- AI Fire
- Posts
- 🤯 AI Crossed Another Line
Humanoids just walked into real factories. Reddit thinks Gemini is acting strange. And a new AI learned physics from scratch like a student with unlimited time. The pattern here is not obvious. Yet.
What are on FIRE 🔥
IN PARTNERSHIP WITH SPINE AI
You're paying for ChatGPT, Claude, Midjourney, and Perplexity. But you're still copy-pasting between tools, losing context, and wasting time.
Spine AI gives you 300+ AI models in one visual workspace where you can:
Run 5+ different models in parallel to critique your strategy
Branch ideas without polluting your original thread
Go from research → prototype → presentation without switching tools
For product managers, designers, and founders who think visually: This is your AI-amplified thinking workspace.
→ Result: What takes 3 hours across multiple tools now takes 20 minutes in Spine.
AI INSIGHTS
UBTECH just shipped hundreds of Walker S2 humanoids into real factories across Shenzhen. Not demos. Real production lines. Orders already passed 800M yuan, with buyers like BYD, Geely, FAW Volkswagen, Dongfeng, and Foxconn. UBTECH plans 500 units delivered by December.
The edge: Each Walker S2 runs UBTECH’s industrial AI stack for vision, navigation, and task planning. It can map factory floors, spot objects, adjust grip strength, and plan moves without human help. The robots share data, so each shift makes the whole fleet smarter.
The advantage: Walker S2 swaps its own battery in minutes. Full 24/7 cycles. Zero downtime. Early tests show strong performance under real factory pressure.
The numbers: Humanoids now make up 30% of UBTECH’s sales. H1 2025 revenue hit 621M yuan. Stock is up 150% to HK$133. Citi and JPMorgan expect it to climb past HK$170.
Why it matters: This is the first real AI-powered humanoid rollout at scale. AI didn’t just write code this year. It stepped onto factory floors. Humanoids aren’t a hype demo anymore. They’re entering the workforce at scale. UBTECH may be leading the next wave of AI-driven automation.
PRESENTED BY BELAY
Q4 is the perfect window to turn this year’s numbers into a clear, actionable forecast aligned with your goals. Set your business up for a stronger 2026 with BELAY’s new guide.
AI SOURCES FROM AI FIRE
1. 5 powerful Google AI Studio tricks you should use for $0. This post shows 5 amazing things you can build, create, and do right now
2. Crafting a GPT brain for automated Solana memecoin trades. Given a small budget, the custom-built agent navigated the memecoin market with logic
3. The modern guide to lead an AI workforce for productivity. We use parallel agents for strategic AI for planning to seriously speed up your projects
TODAY IN AI
AI HIGHLIGHTS
🍃 Reddit users are arguing whether Gemini 2.5 Pro was “nerfed” or if pro-innovation bias is fooling everyone. Some swear they’re seeing early Gemini 3.0 behavior. Full post here.
⚡ Spending on data centers just passed new oil exploration. OpenAI, Meta, Anthropic are pushing trillion-level builds. Solar-powered sites may explode next. More details here.
🌲 Google DeepMind released a global 10-meter AI map of Natural Forests that separates real forests from plantations with 92.2% accuracy. Big win for deforestation-free supply chains. Dataset here.
🔬 China’s new AI-Newton learned physics laws from scratch. It rediscovered Newton’s second law and stored concepts for later tasks. Symbolic regression is getting real. Breakdown here.
🛡️ Yoshua Bengio - one of the godfathers of AI - warns AI safety risks are already real. Malicious use is happening today, and future models need safety built-in. Interview here.
💰 AI Daily Fundraising: Obello raised $8.5M to scale its AI creative platform used by Apple, Nike, Netflix, and Instagram. The round was led by Obvious Ventures with support from Baukunst, AVV, and Preview Ventures. Its GLAM engine helps brands automate on-brand design across every channel.
HOT PAPERS OF THE WEEK
Weight-Sparse Transformers Have Interpretable Circuits
OpenAI trains transformers where 99.9% of weights are zero, uncovering small, human-readable circuits - 16× smaller yet fully interpretable. Scaling them without losing clarity is the next challenge. (OpenAI)Aligning Machine and Human Visual Representations
DeepMind aligns AI vision with human perception using AligNet, reorganizing model features by human categories. The result: better robustness and human-like reasoning. (Google DeepMind)Intelligence per Watt: Measuring AI Efficiency
Stanford and Together AI define Intelligence per Watt (IPW) - a new benchmark merging accuracy and energy. Local models now handle 88.7% of queries with 5.3× efficiency gains since 2023. (Stanford • Together AI)Omnilingual ASR: Speech Recognition for 1600+ Languages
Meta’s open-source Omnilingual ASR supports 1600+ languages - 500 brand new to ASR. It outperforms Whisper and enables zero-shot recognition with just 10-100 samples. (Meta FAIR)AlphaProof: Olympiad-Level Formal Math Reasoning
DeepMind’s AlphaProof becomes the first AI to win a silver medal at the IMO, solving 3 of 5 problems using RL-based theorem proving inside Lean. (Google DeepMind)
NEW EMPOWERED AI TOOLS
⌨️ ClipboardAI runs AI prompts anywhere with a shortcut - fix, translate, or rewrite text instantly.
🗓️ Planndu turns thoughts into structured tasks, combining AI to-dos, notes, timers & routines in one app.
📊 Repo Visibility Analyzer boosts GitHub SEO, auditing repos to improve discoverability on Google & LLMs.
🎥 Sendr personalizes outreach at scale with lip-synced AI videos, smart automations & 479M+ B2B contacts.
AI DISCOVERY
ValueCell is a clean demo of how LLMs + RAG + memory can run real trading, not theory. It’s open source with 5.8K GitHub stars.
One orchestrator handles planning and memory. Individual agents handle their own tasks. Simple and modular.
It ships with 3 key agents:
DeepResearch Agent - Pulls filings, analyzes docs, and gives clear insights.
Strategy Agent - Runs multi-asset crypto strategies and executes real orders.
News Agent - Schedules automatic news scans like “Tesla every 5 minutes.”
It supports OpenRouter, SiliconFlow, Google, and OpenAI. Switch models by editing .env.
For trading, it connects to OKX and Binance. Add API keys, set prompts, and the agent trades with guardrails.
The dashboard shows virtual/live modes, logs, and equity curves. It also includes US, HK, China, and crypto data so one setup works across markets.
Why it matters: This is one of the clearest use cases of LLMs + RAG + memory put into real financial workflows. And while everyone is busy asking AI for charts or summaries, ValueCell pushes AI into actual execution. Agents aren’t just talking. They’re researching, scheduling, deciding, and placing trades.
We read your emails, comments, and poll replies daily
How would you rate today’s newsletter?Your feedback helps us create the best newsletter possible |
Hit reply and say Hello – we'd love to hear from you!
Like what you're reading? Forward it to friends, and they can sign up here.
Cheers,
The AI Fire Team
Reply