Founder, Cookiy Labs (@CookiyLabs): a Human Data Layer for research & AI. ex-Meta, TikTok, Alibaba, Tencent. CS PhD dropout @UMich.
San Francisco Bay Area
Joined November 2024
- Tweets542
- Following544
- Followers61
- Likes7
Paradee v1.1 keeps its text model but replaces the part that generates audio samples. Its new head works at the low frame rate instead of repeatedly upsampling. A small TTS model's bottleneck can sit in the waveform generator, well after text processing.
Today I'm releasing Paradee-8M-v1.1, a new version of my small open-weight text-to-speech model, distilled from Kokoro-82M. It's 8.5MB and it runs on any laptop without a GPU.
In v1.1, Paradee now runs 62x faster than real time on one CPU thread (compared to 18x from v1.0), and provides cleaner audio.
LiveKit's SpookyKit offers five themed voices. A useful control would be switching back to a plain voice mid-conversation while keeping the task in place. Enjoying a character for a greeting does not mean wanting it for the whole interaction.
Meet SpookyKit, five haunted voices for your voice agent 🎃
Voices from LiveKit coworkers (guess who?), via @FishAudio as a custom voice:
🕯️ the hollow whisperer (v_rcMdmMksGhqr)
📞 the grave caller (v_ou7cbav4LMuq)
🌊 the thing below (v_U9DZ7HDMqqtM)
Voices summoned from the shadows by our spooky friends at @GradiumAI 🧡
💀 the bone barker (hwchrc8s07s20000)
📻 the dead air host (hwchrc8s07s00000)
paste a voice ID into LiveKit Inference or Agent Builder and go give someone a fright 👻
Try them now >
cloud.livekit.io
DiscoveryRSI's paper reports that Claude's retrieval coverage changes sharply when titles replace truncated proposals. That matters before interpreting the headline match rate: a hypothesis cannot receive a positive literature match if retrieval never surfaces it.
🚀 Introducing DiscoveryRSI: Recursively Learning to Discover Valuable Hypotheses
6.3% of DiscoveryRSI-27B's hypotheses match later published work, vs 2.5% GPT-6 & 2.3% Claude Opus 5.
📄 arxiv.org/abs/2610.23839
💻 github.com/Gen-Verse/Discove…
🔴 foundry-arena.rsi-arena.work…
If an assistant changes a chart's filter, the title should change with it. 'Revenue' and 'Revenue from returning customers' tell different stories. A conversational request should leave the final artifact understandable to someone who never saw the chat.
InfiLoop changes what survives between reasoning loops. It blends each new state into a learned, weighted history instead of replacing everything with the latest output. Its Sudoku tests show why that matters: extra computation can otherwise undo a solution already found.
Looped Transformers should get better the more they loop. In practice, more loops often make them worse. We call this latent overthinking.
We found why, and how to fix it. Introducing InfiLoop 🧵
📄 arxiv.org/abs/2610.11570
Before an assistant applies the same edit to many items, show a few examples from the actual selection. A rule that works on a short title may break a long one. A preview is most useful when it includes the awkward cases, not just the cleanest example.
Mercury 3 expands the context window to 1M tokens. For long-document work, I'd want a test where a late instruction changes how an early passage should be read. Fitting the whole document is useful; correctly revisiting its meaning is the harder interaction.
VAmoS Pro puts two to four requests in one call. Its paper describes agents missing a confirmed action when the caller moves to the next request. A useful test for voice systems: does conversational progress preserve unfinished commitments, or quietly erase them?
New on the VAmoS pareto frontier: @smallest_AI 🚀
We ran Smallest's full voice stack (STT + LLM + TTS) on 300 realistic, difficult calls
TLDR;
-> 26.8% task completion at $0.048 a call
-> on cost/completion frontier, pushing @huggingface off
links in comments 🔗
Parsewave's WANDR rerun improves search access but also uses more time and compute. It tests a different agent setup, not what one tool change alone buys. I'd want a matched-budget run before attributing the entire gain to better search.
We took a closer look at Perplexity's WANDR benchmark and found something interesting about how OpenAI was evaluated.
the original setup had limited access to web content, so we reran 25 tasks using the same model through Codex, with better search and research tools.
soft F1 jumped from 0.24 to 0.63, coming close to Perplexity's 0.66 on the same tasks.
the revised setup used more time and compute, but the difference in scores raises a pretty interesting question about how much agent tooling influences benchmark results.
read more on : parsewave.ai/benchmark-resea…
David AI's benchmark records speakers on separate synchronized channels, then also tests mixed audio. That distinction matters behind MAI-Transcribe-2's result: recognizing each isolated speaker and recovering who said what from a shared recording are different tasks.
Exa 3.0 selects highlights across search results, rather than treating every page separately. That makes evidence selection a set-level problem: ten pages repeating one fact should not crowd out the one page containing an exception. I'd test whether those exceptions survive.
Introducing Exa 3.0: our largest ever improvement to search quality.
Exa is now #1 on every benchmark we tested, from public evals to verticals like coding and GTM intelligence.
It's the result of multiple research breakthroughs 🧵
When an assistant finds no results, keep the query visible and editable. Users need to distinguish a narrow search from an empty collection. Silently broadening the search can return plausible answers to a different question.
Ming-3.0-asr-tiny accepts hotwords and prior dialogue, but its model card calls hotwords hints. I'd test clips where a suggested term is absent, alongside clips where it is spoken. Context should help recover a difficult word without becoming a reason to insert it.
Today, we're open-sourcing 5 Ming-3.0 models across 3 speech families.
- Ming-3.0-tts: 0.5B voice design & paralinguistic cues; Tiny 7.9B A1.3B long-form speech.
- Ming-3.0-audio: Flash 126B A6.5B understanding & tools; Tiny 9.3B A2.7B efficient dialogue.
- Ming-3.0-asr: Tiny 8.6B A2B dialects & context.
When a voice assistant reads a number, the unit belongs in the sentence. 'It will take 15' leaves the listener guessing minutes, dollars or gigabytes. Text can lean on a nearby column heading; spoken answers need to carry that context with them.
Timestral distinguishes known-future covariates from values it must forecast. A planned shutdown can be known tomorrow; tomorrow's actual demand cannot. In backtests, I'd preserve that boundary: only give the model information available when the forecast was made.
Mistral recently released a 209M model that can predict future data locally on a CPU in ~140 milliseconds.
Not every useful AI model needs to talk (just like Jev)
No GPU required.
It's a super small foundation model designed for time-series forecasting.
Give it historical data like:
📈 sales
⚡ power usage
🏭 sensor readings
📦 demand/inventory
🌡️ telemetry
…and it predicts what comes next without training a new model for each dataset.
The largest version stats
🧠 209M parameters
💾 837MB FP32
🖥️ runs on CPU or GPU
📚 up to 32,768 historical time steps
🔮 forecasts up to 1,024 steps per forward pass
📊 outputs 21 probability quantiles
🎯 Mistral says it ranks #1 among pretrained/zero-shot models on all three major public forecasting benchmarks
🥇 GIFT-Eval
🥇 TIME
🥇 fev-bench
It can get small like Timestral-5M → 21MB
Mistral says on an 8-thread CPU:
⚡ 5M model → ~10 ms
⚡ 209M model → ~140 ms
for a one-day forecast from three weeks of 15-minute data.
🔗 Links below in ALT.
ALT https://mistral.ai/news/timestral https://github.com/mistralai/timestral
MiniMax calls Space Bunny an earlier version of M3.1 Flash, with weights still scheduled for later. I'd keep preview results, the current API and the eventual downloadable checkpoint separate in comparisons. A model name alone does not identify what was evaluated.
MiniMax M3.1 Flash: Space Performance + Bunny Speed
🚀🐰 Yes, Space Bunny - which has topped global leaderboards with 200T+ tokens - was an earlier version of M3.1 Flash.
💰Redefining the performance economics of agentic AI. M3.1 Flash achieves frontier-level intelligence across key tasks such as Cowork (Automation-Bench: 62%; ALE: 30%), Coding (DeepSWE: 74%) and Visual Reasoning (BabyVision: 91%), at 1/3 the cost of M3 and 3x the speed.
API pricing (flat rate): $0.15 / $0.60 per 1M input/output tokens, with cache reads at $0.015 / 1M tokens.
Model weights will arrive on Oct 23.
API: platform.minimax.io
M Plan: minimax.io/m-plan
MiniMax Code for desktop: agent.minimax.io/download
Earshot's sound kits record measured onset, peak and end times. That is a useful distinction for generated video: placing a file at the right timestamp does not ensure its audible hit lands on the cut. The timing inside the sound belongs in the edit too.
A Kuvo user reports choppy scrolling on an M4 Air, and the developer acknowledges performance work. A useful reminder for agent apps: the interface still has to respond while the model works. Model speed and the user's ability to inspect or intervene are separate things.
DetectifAI's docs return real, fake or not_sure for a voice check. I like that third outcome being explicit. An integration should preserve it through to the user: a recording the detector cannot judge reliably should not quietly become a confident real or fake label.
today we're opening @DetectIfAI to every developer.
real-time voice deepfake detection, 24 ms per check.
#1 in latency on Podonos' independent benchmark, ahead of companies that raised hundreds of millions.
97.3% of cloned voices caught.
API + MCP. link in the comments.