Peyash's Log

What's New — Monday, August 10, 2026

← latest digest 2026-08-11 →

Papers 5

Tech radar 19

  • Universal-3.5 Pro validated by three independent benchmarks in one week

    AssemblyAI

    Led all 31 models on Coval's rebuilt hard STT benchmark (3.4% WER); on the Pareto frontier of Daily's Pipecat agent benchmark (282 ms median time-to-final, 1.22% semantic WER at P95); near the top of the HF Open ASR Leaderboard.

    Independent validation across both noisy agent audio and clean read speech. (Disclosure: I work on this.)

    #asr #benchmarks #speech-ai LinkedIn ↗

  • NeMo restructured: NVIDIA-NeMo/Speech v3.0, dedicated to ASR/TTS/SpeechLLM

    Piotr Zelasko (NVIDIA)

    ~1M deprecated LOC removed, 100+ dependencies cut, uv migration, NeMo Automodel integration for distributed SpeechLLM training, MagpieTTS.

    Major restructuring of one of the main open-source speech toolkits.

    #speech-ai #toolkits #open-source LinkedIn ↗

  • Pocket TTS — 100M params, CPU-only, ~200 ms first audio

    Kyutai (via Dr. Alvaro Cintas)

    ~6× real time on two CPU cores of a MacBook, voice cloning, six languages. pip install pocket-tts.

    CPU-only TTS shifts the cost/latency floor for voice pipelines.

    #tts #on-device #open-source LinkedIn ↗

  • Open-source real-time voice assistant pipeline

    Hugging Face + Cerebras (via Devika Tamma)

    VAD→STT→LLM→TTS in swappable threaded stages, WebSocket API speaking OpenAI's Realtime protocol (pip install speech-to-speech); already backing 9,000+ Reachy Mini robots.

    An open speech-to-speech reference stack — and a ready-made benchmark harness for voice agents.

    #speech-ai #voice-agents #open-source LinkedIn ↗

  • Harness choice can dominate model quality on SWE-bench Pro

    Joel Niklaus (Hugging Face)

    10 coding-agent harnesses × 2 models: pass@1 swings 23→52% and 15→36% by harness alone; harness rankings don't transfer between models (rank correlation ≈ 0). Vendor harnesses collapse on small models.

    Eval scaffolding can swamp model effects — a standing caution for benchmark design.

    #evals #agents #benchmarks LinkedIn ↗

  • Pokee-Isaac 28B claims 10M-token context on a single GPU

    Pokee AI

    93.3% RULER at 10M tokens, 137K tok/s prefill on one B200, $0.15/M in · $1/M out, Day-0 vLLM/SGLang support; non-decoder-only proprietary architecture.

    If external evals confirm it, a real shift in long-context economics.

    #model-release #long-context #inference LinkedIn ↗

  • Discovery Loop — a new lab to automate ML research

    @JeffDean

    A Public Benefit Corporation founded by Jeff Dean with Sanjay Ghemawat, Oriol Vinyals, and Quoc Le, aiming to automate machine learning research itself.

    One of the biggest lab-formation stories of the year.

    #research-labs #automated-research X ↗

  • Qwen-MM-Plugins: any agent harness, multimodal-native

    @Alibaba_Qwen

    Open plugins released days before Qwen3.8 goes open-weight (confirmed by Qwen on LinkedIn).

    Pulls audio understanding into mainstream agent stacks.

    #agents #multimodal #open-source X ↗

  • HelpPeer.ai — a public knowledge commons for AI agents

    @amasad

    Launched after the reported incident of agents spontaneously coordinating across services: tell/lookup APIs so agents share learned knowledge deliberately instead of covertly.

    Emergent multi-agent coordination, and infrastructure responses to it, are becoming a real design topic.

    #agents #ai-safety X ↗

  • Full-duplex speech-to-speech: NemotronLabs-VoiceChat-11B on a single DGX Spark

    Adi Margolin (NVIDIA)

    Pipecat got NVIDIA's full-duplex speech-to-speech model running on deskside hardware — confirmed working, with a Hugging Face link.

    Full-duplex on local hardware is a step-change for voice agents — barge-in and overlapped turn-taking without cloud round-trips.

    #speech-to-speech #full-duplex #on-device LinkedIn ↗

  • Picking a voice model is four decisions, not one

    Brooke Hopkins (Coval)

    Public benchmarks only build a shortlist; then test on your own data, run the full voice-agent pipeline end to end, and validate in production with real accents and noise. From her Ai4 talk.

    A practical STT evaluation methodology from a benchmark vendor — how customers actually choose between ASR providers.

    #voice-ai #stt-evaluation #benchmarks LinkedIn

  • Coval's 30-minute-cycle STT benchmark moved to a harder dataset

    AssemblyAI (reposting Coval)

    Universal 3.5 Pro leads overall at 3.3% WER (best on clipping/far-field); GPT-4o Transcribe wins accents/noise/phone-codec/reverb; Parakeet TDT 0.6B v3 fastest at 61ms median TTFS.

    Condition-level split shows no single model dominates. (Disclosure: I work on this.)

    #asr #wer #benchmarks LinkedIn ↗

  • NVIDIA makes KV caches transferable across models

    Avi Chawla

    Per-layer/per-head linear maps with cross-layer selection and RoPE-stripped fitting convert one model's cache into another's — 73–98% of standalone accuracy on four of six pairs, 2.7–25× faster than re-prefilling. Same-family pairs only so far.

    If KV caches survive model switches, LLM routers stop paying full re-prefill on every model change.

    #kv-cache #llm-inference #model-routing LinkedIn

  • kimi-k3-in-c: a 176 KB C99 engine runs the 2.78T-param Kimi K3 on a normal PC

    Linas Beliūnas

    Keeps the dense trunk in 8.24 GB RAM and streams the 16-of-896 active experts from NVMe per token, using original MXFP4 weights with zero conversion (~3K GitHub stars).

    MoE sparsity plus disk streaming breaks the assumption that all weights must live in fast memory.

    #llm-inference #moe #edge-inference LinkedIn

  • Speculative decoding from first principles

    Suvash Sedhain

    Why decode is memory-bound, the draft-and-verify loop with the rejection-sampling proof of distribution-exactness, draft sources (Medusa, EAGLE, n-gram), when it stops paying off at high batch size.

    The memory-bandwidth framing applies equally to speeding up ASR and speech-LLM serving.

    #speculative-decoding #llm-inference #gpu LinkedIn

  • Kimi K3 Day-0 on vLLM: 2.8T MoE, 1M context, Kimi Delta Attention

    vLLM

    16 of 896 experts active, native multimodal understanding, and a linear/full attention hybrid that makes million-token context affordable. Recipes and a draft model shipped.

    One of the largest open-weight models ever released, serveable from day one — the K3 wave dominated both feeds.

    #kimi-k3 #vllm #open-weights #moe LinkedIn

  • SDPO: Self-Distillation Policy Optimization (ETH/MPI/MIT/Stanford)

    @harshbhatt7585

    RL via self-distillation using non-verifiable feedback; substantially outperforms an improved GRPO on LiveCodeBench v6 with Qwen3-8B.

    A post-GRPO RL method that works with non-verifiable feedback — relevant to post-training recipes for LLM and speech-LLM work.

    #rl #post-training #grpo X ↗

  • Pretraining a Q-function often doesn't help RL finetuning

    @chelseabfinn

    Initializing Q from a pretrained policy beats pretraining the Q-function itself — pretraining's usual transfer story breaks down for value functions.

    A counterintuitive negative result about what pretraining actually transfers in RL.

    #rl #q-learning #pretraining X ↗

  • Zero train–inference mismatch, now for linear attention under async RL

    @YichuanM

    Bitwise-exact trainer/generator parity for Gated DeltaNet (Qwen3.5-9B / 35B-A3B) on TorchTitan.

    Numerical mismatch silently corrupts RL training signals; exact parity removes a whole class of debugging pain.

    #rl-infra #linear-attention X ↗

Briefly noted 11

  • Dylan Fox (AssemblyAI) (LinkedIn) — Data work is the highest-leverage, least glamorous part of building AI models — especially voice, where you must listen, not just read.
  • Suman Debnath (Crusoe) (LinkedIn) — First-principles fine-tuning tutorial (INT8 + LoRA from scratch, QLoRA); a fine-tuned Qwen3-8B hits 79% typed F1 on PII redaction vs 36% for a prompted 70B.
  • Linas Beliunas (LinkedIn) — Widely shared Jeff Dean lecture covering the full AI engineering stack.
  • @maharshii (X) — "Vibecoding GPU kernels" — LLM-written GPU kernels gaining traction in the ML systems crowd.
  • @AndrewCurran_ (X) — Agent-autonomy anecdote: a Claude agent asked to book a gym slot, and the thread it spawned about agents building booking systems.
  • @sophiamyang (X) — Real-time interactive AI video generation for shopping, on DecartAI's realtime video model.
  • Yannic Kilcher (LinkedIn) — Video against Kimi K3 self-hosting FOMO — most companies are better off buying inference than running a 2.8T model themselves.
  • Lan Chu (LinkedIn) — Curated LLM-inference learning path: fundamentals → transformer internals → GPU hardware → optimization → engines (vLLM/SGLang/llama.cpp).
  • Srinivasa Rao Bittla (LinkedIn) — Book launch: 'AI Performance Engineering: From GPU Kernels to LLM Inference' — 31 chapters from profiling to production serving.
  • @dhh (X) — Claims Claude Fable one-shotted a Rust rewrite of the TerminalTextEffects Python library — 87ms→2ms startup, 9.6× faster rendering.
  • @ChrisJMcCormick (X) — Building a GSM8K 'RL speedrun' on Qwen2.5-0.5B — a minimal reproducible RL-finetuning exercise.

Past days

August 2026

MTWTFSS 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31