🚀 SpaceXAI and Cursor Prepare to Launch Their First Jointly Developed AI Model 1. SpaceXAI and Cursor are reportedly preparing to ship their first jointly developed artificial intelligence model. 2. The unannounced system stems from a strategic partnership between Elon Musk's newly unified AI entity and the popular AI coding startup. 3. Industry insiders suggest the specialized model could be deployed to select users as soon as this week. Bobbi thinks: Merging SpaceX's raw engineering data with Cursor's developer interface could create a uniquely powerful tool for complex hardware and physics simulations. Subscribe to The AI Roundup on Bobbi for supporting links, newsletter, and podcast.

OpenAI has completely dominated AI voice and conversational prompting. No one else even comes close. GPT is the only AI I use for voice. And even when I'm working in Claude or other tools, I'll just import the transcript across, as it's easier to voice-dump in GPT. Really excited to test this new model - the existing one was already great. You can dunk on OpenAI for every other product except voice. They are clear winners here.

I think many people have underestimated this news. It’s rumored that SpaceX and xAI are collaborating with Cursor to develop their own AI model, moving away from reliance on third-party models. Many people’s first reaction is that this is just another new AI model. But I believe the real focus isn’t the model itself—it’s that they’re now building their own AI production chain. xAI has the computing power; Cursor has the developer access point. Previously, the competition was about model capability—who was the smartest. But I think the real battle moving forward will be about who becomes the indispensable tool in developers’ daily workflows. One more crucial point: I heard the new model will be trained using the Colossus supercomputing infrastructure, aiming to match GPT-5.5 and Opus at a significantly lower cost. If the model truly succeeds in driving down costs, then in the future, enterprises choosing AI won’t just compare who’s the smartest— but who’s the cheapest, fastest, and most user-friendly. After all, the biggest pain point in AI today is the constant burn of token costs. So I believe the next phase of AI isn’t just about model competition. The real battleground is becoming the ecosystem developers can’t live without.

Not having had “early” access to GPT 5.6 for the last two months as one of the new signs of permanent underclass status

A week of HUGE events. Big AI companies announced new models. Open AI announces GPT 5.6 Sol, Terra, Luna Fable-level models to launch on Thursday. SpaceXAI to launch Grok 4.5 tomorrow. Hopefully they will be released for free users too.

[Topics in the Past Hour] ・CFTC Chair explicitly denies CBDC implementation under the Trump administration ・Russia announces ban on diesel exports to ensure domestic supply ・OpenAI prepares to launch new feature: “The All-New ChatGPT Voice”

ANOTHER MASSIVE WIN FOR LOCAL AI. Kyutai quietly dropped a text to speech model that runs on your CPU. 100 million parameters. No GPU. Zero API fees. It bypasses the usual bottlenecks entirely. → 6x faster than real time on a Mac → ~200ms latency to the first audio chunk → Voice cloning from just 5 seconds of audio → 6 languages out of the box 100% free. MIT licensed. Trained only on public data. Voice cloning just became a laptop feature that never touches a server. Bookmark this. Follow @cyrilXBT

DeepSeek-V4 can now run fully local via Unsloth GGUFs Here's what the quantization ladder actually looks like: ⁠1-bit: 92GB total memory 2-bit: 102GB ⁠3-bit: 110-135GB — runs on a Mac with unified memory 4-bit near lossless: 162GB ⁠Q8 lossless: 169GB A few things worth understanding about the Flash variant specifically: •⁠ ⁠284B parameters total, but only 13B active at any time — MoE architecture means you're not running the full model, just the relevant experts •⁠ ⁠The lossless 8-bit GGUF is 162GB, but the 3-bit drops to 103GB with no meaningful quality loss Run it via Unsloth Studio or llama.cpp. Guide and GGUF files are live on HuggingFace. The 1T parameter frontier model that was paywalled last quarter is now running on consumer hardware. How long before local DeepSeek outperforms hosted mid-tier APIs?