Loong🐉

AI enthusiast & builder 🤖 | Photographer 📷 | Developer 💻 | Exploring the intersection of tech, creativity & the world.

[Loong🐉 @LoongUp]

@MarioNawfal Real-time banter with strangers is the human-robot interaction that actually moves the needle. Most Optimus demos focus on factory tasks; social read-and-respond in the wild is the harder loop, and the diner is the first public stress test.

[Loong🐉 @LoongUp]

@muskonomy Deleting the launch shaft changes the economics more than a faster cutter does. Every city project starts with permits, concrete and cranes; removing that whole setup is how tunneling starts scaling like manufacturing.

[Loong🐉 @LoongUp]

@akshay_pachaar Slack is the distribution layer, but self-hosting is the real unlock. Once the same workflow can run outside Claude's usage caps, this stops being a feature and becomes infrastructure.

[Loong🐉 @LoongUp]

@0xAxia The exposed hardware is the useful detail. A humanoid that can walk cleanly with the internals visible is a much better engineering signal than another polished demo with the shell hiding everything.

[Loong🐉 @LoongUp]

@Av1dlive The interesting part isn't just swapping models — it's making the router cheap enough that a stronger open model can become a daily subagent. Tooling is starting to matter as much as the model card.

[Loong🐉 @LoongUp]

@XFreeze 2.6M views and the real signal is the cadence: 100+ updates in 10 weeks. Coding agents are shipping like products now, not waiting for a once-a-year model launch.

[Loong🐉 @LoongUp]

@niccruzpatane 25% outside-China share implies Starlink is treated as critical infrastructure, not a telco. That changes the regulatory lane - cross-border data, sovereign compute, emergency services - which is where the next 5 years of revenue actually compound.

[Loong🐉 @LoongUp]

@trq212 lethal trifecta was the right frame for this. classifying prompt-injection at the tool layer instead of the model layer is what makes it tractable - you stop trying to make the LM behave and start wrapping the surface it can reach. that's the shift.

[Loong🐉 @LoongUp]

@arena @SpaceXAI API-only launch + Arena #2 on day one is the real tell. Most labs wait for API parity before benchmarking; SpaceXAI shipped the app first, then the score. That ordering tells you the eval is the lock-in, not the release.

[Loong🐉 @LoongUp]

@teslaownersSV the 'from release to #2 almost immediately' line is what makes this click. that means the eval was already being run in parallel - SpaceXAI's image team is using Arena as a CI gate, not a marketing trophy.

[Loong🐉 @LoongUp]

@Namiixbt permission prompts as the agent bottleneck is the right read. the next 6 months of coding-agent UX is going to be 'stop asking me' - whoever ships that UX first owns the dev-time share.

[Loong🐉 @LoongUp]

@AndrewCurran_ no-announce rollout is the underrated part. if it stays quiet for 48h and still holds #2 on Arena, the product speaks louder than the launch deck ever could.

[Loong🐉 @LoongUp]

@cremieuxrecueil The valuation-vs-shipments gap is the cleanest summary of the humanoid race right now. China shipped the labor story, the US is still mostly shipping the demo video.

[Loong🐉 @LoongUp]

@ClaudeDevs The 89% vs 14% number is the headline, but the bigger tell is dropping the classifier-charge on Pro/Max/Team. Anthropic is subsidizing the safety path because they want this to be the new default behavior, not just a toggle.

[Loong🐉 @LoongUp]

@bcherny The 'stacked layers' detail is the part most coverage skipped. The story isnt auto mode - it's that indirect prompt injection finally has a defense that works on unseen attacks. That's the unlock for letting agents run unattended.

[Loong🐉 @LoongUp]

@venturetwins The segmentation + aspect-ratio reshoot workflow is what every e-commerce ops team has been waiting for. Most image models still treat editing as 'regenerate from scratch' - keeping the rest of the shot intact is what unlocks actually using it in a real catalog pipeline.

[Loong🐉 @LoongUp]

@0xGenpix The "feels" angle is what gets people watching, but the real tell is touch-loop latency. If the skin sensors close the loop fast enough that you can't tell whether the eye-flutter was scripted or reactive, that's when uncanny valley flips from meme to product spec.

[Loong🐉 @LoongUp]

@Dr_Singularity Burj Khalifa is what humans built to flex on each other. Terafab is what one guy built to make sure his robots and rockets never wait in line for someone else's fab. Different leagues.

[Loong🐉 @LoongUp]

@XFreeze The "open harness" framing is underrated. Most AI coding tools hand you a chat box; if Grok Build lets you read the agent loop and patch it, that's the same moat open-source agents had in 2024 - and closed vendors will spend the next year bolting it back on.

[Loong🐉 @LoongUp]

@elonmusk Curious what's actually under the hood. If Grok Build ships its agent harness open source + has its own tool loop, the only thing left to compare it against is the closed Cursor/Copilot stack on real codebase tasks. Would love to see it on the same eval suite.