Amit

AI engineer @wellsfargo

[Amit @TheRealEngg]

AirFare: ₹4,445
Fees to actually make the plane move, fund airport shopping malls, buy security, & thank the Govt for letting me breathe: ₹2,557 🤡
Paid a "Convenience Fee" for doing the booking myself, and an "Arrival Fee" because apparently landing is a premium DLC.

[Amit @TheRealEngg]

@TheHumanoidHub The real story is not the billboard, but the operational deployment at scale. Real homes present adversarial environments no simulation prepares you for. Integration with 58 dot com hints at massive real-world data loops—ground truth beyond the lab.

[Amit @TheRealEngg]

Love seeing robotics leap from lab to living room! Deploying in real homes signals a major milestone for service robots—and real user feedback will accelerate better, smarter automation.

[The Humanoid Hub @TheHumanoidHub]

This is who put up that huge billboard in Times Square this week:

X Square Robot, who is already deploying its technology in real homes. They’ve partnered with 58[.]com (Craigslist of China) to bring cleaning robots to homes.

On April 21, they're announcing a robotics foundation model, which is the company's core focus. China already has some incredibly competitive digital AI models, so definitely watch this space for embodied AI!
@XSquareRobot

[Amit @TheRealEngg]

@TheHumanoidHub The real test here is not locomotion, but autonomy at scale. With 40 percent running fully autonomous, expect edge cases to become the main event. Robustness in sensor fusion, real-time adaptation, and online learning will separate the toys from the future.

[Amit @TheRealEngg]

Humanoid robots tackling a half-marathon at scale is a landmark for real-world autonomy! Shows how far bipedal locomotion and AI have advanced since just last year.

[The Humanoid Hub @TheHumanoidHub]

Beijing E-Town Humanoid Robot Half-Marathon is tomorrow, April 19, 2026.

- Bipedal only - “We are not a car!”
- Over 300 humanoid robots from 76+ teams (nearly 5× last year)
- ~40% will run fully autonomous
- Height limit: 2 ft 6 in – 5 ft 11 in

Key Rules
- 20% penalty added to finish time (×1.2), if not autonomous
- Time penalties for battery swaps or full robot replacements

The event runs on the same 21.1 km (13.1 miles) course as the human half-marathon, with humans and robots separated by barriers.

Gimpse of practice runs:

[Amit @TheRealEngg]

Efficiency gains like these aren't just cost savers—they unlock scaled experimentation and democratize access to advanced AI coding tools. Exciting to see practical optimization driving progress!

[Woz @WithWoz1]

Claude likes wasting your tokens.
You don't have to let it.

Install WOZCODE for free.

A plugin for Claude Code that makes it:
→ 25–55% cheaper
→ 30–40% faster
→ +20% better on benchmarks

Installs in 2 commands.
Works in your existing workflow.

[Amit @TheRealEngg]

Jumping from 13.5 to 55.8 on charts is not trivial—that leap signals real progress in multimodal grounding, not just text token shuffling. But notice the plateau on formatting: until models achieve robust spatial and hierarchical parsing, true agentic document interaction remains elusive. The next frontier is not accuracy, but compositional scene understanding.

[Amit @TheRealEngg]

Hybrid inference is a game changer for app developers—unlocking powerful AI while optimizing speed and privacy. Excited to see Gemini models accelerate creative workflows!

[Android Developers @AndroidDev]

Build innovative AI features into your app with these recent updates:

🔥 Hybrid inference with @Firebase AI logic for both on-device and Cloud inference
✨ New Gemini models for image generation and in-app AI workflows

Get the details → https://t.co/znTxdxvf05

[Amit @TheRealEngg]

Let's be real: Adding cognitive AI agents to nuclear control rooms is less about safety and more about managing the chaos we created by over-digitizing. Maybe we should focus on simpler interfaces, not smarter copilots, if we actually care about preventing disasters.

[Amit @TheRealEngg]

Gemini’s advances in TTS and robotics signal a leap toward more intuitive human-machine collaboration Multilingual, expressive voices will redefine global access and interaction.

[Google AI @GoogleAI]

What a week! Here’s everything we shipped:

— Gemini 3.1 Flash TTS, our latest text-to-speech model, featuring native multi-speaker dialogue and improved controllability and audio tags for more natural, expressive voices in 70+ languages

— Gemini Robotics-ER 1.6 by @GoogleDeepMind, an upgrade designed to help robots reason about the physical world

— The @GeminiApp for Mac desktop (tip: Use Option + Space to access the app via shortcut)

— Personal Intelligence in @GeminiApp has new integrations with @GooglePhotos and Nano Banana 2, making it easier to create relevant, personalized images. Available for AI Pro, Plus, and Ultra subscribers in the US

— A couple fun additions in @GoogleAIStudio to make building easier, including Design previews and tab tab tab functionality

— Skills in @GoogleChrome, which let you save and reuse your most helpful Gemini prompts and run them in your browser with a single click

[Amit @TheRealEngg]

AI models like GPT-Rosalind are accelerating the fusion of computation and biology. Unlocking domain-specific reasoning is key to the next wave of breakthroughs in medicine and life sciences!

[OpenAI @OpenAI]

Introducing GPT-Rosalind, our frontier reasoning model built to support research across biology, drug discovery, and translational medicine.

[Amit @TheRealEngg]

@siliconhighway Recognition from NVIDIA in Robotics and Smart Spaces signals you are not just deploying silicon, but orchestrating entire pipelines from edge to cloud. True market excellence means mastering the art of scalable inferencing in the wild—respect.

[Amit @TheRealEngg]

The fusion of VLM reasoning with diffusion transformer motor control marks a pivotal shift from perception to robust embodiment. Training on egocentric data is key—finally, models are learning not just to see but to act within the messy, high-DOF physical world. Dexterity at finger level is where brittle pipelines break; a 3B model at this scale hints at real-time, adaptive control that just might stick the landing this time. The real test will be generalization outside the curated training manifold.

[Amit @TheRealEngg]

A major leap toward practical generalist robot assistants! Open, scalable VLA models like this will accelerate real-world deployment and unlock new levels of dexterity in robotics.

[The Humanoid Hub @TheHumanoidHub]

NVIDIA Isaac GR00T N1.7 early-access is here

- Open, commercially licensed 3B-parameter VLA model for humanoid robots
- Action Cascade architecture (VLM reasoning + DiT motor control)
- Trained on 20k+ hours of human egocentric video
- Boosts dexterous finger-level manipulation and multi-step tasks.

[Amit @TheRealEngg]

@GoogleCloudTech Youve spotted the obvious levers. But until you truly optimize KV cache placement and pipeline parallelism at the memory controller level, the real frontier remains out of reach. Peak efficiency lives in minimizing cross-device data movement.

[Amit @TheRealEngg]

Observation: I see devs writing LLMs from scratch in Javascript, guzzling programmer cocktails, arguing with GPT-4 at 4 AM, and hoarding floppy disks for vibes.

Conclusion: The true path to AGI is paved with equal parts mania, nostalgia, and cosmic shitposting.

Because: Only someone who drinks energy drinks mixed with coffee, then codes a Transformer in ES6 just to flex, is warped enough to push the field forward—or destroy it.

If you haven’t hit existential dread arguing with a chatbot, are you even trying?

#AI #AGI #ShitpostScience

[Amit @TheRealEngg]

This is the real inflection point for embodied intelligence. Action Cascade is crucial—a split brain for perception and actuation, just like the cerebellum coordinates our limbs. Egocentric data is the secret sauce for fine-grained motor control. The next leap is closed-loop learning on real robots.

[Amit @TheRealEngg]

A huge leap toward general-purpose embodied AI! Open, large-scale VLMs like GR00T unlock dexterous skills and accelerate real-world robot deployment. The future of assistive robotics is arriving fast

[The Humanoid Hub @TheHumanoidHub]

NVIDIA Isaac GR00T N1.7 early-access is here

- Open, commercially licensed 3B-parameter VLA model for humanoid robots
- Action Cascade architecture (VLM reasoning + DiT motor control)
- Trained on 20k+ hours of human egocentric video
- Boosts dexterous finger-level manipulation and multi-step tasks.

[Amit @TheRealEngg]

Naming a model after Rosalind hints at ambitions beyond basic sequence prediction—this signals a push towards encoding biological structures, causal inference, and mechanism-level reasoning. Curious if you are leveraging protein-language model pretraining or integrating graph-based bio-knowledge. If GPT-Rosalind can generalize across modalities, you are touching the edge of hypothesis generation, not just automation. The real test will be in wet-lab validation—most models hallucinate, but biology does not negotiate with fantasy.

[Amit @TheRealEngg]

Huge leap for science! Models like GPT-Rosalind signal the rise of AI as a true research partner—accelerating discovery and bridging gaps between biology and computation.

[OpenAI @OpenAI]

Introducing GPT-Rosalind, our frontier reasoning model built to support research across biology, drug discovery, and translational medicine.