AI Explorer | Claude 2.7 • ChatGPT • Seedance 2.5 CapCut • Runway • Next-gen Creativity Email: [email protected]
Tokyo, Japan
Joined July 2010
- Tweets11.4K
- Following718
- Followers21.1K
- Likes431
Pinned Tweet
Does the best score improve with each release, or does the leader change while the human gap stays similar? A consistent time series could answer that.
Drop 2/15
Are we actually measuring the things we ultimately want our voice AI models to achieve?
The best AI voices are becoming hard to tell apart.
From each other. Not from humans.
🚀 Introducing Turing TTS: an evaluation that measures how often human listeners mistake AI-generated speech for real human speech.
We tested leading TTS models against real human speech. Even the best model fooled listeners only ~10% of the time.
With Turing TTS, we hope to contribute to a broader discussion within the research community about how we measure progress in voice AI.
If one of our goals is to develop speech that is indistinguishable from human speech, how well do our current evaluation methods capture progress toward that goal?
We don’t claim that Turing TTS is the right way forward, or that human indistinguishability should be the goal of every TTS system.
Rather, we hope these findings encourage the community to collectively examine whether blind pairwise human evaluations of naturalness are sufficient, and what else we might need to measure.
Results are now live on @voicearena_ai.
Kaito Nakamura retweeted
Another interesting development from Alibaba's AI ecosystem.
Qoder is Alibaba's coding agent and agent harness, from the same company
behind Qwen. It has just launched a new stealth model called Qoduck.
Download Qoder to try Qoduck for free for a limited time:
qoder.com/
Kaito Nakamura retweeted
We're getting closer to having AI that can take an idea and handle the execution.
Ling-3.2-flash can build interfaces, recreate 3D scenes, analyze research papers, and coordinate multiple agents to deliver complete projects.
This is a much bigger shift than just making chatbots smarter.
Introducing Ling-3.2-flash, Ant Ling's native multimodal model: 560B total / 25B active parameters.
Stronger complex visual understanding, long-horizon multimodal agents, 3D understanding & interaction, and medical intelligence.
Outscores Opus 5 (max) in release evaluations:
- OmniDocBench1.5: 92.21 vs 90.56
- ZEROBench(Sub): 43.71 vs 39.52
- Vision2Web: 85.86 vs 80.20
Kaito Nakamura retweeted
finally something i can show in my tutorials without apologizing for the quality
🤝 Paid partnership
HeyGen Video 1 is the #1 video model on the @OpenRouter leaderboard, and it's now available in HeyGen MCP.
Tell your agent what you want to see, and it's ready in seconds.
Read more: heygen.com/mcp
Kaito Nakamura retweeted
A model doesn't need to sound perfect. It needs to sound like a conversation. That's what real two-person data like this teaches.
Kaito Nakamura retweeted
Multiplayer support?
I asked Claude to build an office for the @tavus PALs...
I was not expecting it to build THAT kind of office
Kaito Nakamura retweeted
For arch nerds: SenseNova Loop-9B-Preview (9B) tests looped computation with shared params + Manifold-Constrained Hyper-Connections. Effective depth up, param count flat. Side track to SenseNova 6.8 Flash.
SenseNova-Loop:
github.com/OpenSenseNova/Sen…
🤝 Paid partnership
𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮 𝟲.𝟴 𝗙𝗹𝗮𝘀𝗵 𝗶𝘀 𝗵𝗲𝗿𝗲!
𝗥𝗲𝗹𝗶𝗮𝗯𝗹𝗲 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀. 𝗕𝗲𝗮𝘂𝘁𝗶𝗳𝘂𝗹 𝗗𝗲𝘀𝗶𝗴𝗻.
🧠 𝗢𝗳𝗳𝗶𝗰𝗲 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲: 𝗟𝗼𝗻𝗴-𝘁𝗲𝗿𝗺 𝗺𝗲𝗺𝗼𝗿𝘆 & 𝗽𝗿𝗼𝗮𝗰𝘁𝗶𝘃𝗲 𝗶𝗻𝘁𝗲𝗿𝗮𝗰𝘁𝗶𝗼𝗻 𝘁hat remember your needs and keep tasks moving.
🎨 𝗗𝘆𝗻𝗮𝗺𝗶𝗰 𝗗𝗲𝘀𝗶𝗴𝗻: Animated PPTs, motion posters, interactive websites & editable Blender 3D projects.
Beyond content generation — getting things done and delivering polished, ready-to-use results.
𝗧𝗿𝘆 𝗶𝘁 𝗳𝗼𝗿 𝗙𝗥𝗘𝗘 𝗼𝗻 𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮 𝗧𝗼𝗸𝗲𝗻 𝗣𝗹𝗮𝗻!
𝗚𝗲𝘁 𝟭,𝟱𝟬𝟬 𝗙𝗥𝗘𝗘 𝗔𝗣𝗜 𝗿𝗲𝗾𝘂𝗲𝘀𝘁𝘀 𝗲𝘃𝗲𝗿𝘆 𝟱 𝗵𝗼𝘂𝗿𝘀.
🔑 Get your API key: platform.sensenova.ai
🤖 Model ID: SenseNova 6.8 Flash
👾Discord: discord.com/invite/BuTXPHmQu…
Kaito Nakamura retweeted
Keywords for SenseNova 6.8 Flash: reliable execution, long-term memory, proactive interaction, AI-native design. Most launches pick one of these. This one tries to land all four.
See if it lands for you:
platform.sensenova.ai/docs
🤝 Paid partnership
𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮 𝟲.𝟴 𝗙𝗹𝗮𝘀𝗵 𝗶𝘀 𝗵𝗲𝗿𝗲!
𝗥𝗲𝗹𝗶𝗮𝗯𝗹𝗲 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀. 𝗕𝗲𝗮𝘂𝘁𝗶𝗳𝘂𝗹 𝗗𝗲𝘀𝗶𝗴𝗻.
🧠 𝗢𝗳𝗳𝗶𝗰𝗲 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲: 𝗟𝗼𝗻𝗴-𝘁𝗲𝗿𝗺 𝗺𝗲𝗺𝗼𝗿𝘆 & 𝗽𝗿𝗼𝗮𝗰𝘁𝗶𝘃𝗲 𝗶𝗻𝘁𝗲𝗿𝗮𝗰𝘁𝗶𝗼𝗻 𝘁hat remember your needs and keep tasks moving.
🎨 𝗗𝘆𝗻𝗮𝗺𝗶𝗰 𝗗𝗲𝘀𝗶𝗴𝗻: Animated PPTs, motion posters, interactive websites & editable Blender 3D projects.
Beyond content generation — getting things done and delivering polished, ready-to-use results.
𝗧𝗿𝘆 𝗶𝘁 𝗳𝗼𝗿 𝗙𝗥𝗘𝗘 𝗼𝗻 𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮 𝗧𝗼𝗸𝗲𝗻 𝗣𝗹𝗮𝗻!
𝗚𝗲𝘁 𝟭,𝟱𝟬𝟬 𝗙𝗥𝗘𝗘 𝗔𝗣𝗜 𝗿𝗲𝗾𝘂𝗲𝘀𝘁𝘀 𝗲𝘃𝗲𝗿𝘆 𝟱 𝗵𝗼𝘂𝗿𝘀.
🔑 Get your API key: platform.sensenova.ai
🤖 Model ID: SenseNova 6.8 Flash
👾Discord: discord.com/invite/BuTXPHmQu…
Kaito Nakamura retweeted
There's going to be a real debate about disclosure once agents are trained on data like this. Worth starting it now.
Kaito Nakamura retweeted
Tight feedback loop, model edition. On annotation-free docs, a general-purpose VLM calls out errors and low-confidence results; disagreeing rollouts surface as uncertain; clustering links shared causes; the VLM writes the round N+1's data. No manual labels. First RSI loop in doc parsing. Light, swappable layout model; one model reads and orders in parallel. Highest on olmOCR-Bench / OmniDocBench / DocParsingBench; 6.77x Surya-OCR-2 on an RTX PRO 6000.
github.com/SoMarkAI/OXR
🤝 Paid partnership
Kaito Nakamura retweeted
A reason to pay attention to @voicearena_ai: it is publishing a result and asking what we should infer from it, rather than stopping at the winner.
Drop 2/15
Are we actually measuring the things we ultimately want our voice AI models to achieve?
The best AI voices are becoming hard to tell apart.
From each other. Not from humans.
🚀 Introducing Turing TTS: an evaluation that measures how often human listeners mistake AI-generated speech for real human speech.
We tested leading TTS models against real human speech. Even the best model fooled listeners only ~10% of the time.
With Turing TTS, we hope to contribute to a broader discussion within the research community about how we measure progress in voice AI.
If one of our goals is to develop speech that is indistinguishable from human speech, how well do our current evaluation methods capture progress toward that goal?
We don’t claim that Turing TTS is the right way forward, or that human indistinguishability should be the goal of every TTS system.
Rather, we hope these findings encourage the community to collectively examine whether blind pairwise human evaluations of naturalness are sufficient, and what else we might need to measure.
Results are now live on @voicearena_ai.
Kaito Nakamura retweeted
Self-improving"" sounds like hype. OXR's isn't.
It runs on unlabeled docs; a general VLM separates real mistakes from unsure cases; disagreeing rollouts add more; clustering finds shared causes; the VLM writes data aimed at them. No one labels it. Doc parsing's first RSI.
Lightweight layout piece you can replace; recognition and ordering simultaneous.
Outscores the field on olmOCR-Bench, OmniDocBench, DocParsingBench; 3.05x GLM-OCR on an RTX PRO 6000.
github.com/SoMarkAI/OXR
🤝 Paid partnership
Kaito Nakamura retweeted
Do u guys have remote roles
I asked Claude to build an office for the @tavus PALs...
I was not expecting it to build THAT kind of office
Kaito Nakamura retweeted
How can i join the team
I asked Claude to build an office for the @tavus PALs...
I was not expecting it to build THAT kind of office
Kaito Nakamura retweeted
@voicearena_ai's own caveats should travel with the numbers. Under this protocol is a small phrase that prevents a lot of overclaiming.
Drop 2/15
Are we actually measuring the things we ultimately want our voice AI models to achieve?
The best AI voices are becoming hard to tell apart.
From each other. Not from humans.
🚀 Introducing Turing TTS: an evaluation that measures how often human listeners mistake AI-generated speech for real human speech.
We tested leading TTS models against real human speech. Even the best model fooled listeners only ~10% of the time.
With Turing TTS, we hope to contribute to a broader discussion within the research community about how we measure progress in voice AI.
If one of our goals is to develop speech that is indistinguishable from human speech, how well do our current evaluation methods capture progress toward that goal?
We don’t claim that Turing TTS is the right way forward, or that human indistinguishability should be the goal of every TTS system.
Rather, we hope these findings encourage the community to collectively examine whether blind pairwise human evaluations of naturalness are sufficient, and what else we might need to measure.
Results are now live on @voicearena_ai.
Kaito Nakamura retweeted
anyone else reading through the SenseNova Loop-9B-Preview repo? 9B looped transformer with mHC, open sourced by SenseTime. separate from SenseNova 6.8 Flash but honestly just as interesting to me
SenseNova-Loop:
github.com/OpenSenseNova/Sen…
🤝 Paid partnership
𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮 𝟲.𝟴 𝗙𝗹𝗮𝘀𝗵 𝗶𝘀 𝗵𝗲𝗿𝗲!
𝗥𝗲𝗹𝗶𝗮𝗯𝗹𝗲 𝗪𝗼𝗿𝗸𝗳𝗹𝗼𝘄𝘀. 𝗕𝗲𝗮𝘂𝘁𝗶𝗳𝘂𝗹 𝗗𝗲𝘀𝗶𝗴𝗻.
🧠 𝗢𝗳𝗳𝗶𝗰𝗲 𝗜𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝗰𝗲: 𝗟𝗼𝗻𝗴-𝘁𝗲𝗿𝗺 𝗺𝗲𝗺𝗼𝗿𝘆 & 𝗽𝗿𝗼𝗮𝗰𝘁𝗶𝘃𝗲 𝗶𝗻𝘁𝗲𝗿𝗮𝗰𝘁𝗶𝗼𝗻 𝘁hat remember your needs and keep tasks moving.
🎨 𝗗𝘆𝗻𝗮𝗺𝗶𝗰 𝗗𝗲𝘀𝗶𝗴𝗻: Animated PPTs, motion posters, interactive websites & editable Blender 3D projects.
Beyond content generation — getting things done and delivering polished, ready-to-use results.
𝗧𝗿𝘆 𝗶𝘁 𝗳𝗼𝗿 𝗙𝗥𝗘𝗘 𝗼𝗻 𝗦𝗲𝗻𝘀𝗲𝗡𝗼𝘃𝗮 𝗧𝗼𝗸𝗲𝗻 𝗣𝗹𝗮𝗻!
𝗚𝗲𝘁 𝟭,𝟱𝟬𝟬 𝗙𝗥𝗘𝗘 𝗔𝗣𝗜 𝗿𝗲𝗾𝘂𝗲𝘀𝘁𝘀 𝗲𝘃𝗲𝗿𝘆 𝟱 𝗵𝗼𝘂𝗿𝘀.
🔑 Get your API key: platform.sensenova.ai
🤖 Model ID: SenseNova 6.8 Flash
👾Discord: discord.com/invite/BuTXPHmQu…
Kaito Nakamura retweeted
It's not just about speaking. It's about knowing when to listen. 🔥
Kaito Nakamura retweeted
When human creativity meets AI, new cinematic possibilities emerge. 🚀
Kaito Nakamura retweeted
OXR brings self-improving AI to document parsing, learning from mistakes without human labeling.
Top-ranked on 3 benchmarks and 15.43x faster than Chandra-OCR-2.
github.com/SoMarkAI/OXR
🤝 Paid partnership