的陷阱,专门用来阻止 AI 辅助逆向工程。
于是他做了个实验:
• Astra:拒绝分析
• Astra + Daybreak Blue:仍然拒绝
• GLM-5.3:成功分析
这就有意思了。
传统代码混淆是通过复杂控制流、加密字符串、无意义变量名,让人类和工具难以理解代码。
而这种新思路,直接瞄准了大模型的安全机制。
代码本身可能完全无害,但只要在函数名里塞进足够多的危险词汇,部分 AI 就可能主动放弃分析。
更讽刺的是,模型越严格遵守某些安全规则,面对这种混淆时反而可能越容易受到干扰。
当然,目前还没有证据证明开发者一定是故意这样设计,也不能说明这种方法对所有模型都有效。
但它揭示了一个值得关注的趋势:随着越来越多的逆向工程依赖 AI,未来的代码混淆技术可能不再只针对编译器、反编译器和人类研究员,还会专门针对大模型的安全判断机制。
以前是让代码难以理解,现在甚至可以尝试让 AI 不敢理解。
Small CPU TTS bench updates
- Now with UTMOS
- PiVo-1.4M: 5 ms to first audio, 9x real time
- Inflect-Micro-v3: 11x real time, 0.8% WER
- Supertonic 3: UTMOS 4.47
Kokoro-82M missed real time.
erogol.com/small_tts_bench/
A 4-min explainer on "Weight-Space Programming":
• MIT's Sparks of WSP (GPT-6 Astra)
• @roicohen9: one paragraph in, a Qwen weight update out
• @HyunjongOk: sparse tables, ~1,800× fewer params
(Narrated by Phonon 31M )
erogol retweeted
1/ We still don’t have a simple picture of how even 2-layer MLPs learn features. In a new paper with @berkott3 and @learning_mech, we propose a simple, solvable model of shallow MLP training, and find that real MLPs roughly follow it. 🧵
erogol retweeted
⚠️ Numerical issues can go on silently for decades like in the softmax/CE! Similar issues are present in other widely-used divergences like the KL and JSD
I just posted this paper overviewing what they are and how to fix them:
📚 arxiv.org/abs/2611.18511
erogol retweeted
🚀 Stop using Transformers for weight-space programming
If you are not training with gradients on a GPU, why default to a Transformer?
When an LLM writes network weights directly, letting it also design the architecture achieves:
📉 ~1,800× fewer parameters at matched accuracy on ARC (AI2 Reasoning Challenge), compared to the best Transformer the same LLM wrote.
⚡ Up to 17× faster GPU inference with a memory architecture built for these tables, compared to a Transformer of the same size.
🧵 Paper & code at the end 👇
Added tiny-voca to my small CPU TTS benchmark.
4.5M params, Apache-2.0. On 1 CPU thread:
- 117 ms to first audio
- 37x real time
- 3.3% WER
- 666 MB memory
2nd fastest of 13 models. Phonon 31M stays my pick for live voice.
Hear all: erogol.com/small_tts_bench/
Small CPU TTS Benchmark
12 tiny TTS models on 1 CPU thread, 60 audio samples.
My Pick -> Phonon 31M @GradiumAI. 82 ms to ttfa, 4x rtf, 1.6% WER, 363 MB, Apache-2.0.
Pocket TTS, same box: 286 ms, 1.2x, 1347 MB.
Let me know any models i missed
erogol.com/small_tts_bench/
Here is a short video about "Decode Before You Hear" paper that speeds up chunk based speech to speech models significantly
arxiv.org/abs/2611.06982
Machine Learns - 81 🤖
• Weight-Space Programming: a closed LLM edits raw weights
• WaveGen: TTS with no codec or vocoder
• Decode Before You Hear
• Latent Sparse Attention
• Agentic Backdoors
• Mistral Large 4
• My small CPU TTS benchmark
erogol.substack.com/p/machin…
6
404
New best val loss in BlaGPT (3.1106):
- Aurora optimizer beat Muon (-0.0294 loss)
- RoPE + RoVE beat GRAPE-A with 0 extra params & faster steps
- GatedNorm + Engram x20 stacked cleanly
Code & logs: github.com/erogol/BlaGPT
4
248
no gold ? he'd love that !
Cumhurbaşkanımız Sayın @RTErdogan, Apple Yönetim Kurulu Başkanı Tim Cook'u Cumhurbaşkanlığı Külliyesinde kabul etti.
161
tried of telling the same things to AI agents
so I built "kechup" a cli task manager that lives in the repo and git
- each claim, note and close is a git commit, agent ID as author
- a task closes only when its REPORT.md is committed
- git revert of done work reopens the task
My agents tracked 190+ tasks in 14+ projects with it in 3 weeks.
Reduced token usage by ~30% by
- no need for a context for onboarding
- no false "done" flags without evidence
- easy switch between harnesses
github.com/erogol/kechup
4
46























