Dev Shah
@0xDevShah
Followers
4K
Following
13K
Media
192
Statuses
2K
Audodidact | Devrel-GTM-Agents-Prompting @resembleai | https://t.co/FiJLqfJUpH
San Francisco
Joined September 2019
This is the DeepSeek moment for Voice AI. Today we’re releasing Chatterbox Turbo — our state-of-the-art MIT licensed voice model that beats ElevenLabs Turbo and Cartesia Sonic 3! We’re finally removing the trade-offs that have held voice AI back. Fast models sound robotic.
161
381
4K
We'll know we've achieved AGI when a model can build-test-iterate to solve problems it hasn't seen before by bootstrapping architectures and frameworks. Just like humans, we do not use raw intelligence to solve problems. We instead design methods, tools, and frameworks. At a
AI will increasingly turn R&D into a high speed automated search process, where models propose designs, automation executes tests, and results feed back into the next round. This will compress timelines and reduce costs in domains with automatable experiments and clear metrics,
0
3
13
0
1
7
the techno optimists and ai safety specialists stopped being able to inhabit the same reality because one sees capability gains, the other sees alignment debt compounding
0
1
6
it's hard to take any benchmark seriously that claims Gemini 3.0 flash is better than Opus 4.5. Nothing beats Opus, it is the best we have!
18
4
59
needing a Metr graph to notice ai acceleration is wild
We estimate that, on our tasks, Claude Opus 4.5 has a 50%-time horizon of around 4 hrs 49 mins (95% confidence interval of 1 hr 49 mins to 20 hrs 25 mins). While we're still working through evaluations for other recent models, this is our highest published time horizon to date.
0
0
6
We'll know we've achieved AGI when a model can build-test-iterate to solve problems it hasn't seen before by bootstrapping architectures and frameworks. Just like humans, we do not use raw intelligence to solve problems. We instead design methods, tools, and frameworks. At a
AI will increasingly turn R&D into a high speed automated search process, where models propose designs, automation executes tests, and results feed back into the next round. This will compress timelines and reduce costs in domains with automatable experiments and clear metrics,
0
3
13
the future is gonna be wild. we'll have humanoid robots like @Tesla_Optimus, running on @grok brain, controlled directly by our thoughts via @neuralink. We'll have @SpaceX taking us interplanetary, full self driving everywhere, ai everywhere, and we'll be cruising the cyber
Try Tesla self-driving. It’s a gamechanger!
1
2
9
it's hard to take any benchmark seriously that claims Gemini 3.0 flash is better than Opus 4.5. Nothing beats Opus, it is the best we have!
18
4
59
Elon seems to have particularly internalized 2 beliefs - > Long time horizons make short-term social costs irrelevant. Most actors and entities think in quarters or years, where social cost compounds faster than they can build escape velocity. But because Musk thinks in decades,
a lot of people think they can act like elon, then they do, and immediately find out that they can't handle the consequences. many such cases
0
3
11
Identity certainty correlates with better functioning across domains, but the hidden tradeoff is that you're paying computational costs to avoid the anxiety of maintaining multiple equilibria simultaneously. the people who "find themselves" aren't discovering truth, but are
0
1
8
Gemini only flags generated content that includes synthID, which means it has to come from Google's own tools. Resemble AI doesn't have that limitation. It can detect deepfakes across videos, audios, and images, no matter what models created them. Powered by Detect-3B Omni and
You can now ask @GeminiApp: "Is this video made with AI?" 🔍 Upload the file and it will check for the SynthID watermark to help verify if it was created or edited by Google tools. Find out more → https://t.co/rK9ucF8KFA
5
7
71
We released Chatterbox Turbo on Monday. Here's @theonlysaqib and Tedi on what we built and why. ~100ms latency. 6x faster than real-time. Expressive sound tags. PerTh watermarking on every output. The full conversation on the engineering and the thinking behind it.
1
1
13
Meta is dropping Avocado and Mango, OpenAI is dropping Strawberry and Garlic, Google is dropping ditto and banana we're two fruits away from AGI
Meta is dropping two AI models - Avocado and Mango sometime next year It will literally be a HUGE MIRACLE if they come anywhere close to top models. They are almost 18 months behind which may as well be a decade in AI land
0
1
11
Time travel through latent space.
Check out Nano Banana Pro in action, right in Google Search 🍌✨👀 After seeing folks visualize coordinates, we tried the prompt “Visualize 40.7422° N, 73.9880° W in 1916” and reran it several times, adding a decade with each go. Here’s a GIF compilation:
0
2
22
enterprise software is sticky because it is hard to install. we'll know agi has arrived when it can integrate salesforce in 5 mins and it can rip it out in 5 mins. churn drops to zero in value, but skyrockets for lock-ins
0
2
7
elon only reports to the simulation overlords
@jasonlk I was CEO and Peter reported to me, so could not fire me. It was a palace coup by most (not all) of the exec team and most of the board, who were worried that my decisions were too risky. I was the largest shareholder in the company. There was nothing anyone could have done to
0
1
6
Why hasn't @Microsoft @github added a "Talk to Repo" or something button on every github repo? Right now I have to get the repo into cursor or chatgpt or etc for it to scan it and qa about it?
36
3
263
Gemini only flags generated content that includes synthID, which means it has to come from Google's own tools. Resemble AI doesn't have that limitation. It can detect deepfakes across videos, audios, and images, no matter what models created them. Powered by Detect-3B Omni and
You can now ask @GeminiApp: "Is this video made with AI?" 🔍 Upload the file and it will check for the SynthID watermark to help verify if it was created or edited by Google tools. Find out more → https://t.co/rK9ucF8KFA
5
7
71
Thanks @googlecloud for the feature! We've deployed Gemini 3.0 Flash and are seeing excellent performance and compatibility across all our systems - from voice agents to forensic grade deepfake detection systems. Great work, team!
Gemini 3 Flash unlocks new and more efficient use cases for companies across industries 🧵↓
0
2
11