Doing ML Web - https://teapaw.org/t.co/yxKAwSSkgR Substack - https://teapaw.org/t.co/W9Qg4M3AZg
Joined October 2008
- Tweets8.7K
- Following580
- Followers1.5K
- Likes6K
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 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
3
214
no gold ? he'd love that !
Cumhurbaşkanımız Sayın @RTErdogan, Apple Yönetim Kurulu Başkanı Tim Cook'u Cumhurbaşkanlığı Külliyesinde kabul etti.
146
The Kimi K3.1 Pro is a beast!
It’s enough for 99.9% of AI work in a company and there’s no reason to pay premium for Claude or OAI.
I’ve tried to find any shortcomings but I can’t seem to find any.
2
5
555
Gemini iOS app is embarrassingly bad.
I tried to use it after a long time.
Got 3 connection errors after a long wait
Now I reached my limit and can’t ask anything
200
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
43
Die KI-Steuer kam noch vor der KI selbst.
Jetzt wird es richtig lächerlich: Die Koalition debattiert eine KI-Steuer
Vor zwei Wochen brachte Digitalminister Wildberger im Tagesspiegel die Besteuerung von KI-Umsätzen ins Spiel, um den Umbau des Arbeitsmarkts zu finanzieren. Kurz darauf ruderte er bei RTL zurück, erst müsse Deutschland die Chancen der Technologie nutzen.
Gestern hat CDU-Vize Karin Prien den Vorstoß wiederbelebt. Im Stern: Wenn Wertschöpfung im Wesentlichen über Künstliche Intelligenz funktioniere, müsse das auch besteuert werden, anders gehe es nicht.
Immerhin: Merz steht der Sache kritisch gegenüber. Beschlüsse oder Gesetzentwürfe gibt es keine.
Was es auch nicht gibt: eine Definition.
Was genau ist ein KI-Umsatz?
Der eines Unternehmens, das ein Modell trainiert?
Der eines, das eines nutzt?
Jede Software enthält heute KI-Komponenten. Entweder trifft die Steuer irgendwann jeden Softwareumsatz, oder sie trifft genau die vier US-Konzerne, die hier ohnehin das meiste Geschäft machen.
Und das ist der eigentliche Punkt. Deutschland hat keine Frontier-Modelle, keine nennenswerte Rechenkapazität und rund 5 Prozent Anteil an der globalen KI-Infrastruktur. Diskutiert wird gerade die Verteilung einer Wertschöpfung, die hier niemand erzeugt.
Man beschließt die Ernte, bevor man gesät hat. Mit dem Wissen, dass es am Ende ohnehin nichts zu essen gibt.
117
erogol retweeted
We have discovered a substantial number of vulnerabilities in the German legal system using LLMs.
These may enable access to confidential documents, travel without an otherwise required visa, enforceable leave for soldiers at NATO’s JSEC, and more.
Similar to computer systems, the written law can contain vulnerabilities that invite exploitation. In our blog post, we share a selection of vulnerabilities, and present potential exploits for each. To our knowledge, none of these have been previously recorded. We share only those vulnerabilities whose exploitation we believe would not cause serious harm.
blog post: felixringe.com/legal-exploit…
Machine Learns - 80 🤖
• Clover: block-diffusion coding LLM
• KronLSTM: linear attention with a convergence proof
• RubatoCodec: adaptive speech tokens
• Speculative Forcing
• MAI-Base-2 pre-training report
• Normalizing Flow Maps
• kechup (mine)
erogol.substack.com/p/machin…
Ming-3.0-tts: open weights, MIT. #1 overall on Artificial Analysis Pronunciation Robustness.
LM → DiT flow head → continuous 12.5Hz latents. No discrete tokens. 44.1kHz.
0.5B dense + 7.9B MoE (1.3B active). 7 languages. 14 paralinguistic tags.
github.com/inclusionAI/Ming-…
Fish Speech Drama 3 deep dive: 35.9B open-weight MoE TTS, ~3B active. 7.8x the size of Fish S2, yet faster: 271 vs 177 frames/s on one H200.
Why: 6.0 GB of weight reads per frame vs 15.6 GB. 13 serial layer passes after the backbone vs 40.
erogol.substack.com/p/fish-a…
erogol retweeted
🪗 Your Transformer Is Secretly Elastic 🧵👇
Transformer families come in a few fixed sizes, so you commit to one for your hardware and latency constraints, even as query loads vary. What if a single pre-trained model could serve every budget?
We learn one importance ranking over every MLP neuron and attention head, once, with frozen weights. At inference time, the model can then be pruned to any computational budget by dropping its least important units, without retraining.
Across 27 vision and text encoders spanning 22M to 8B parameters, performance degrades smoothly up to 60% sparsity. For example, pruning DINOv3 ViT-H+/16 to 50% sparsity removes 427M parameters, for only a 0.8% drop in ImageNet-1k linear probing.
Super fun project in collaboration with Dheeraj Varghese (@dhevarghese), supervised by Gertjan J. Burghouts (@gjburghouts) and Cees Snoek (@cgmsnoek)
🌐 Project page: secretly-elastic.ashita.nl
Models are terrible at planning.
Ask a model to plan something you don't fully understand, and it over-specs every time.
Adding more models on top makes it worse.
The only thing that works: you write the plan. The model follows it exactly.
1
233
Machine Learns - 79 🤖
• Diffuplex: diffusion LM, full-duplex speech
• Audio token attention, predicted before the LM
• UltraBERT: encoders saturate at 1-2T tokens
• Looped DiT
• Level-of-Token Diffusion
• VISTA
• EmbeddingGemma 2, Clef, PixelUMM
erogol.substack.com/p/machin…
Drama 3: open-weight 35B MoE TTS. 16 speakers, full scenes in one pass.
3B active, hybrid linear attention. Codebook-MTP decoder, 10 causal codebooks. Voice cloning, emotion tags, edit-as-continuation.
89ms TTFA. 71 languages. SGLang-Omni day one.
huggingface.co/fishaudio/dra…
erogol retweeted
🧭 New model: Typed Decision, built on EmbeddingGemma 2
Give it a state + any schema of questions → typed, calibrated answers (choice · score · noul). No text generation, no output parsing.
76.4% on typed-decisions, ~6 pts behind a 27B LLM with ~100× fewer params.
🤗 huggingface.co/mrm8488/embed…
