Ailing

🌐 #AINavigator

[Ailing @Sushiegrows]

Checked all trendy AI infras/apps, gonna throw out some takeaways how startups grow...

[Ailing @Sushiegrows]

4/4: Abstraction helps the model ignore noise and focus on the one key factor in an environment.

For complex tasks, a step by step structure works best: set a verifiable reward at each step, just like how humans solve difficult problems.

[Ailing @Sushiegrows]

1/4: Is reinforcement learning the right path to AGI?

After watching the Silicon Valley 101 discussion, one point stood out: with a high quality simulator and a rich reward system, RL can create innovation that goes beyond LLMs and break out of the pre-training paradigm.

[Ailing @Sushiegrows]

3/4: Some experts think this is doable for small models because they are cheaper to train.

Abstraction and hierarchy may be the next breakthroughs for RL.

[Ailing @Sushiegrows]

The next real opportunities may be in the upstream of AI. Not just GPUs, but also memory and storage, special purpose chips, and even electricity supply. Semiconductor storage is another area worth watching.

[Ailing @Sushiegrows]

Is there really an AI bubble right now?

I read an analysis today comparing the current AI boom with the tech boom in 2000. Based on valuations, earnings quality, and actual investment activity, it is still too early to call this a bubble.

[Ailing @Sushiegrows]

2) Quantum computing for ML — as quantum computers mature, they may eventually power next-generation AI with unprecedented compute.

Plus, quantum technologies could help solve the massive energy consumption problem in today’s AI compute centers.

[Ailing @Sushiegrows]

QPU (Quantum Processing Unit) is the core of quantum computers.

So how does QPU + AI come together?
1) AI for Science — for example, Google once used quantum AI to optimize error-correcting codes.

More👇

[Ailing @Sushiegrows]

Gemini 3 changes everything. It’s not just an improvement; it's a step-function lead, dominating all benchmarks.

While everyone was betting on RL to break the LLM pre-training barrier, the G3 team credits better pre- and post-training. Maybe the scaling law is still the answer.

[Google Gemini @GeminiApp]

Our most anticipated launch of the year is here.

- Gemini 3, our most intelligent model

- Generative interfaces, for perfectly designed responses

- Gemini Agent, made to complete complex tasks on your behalf

See how Gemini 3 can help you learn, build & plan anything 🧵

[Ailing @Sushiegrows]

@drfeifei @ylecun @Google 5/ 📐 Three directions, three philosophies
• Marble → “Make the world visible”
• Genie 3 → “Make the world explorable”
• JEPA → “Make the world understandable”

[Ailing @Sushiegrows]

Recently, @drfeifei released Marble, a new world model. Here’s a quick look at how the current world-model landscape is shaping up.

Marble vs Genie 3 vs JEPA

Three very different paths toward AGI 🧵👇

[Ailing @Sushiegrows]

@drfeifei @ylecun 4/ 🎞 Genie 3 @Google
• Text prompts generate interactive video worlds
• Solves long-term consistency to keep scenes stable
• Allows agents to explore inside the generated video environment

[Ailing @Sushiegrows]

@mindnetwork_xyz @le_sserafim @chocoha25 @shu_yamino

[Ailing @Sushiegrows]

@mindnetwork_xyz @0xlordachita @hs_tkj @mimiwagmi @tottitti9 thanks mind!

[Ailing @Sushiegrows]

@mindnetwork_xyz 100% HTTPZ！

[Ailing @Sushiegrows]

@mindnetwork_xyz @0G_labs nice