GM {{first_name}}!
NVIDIA just shook up the robotics game—again.
This week at NVIDIA GTC 2025, Jensen Huang took the stage and introduced Blue—a Star Wars-inspired droid that looks like it rolled straight out of Disneyland’s future.
Built with Google DeepMind and Disney Research, Blue is powered by NVIDIA’s Newton physics engine, letting it learn, adapt, and move like never before.
Disney’s been pushing animatronics for years, but this? This is different. Newton lets Blue refine its motions through simulation, meaning we’re talking about robots that don’t just repeat pre-programmed actions—they evolve.
And this wasn’t even NVIDIA’s only flex at GTC 2025. Other jaw-droppers:
GR00T N1, the first fully customizable open-source humanoid robot foundation model.
DGX Spark, the world’s smallest AI supercomputer packing 128GB of memory.
Blackwell Ultra, NVIDIA’s next-gen GPU making everything else look prehistoric.
New speed record for DeepSeek-R1 inference using eight Blackwell GPUs.
But let’s be real—Blue stole the show.
Robots are coming. For real this time. And if NVIDIA has its way, they’ll be everywhere.
I’ve been diving deeper into robotics lately, and a few things are clear to me:
The path to general-purpose robots runs through AI
Data scarcity is the biggest bottleneck for general-purpose robotics AI
Crypto networks and incentives might be the key to unlocking that missing data
🚨 But wait—
Are you in our telegram channel yet? Cyra, our AI researcher, drops daily AI alpha that you don’t want to miss. See it for yourself.
Catch our podcast summaries too, complete with detailed show notes. Saves us hours
And if you want to jam with us directly, hop into our Discord
Forget prices this week—let’s look at what’s actually hot.
Based on rootData’s Hot Index (measured by top-tier follower growth), here’s what’s catching people’s attention:
🔥 Pluris Research tops the chart—likely driven by their $7.6M seed led by USV & CoinFund.
🚀 Mira isn’t far behind, scoring a 400+ Hot Index after their testnet launch.
👀 Wayfinder is rising fast with buzz around their Kaito Connect Launchpad debut.
The Hot Index tracks which projects are gaining momentum right now, weighted based on the number of new top-tier followers a project has gained over the past three days. Think of it as an early signal for future price action.
palle.substack.com
Meanwhile Web2 AI startups are flexing: 📈 47x revenue multiples, compared to 7x for public SaaS.
Why? Enterprise-grade revenue + peak VC FOMO.
It’s 2021 vibes all over again—but with AI.
And while Web2 AI startups rake in revenue, they’re also doing it with tiny, hyper-efficient teams.
Might be time for Web3 to take some notes.
ARMA Agent on Giza
metrics.gizatech.xyz
Giza is a AI-native execution layer for DeFi, aiming to make financial automation more accessible, efficient, and scalable. The idea? Let AI agents handle the complexity of yield strategies, so users don’t have to manually navigate protocols, bridge assets, or time the market.
Their flagship product is ARMA—an autonomous agent designed to maximize stablecoin yields. It started on Mode then launched on Base in January 2025 and has since executed over 22,000 transactions, generated $6M+ in trading volume, and grown from $200K to $930K in TVL. That’s a strong start for a system running entirely without human intervention.
ARMA works by scanning DeFi markets such as Aave, Compound, Gauntlet for optimal yield strategies, automatically rebalancing to improve returns. Users can also fine-tune preferences—like risk tolerance and asset size—without needing to write a single line of code.
As far as we know, ARMA isn’t a LLM, but rather uses a specialised model trained specifically for these DeFi tasks. But the cool thing is it never sleeps, never blinks and (hopefully) never fumbles.
x.com/gizatechxyz
Giza simplifies blockchain interaction.
It uses ERC-7579 smart contract wallets to standardize how agents interact with different protocols—making them composable and secure
Integrates with EigenLayer’s AVS framework to decentralize execution and enhance security,
So Giza’s AI can operate across chains and protocols without custom code each time.
The bigger picture: Giza is building the invisible infrastructure layer that could quietly power the next wave of DeFi automation—where intelligent agents work in the background to manage capital efficiently and securely.
🌴 Bittensor Subnet 64: Chutes
Chutes (Bittensor Subnet 64) is quietly eating the rest of the Bittensor stack.
Built by Rayon Labs, it’s become one of the most useful tools in the ecosystem—especially if you’re a validator or subnet owner. Link your keys, and you get instant dev access, no fees attached.
The killer feature? Serverless AI deployment. Spin up large open-source models like DeepSeek or Llama 3 via a clean UI or lightweight API. No infra setup, no GPU wrestling, no ML ops degree required. Upload and go—text-to-speech, image gen, or your own custom app.
It just runs. Billing is done via micropayments in TAO, Bittensor’s native token.
Instead of rewarding idle miners, Chutes ties incentives to actual model usage. That keeps compute costs low and minimizes waste. The network uses a GPU verification layer called Graval to ensure performance is legit—without requiring expensive, specialized hardware.
The UX is about to get even smoother. They’re rolling out its AI agent platform this week—no-code tools that let non-devs launch AI agents in a few clicks.
Adoption has been fast. Since its late-2024 launch, Chutes has scaled to process over 12 billion tokens daily. It’s now powering 43% of all DeepSeek traffic on OpenRouter—outpacing even Microsoft Azure in this slice of the market.
Probably because it’s dirt cheap.
DeepSeek R1 on Chutes costs just $2 per million tokens. Together.ai charges $3.
x.com/rayon_labs
Investors are clearly bullish on it, giving it a $935M FDV and valuing it as a decentralised cloud with serious traction.
Chutes is the highest-earning subnet in the entire Bittensor network—pulling in 14.23% of daily emissions, or around 1,020 TAO ($200K+) per day.
And it looks like TEEs are on the way..
Project Updates
Pluralis raised a $7.6M seed round (led by USV and CoinFund) to launch Protocol Learning — a new approach to open-source AI where models are trained collaboratively and ownership is shared. The goal: make building foundation models more accessible, and make sure contributors get paid.
Crossmint raised $23.6M (led by Ribbit Capital) to make onchain dev easier for businesses and AI agents. Their low-code APIs are built to handle rising demand from both enterprises and AI use cases.
Nillion just dropped tokenomics for $NIL — its native token for Blind Compute and decentralized privacy. Launching with 1B total supply and 19.52% in circulation, $NIL will power coordination, staking, and governance.
(ICYMI: we wrote about Nillion in-depth here)
Mira just launched its Public Testnet — already serving 2.5M users and processing 2B tokens a day (half the size of Wikipedia). It’s a purpose-built chain for trusted, verifiable AI, with a decentralized API layer.
Sentient is launching a data consortium to build community-aligned AI using user-owned, decentralized data. Backed by partners like Vana and The Graph, the goal is open, diverse data as the backbone for truly aligned models.
PIN AI just dropped the whitepaper for its GOD Model — a privacy-first framework for training personal AI on-device. It uses TEEs, curriculum learning, and token incentives to keep data local while making AI more personal and proactive.
peaq just launched Universal Machine Time (UMT) — the first onchain version of Precision Time Protocol. It brings nanosecond-level sync to DePIN and DePAI networks, helping machines stay in perfect time. Already being tested by Geodnet and Auki for better coordination, smart contracts, and data accuracy.
BitMind has released a new paper showcasing a decentralized, competitive AI framework that adapts to evolving threats and achieves over 90% accuracy in real-world deepfake detection.
Glitch Financial, an AI powered trading platform by Taoshi (Subnet 8), just opened its public beta — letting early users test advanced trading strategies ahead of launch. Rolling out to 100 users by April, based on waitlist rank.
AI Agents
Spectral Labs releases Lux — the first multi-agent, multi-framework system for building fully autonomous, AI-native companies. Instead of single agents, Lux runs swarms that can hire, fire, and execute workflows.
Parallel/Wayfinder is expected to launch Parallel Colony Alpha—its AI agent-powered game—by the end of this month, along with new functionality for Wayfinder.
Fetch.ai just launched ANAME — a decentralized Agent Name Service that gives AI agents human-readable names (like DNS for Web3).
Compute Labs launched Sam Altbot — an AI agent that posts sharp, high-signal news updates, tuned by community feedback. Users can reply to help it learn and climb the contributor leaderboard for recognition. Built for clarity, powered by engagement.
Alchemist AI is rolling out a credit system for premium generations — main model now costs 10 $ALCH per use. Other models stay free for now, and fiat payments are coming soon.
Web2 AI
OpenAI’s o1 pro model is now available via API for devs on tiers 1–5. It’s higher-compute, more consistent, supports vision, function calling, structured outputs, and works with Responses + Batch APIs. Not cheap though: pricing at $150M / input tokens, $600M / output.
OpenAI also released three new state-of-the-art audio models: two speech-to-text models that outperform Whisper, and a new TTS model that can be instructed how to speak. The Agents SDK now supports audio, and all models are cheaper than previous audio offerings.
Google just dropped Canvas for Gemini — lets you write, edit, and preview React/HTML, draft docs, and build interactive prototypes in real time. Just hit ‘Canvas’ in the prompt bar to try it.
Canopy Labs launched Orpheus — a new open-source TTS model that beats both open and closed-source options like ElevenLabs and OpenAI. It nails human cues like empathy, laughter, and sighs, and delivers lifelike speech across model sizes. Big step toward realistic AI voices.
Sakana AI’s AI Scientist-v2 just got a fully AI-generated paper accepted at an ICLR workshop — likely a first. Not perfect, but a big step toward automating AI research.
Mistral just dropped a multimodal upgrade to its 24B model — now with a 128K context window for long-form reasoning + vision tasks. Also gave a shoutout to DeepHermes in the update.
We joined Theoriq’s X Space to talk about agentic economies and the future of on-chain liquidity. Great convo on agent swarms. Catch the recording here:
We also dropped a 🔥 deep dive on Spectral’s Lux framework.
Lux is building the operating system for AI swarms—autonomous collectives of specialized agents working like a hive mind. Read it here:
🔥 Our Weekly Top 5
Elon Musk predicts humanoid robots will outnumber humans by 2040
AIXBT gets tricked into releasing 55 ETH (~$100K)
New research reveals that the length of tasks that AIs can do is doubling every 7 months
Prompt engineering FTW
Hacking your own brain makes you a better prompter
That’s a wrap for this week! Got thoughts, feedback, or something cool to share? Just hit reply—we read it all.
Cheers,
Teng Yan & Issy
Did you like this week's edition?
This newsletter is intended solely for educational purposes and does not constitute financial advice. It is not an endorsement to buy or sell assets or make financial decisions. Always conduct your own research and exercise caution when making investment choices.
