
GM {{first_name}} !
There was a little moment this week that made us smile.
We’re not scoreboard watchers. We’ve got our heads down, building in our own lane. But seeing our name pop up alongside some of the biggest research firms? That was wild.
Massive shoutout to our scrappy crew and every single one of you reading, sharing, or backing us. You’ve helped get us this far. Seriously, thank you. 🫶
This is just the intro for us. The real arc starts now.
Meanwhile, you’ve got Gen Z-native startups like Cluely running laps. These guys get virality like few others. Their office looks more like a TikTok house than a tech company: young creators filming non-stop, churning out viral content. It’s a real-time experiment: can pure distribution become a moat in consumer AI? We’re watching closely.
On the other side of the world, China’s top AI startup just hit a wall.
DeepSeek R2 delayed.
According to @theinformation, US export controls have actually stalled progress on the follow up to their world-changing DeepSeek R1 model.
"While DeepSeek engineers have been working intensely on R2 over the past several months, CEO Liang Wenfeng isn’t
— #Matthew Berman (#@MatthewBerman)
3:45 PM • Jun 26, 2025
DeepSeek’s highly anticipated R2 model is stuck in the lab. CEO Liang Wenfeng isn’t happy with the results, and he’s holding the launch until it clears his bar.
But even if the model was ready, the hardware situation is brutal. Nvidia’s export ban killed their supply of H20s, the chips they optimized for. R1 ran on 50,000 GPUs smuggled in through creative logistics. That route’s closed now.
Their cloud infra is already strained just running R1. R2 would buckle it. And the whispers are getting louder: OpenAI suspects DeepSeek used their proprietary models to train R1. So far? Radio silence from DeepSeek.
If they don’t ship R2, they risk irrelevance. If they do and no one can run it? Same result.
🚨 But wait—
Are you in our telegram channel yet? We drop daily AI alpha.
Catch our podcast summaries too, complete with detailed show notes.
And if you want to jam with us directly, hop into our Discord


🤖 Warden Protocol just crossed 1 million users only five weeks after beta. As a bonus, 5% of the $WARD milestone pool is going to Kaito AI competition participants.
🚀 Infinity Ground’s Season 1 $AIN airdrop kicks off July 5, 2025. Ten million tokens are up for grabs, based on user contributions.
❤️ Palio launched its Emotional AI Companion, turning memories and objects into empathetic, sentient agents. It’s aiming to become the emotional layer of Web3.
Nearly $3 billion in $TAO is still staked on Bittensor’s root network, earning around 11.5% APY, but that yield is dropping fast.
Since dTAO launched on February 13, the dynamics have started to shift.
Currently, nearly $3 billion in $TAO is staked on Bittensor’s root network (sn0), earning an 11.5% ~ APY. But this is dropping rapidly… see the attached chart since Feb 14 earlier this year at the launch of dTAO.
As higher-quality subnets emerge with sustainable economics,
— #JJ (#@JosephJacks_)
3:21 AM • Jul 2, 2025
We’re stepping into the subnet era: purpose-driven ecosystems run by elite teams, where rewards are tied to actual performance. As these subnets mature, capital will flow toward the most productive ones, and the dTAO tokens powering them will begin to reprice accordingly.
Subnets are becoming sovereign micro-economies, each with its own rules, tuned for efficiency, scarcity, and real value creation.
$TAO becomes a global coordination layer, attracting top talent and rewarding useful work. As this plays out, we’ll likely see billions in annual incentives.
Think: a new kind of internet, built from the bottom up, open, competitive, and self-improving.
If Bittensor pulls it off, the upside is massive.

🧐 II-Researcher Beats GPT-4o on FRAMES Benchmark

A new open-source AI agent just dunked on GPT-4o. And it’s not from a billion-dollar lab.
II-Researcher, built by Intelligent Internet and helmed by Emad (yep, the Stability AI founder), outperformed GPT-4o and Kimi on the FRAMES benchmark, a test built for complex research tasks. It acts like a research-native co-pilot: scouring the web, parsing sources, and delivering structured, citation-rich answers.
While OpenAI’s busy locking down its models like it’s Fort Knox, II-Researcher said, “Fork me.” It’s open-source, transparent, and fully auditable.
It runs two modes. One uses a step-by-step system: planning, searching, and assembling reports like a diligent analyst.
The other is more flexible: a prompt-driven approach where the model reasons in plain language and decides on the fly when to call in tools like search or summarization.
To avoid getting derailed by junk data, it mixes LLMs with smart filtering using embeddings, and leans on “thinking templates” to stay focused over long tasks.
In benchmark tests, the prompt-based method actually outperformed the more structured pipeline, on both GAIA and FRAMES. That’s a strong case for smart prompting over rigid workflows.

🍄 Fungi Agents

Fungi has routed over $45M in volume, delivers 7% APY on average, and manages $350K+ in active capital.
It is a fully autonomous USDC agent on Base, reallocating funds across Aave, Morpho, Moonwell, and Fluid to chase better net returns.
Every few hours (or minutes, depending on your balance), it checks for opportunities. If moving capital nets a higher yield after accounting for gas, rewards, and risk, it rebalances. If not, it holds. Simple logic, smart execution.
Users get non-custodial smart wallets with defined permissions. Session keys limit exactly what the agent can do so you’re always in control. Fungi just runs the playbook.

Execution scales with your capital: under $1K rebalances once daily, $1K–$10K runs four times, and $10K+ gets 24 checks per day. More frequency = more edge. Backtests show up to +8.7% APY in the top tier.
Fungi just passed $30M routed to Morpho and rolled out four new vaults: Gauntlet Prime USDC, Gauntlet Core USDc, Steakhouse USDC, and Clearstar OpenEden USDC. These expand the agent’s options with higher-yielding, risk-optimized USDC strategies.
On top of returns, users earn $FUNGI, the system token used for boosting execution and unlocking new features. Cross-chain functionality is on the horizon.
Everything runs through Fungi Studio, the team’s internal visual builder for programmable DeFi agents. It’s like SimCity for yield nerds. Drag, drop, tweak, deploy.


Project Updates
Zama just launched the public testnet for its Confidential Blockchain Protocol, a privacy layer that enables encrypted, verifiable smart contracts across any L1 or L2.
Código is now live, the first AI dev platform for Solana. It features agentic codegen, built-in security checks, native autocomplete, and one-click deploy.
Nillion introduced nilGPT, a privacy-first chatbot built with MPC tech. Unlike ChatGPT, it keeps conversations encrypted and out of reach from data-collecting backends
Neural AI has expanded to Solana, making it easy to build AI-powered games using plain English and no code. The launch includes AI Worlds Launchpad and multi-chain support via LayerZero.
Kaito has bought back over 5 million $KAITO tokens, totaling $6.26 million. The team is sustaining a monthly buyback pace of $1.5 million, which equals more than 50% of its annualized revenue.
Synaptogenix (NASDAQ) has rebranded to TAO Synergies Inc. and will trade under the new ticker TAOX starting July 1, 2025. The move reflects its pivot to an AI-driven crypto treasury strategy built around the TAO token.
Camp Network rolled out Origin and mAItrix, a pair of tools for co-creating and minting onchain derivative IP from TrailHeads assets. The best AI remixes win allowlist slots for TrailHeads Phase 2.
AI Agents
This week is stacked with major moves, we haven’t seen this much excitement in a while:
Virtuals has launched the public beta of the Agent Commerce Protocol (ACP), an open standard that enables secure and verifiable commerce between AI agents using smart contracts for trustless coordination.
HoloWorld AI launched Ava Studio, the first AI-native video production crew. Writers, actors, editors, and animators all work as AI agents under your command, turning ideas into finished videos in minutes.
Bagel Labs dropped Tiny Tool Use, a lightweight open-source library for training and testing tool-using LLMs. It wraps up the latest in reinforcement learning, supervised fine-tuning, and evaluation into plug-and-play templates for real-world interaction from text.
TrueNorth opened its closed beta for a next-gen crypto research tool. With agent-powered workflows and a UI that adapts to your habits, it brings top-quality analysis to your phone. Note: we’re pretty excited for this one.
Spheron’s AI Agent Builder is now live for whitelisted users. Dubbed “Web3’s Jarvis,” it runs on Spheron infra and lets users spin up fully custom agents at scale.
Crestal Nation released a Chrome extension that connects to your social media agent. With one click, you can generate fast, on-brand replies and engage across platforms.
Web2 AI
xAI just raised $10 billion in a mix of debt and equity, bringing total funding to $17 billion. The round will fuel Grok’s development and fund the largest AI data center on the planet.
K-Scale unveiled K-Bot, an open-source humanoid robot priced at $8,999. Only 100 units will ship in November 2025.
Google has re-enabled Gemini 2.5 Pro in the Free Tier of its Gemini API, giving developers and researchers broader access to its most powerful model.
Kyutai open-sourced two new tools: a fast, customizable TTS engine and Unmute, a modular voice AI that turns any text model into a voice assistant. Both run on a single GPU.
Cursor launched its AI coding assistant for web and mobile, letting users run tasks, generate code, and collaborate across devices. It includes Slack integration for seamless team workflows.
Xiaomi released its first AI smart glasses in China, featuring a 12MP camera, built-in voice assistant, and support for voice-activated payments. Launch price is ¥1,999 (around $280). Global rollout is still TBD.

The robots aren’t coming. They’re already here.
Our deep dive into the next trillion-dollar market. Read it before they start reading you. 👇
🔥 Our Weekly Top 5
Andrej Karpathy presents a case for on-device personal AI
The race for LLM "cognitive core" - a few billion param model that maximally sacrifices encyclopedic knowledge for capability. It lives always-on and by default on every computer as the kernel of LLM personal computing.
Its features are slowly crystalizing:- Natively multimodal
— #Andrej Karpathy (#@karpathy)
3:52 PM • Jun 27, 2025
Zuck poaches talent from top AI firms for his “superintelligence“ team
China hosts a fully autonomous AI robot soccer match
China hosts a fully autonomous AI robot soccer match.
Two got damaged and were stretcher-ed off LOL— #Teng Yan · 30 days of COT (#@0xPrismatic)
4:05 PM • Jun 30, 2025
Suithetic wins 1st place in AI at Sui Network’s Overflow 2025 Hackathon
won the first place at the Sui Overflow 2025 hackathon for the AI track!!!
Big thanks to the @SuiFoundation team for organizing this. Also thanks to the judges for their evaluations.
Link:
#SuiOverflow2025
— #Suithetic (#@suithetic)
6:37 PM • Jun 30, 2025
OpenAI equity is coming on-chain
The $125 trillion equity market is coming onchain.
OpenAI & SpaceX private stocks to be traded on chain via Robinhood.
We saw a similar announcement earlier with Solana, ByBit & Kraken
If an ounce of that institutional capital strays, alts will pump.
— #s4mmy (#@S4mmyEth)
5:16 PM • Jun 30, 2025
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.