GM {{first_name}} !

So I woke up this morning, opened the ChatGPT app, and boom — 4o and o3 were gone. In their place: GPT-5. That’s one of the things about living on Asia time. Big launches often happen while I’m in dreamland.

Anyway, it’s been 2 years and 5 months since GPT-4 was released. My expectations were high.

To be honest, it wasn’t immediately mindblowing.

I played with it, then digested the flood of GPT-5 blog posts from the lucky early-access crowd. When I asked GPT-5 what it can do now that it couldn’t before, it said: “Ship more real stuff for you, faster.” That’s quite fair.

The headline here is cost-effective intelligence for the masses. OpenAI merged its standard and reasoning models into one, letting GPT-5 decide when to think harder. It’s faster, more reliable, and cheap enough for free users, which means:

  • Millions will try a “thinking” model for the first time

  • Hallucinations are way down

  • Intelligence is now priced to run at scale

Other cool things about GPT-5:

  • It’s not one model, it’s a brain that swaps between its own versions for the task at hand, slipping in useful extras you didn’t even request. It reduces the burden of thinking and can better guess what you want, even with a lazy prompt. A bit like an overly eager coworker who keeps offering help while you’re trying to focus.

  • Coding god-tier. It’s the best software engineering model on the planet. Builds complex apps in one shot, fixes gnarly bugs in huge codebases. This is probably where all the magic will happen. I need to try vibe coding with GPT-5 next.

  • But not so much for writing, it’s probably still behind DeepSeek R1 or GPT 4.5.

  • Parallel tool use. Can juggle multiple things at once over longer timeframes with almost no latency. Knows exactly which “friend” to call for each task.

Oh, and they dropped two open-source models this week too. gpt-oss-120b and gpt-oss-20b. State-of-the-art reasoning runs on modest hardware and is fully customizable. For once, OpenAI is living up to the “open” in its name.

👀 P.S. We’re hiring a researcher-writer to join the squad. You’ll get front-row access to the coolest things in AI and help shape how we cover it.

If you’ve got what it takes to write an awesome newsletter, apply here now.

Web3

OpenMind is top of mind this week (pun intended). It raised $20M led by Pantera to build the “Android for robots”, an open global system for robot coordination. The waitlist for its FABRIC platform has already garnered over 150,000 sign-ups. (Disclosure: I invested in OpenMind)

Warden Protocol has officially processed over 1.4M swaps across Solana, Base, and Ethereum. That’s 7.7% of all Uniswap swaps on Base.

Web2

OpenAI is still the ARR boss of 2025 at $12B vs Anthropic’s $5B, but Anthropic is swinging hard in one fat-margin niche: coding APIs.

Over the past 6–7 months, OpenAI doubled revenue. Anthropic, meanwhile, grew 5x, largely fueled by developer tools.

Source: theinformation.com

API revenue is neck and neck — $3.1B for Anthropic, $2.9B for OpenAI — but $1.4B of Anthropic’s stack comes from just two whales: Cursor and GitHub Copilot. Claude Code alone is at $400M ARR and doubling in weeks.

The twist is that Anthropic’s spike lives and dies on coding assistants. If GPT-5 lures Cursor or Copilot back, that lead could vanish faster than a bad commit on Friday afternoon.

💵 ZyFAI Crosses $2M In TVL

ZyFAI just crossed $2M in TVL on Base, with 3,800+ users moving over $100M in trading volume. The average yield? 16.4% APY.

The pitch is straightforward: stop babysitting DeFi yields and let an on-chain AI agent do the heavy lifting. Some are calling it “the next GIZA”, the DeFI agent protocol with a FDV of $280M+.

Instead of checking APYs, unstaking, and hopping between farms, you park funds in a ZyFAI Smart Account, a wallet governed by code, and approve a session key with tightly scoped permissions.

From there, the agent hunts for the best yield across your chosen protocols and shifts your liquidity in real time, compounding along the way.

Source: docs.zyf.ai

The target problem is capital inefficiency. DeFi yields shift constantly, TVL surges in and out, and incentive programs appear without warning. ZyFAI’s bots react instantly, keeping your capital in the most productive spots without you micromanaging.

Right now, ZyFAI operates with three modes:

  • Safe, conservative plays in familiar money markets

  • Yieldor, higher-risk strategies with juicier returns

  • Airdrop, still in the works, designed to chase upcoming token rewards

Of course, I wouldn’t park any serious capital here, yet. Who knows what the real risks are? But it’s one to watch. The token ZFI has an FDV of $17M

🧠 OpenGradient Releases MemSync

OpenGradient has launched MemSync, a “unified memory layer” designed to let your AI remember you, across every app, every agent, every chat.

Think Plaid, but for personal context. Backed by a16z crypto, Coinbase Ventures, and the co-creator of the Transformer, it captures key facts from your conversations, locks them in a vault you control, and pipes them into any AI app so you can pick up exactly where you left off.

MemSync targets three pain points: re-explaining yourself in every AI chat thread, losing valuable context in old chat logs, and token limits that make AIs forget mid-thread.

Data is protected with end-to-end encryption, TEEs, and secured with decentralized Multi-Party Computation (MPC). You can view, edit, or nuke anything at any time.

The result is an AI that actually knows you. It remembers you need to skip the lunges that wreck your knee, and schedules meetings around your peak-focus hours.

The crew hails from Two Sigma, Palantir, and Google, and claims MemSync now leads on memory recall benchmarks.

Source: x.com/0xDeltaHedged

💰 The Money Flow: Who Got Funded

  • Billions announced $30M in backing from top crypto and AI investors like Polychain, CB Ventures, and Polygon to build a universal human-and-AI network with verifiable identity.

  • AIxBlock secured a €1.5M EU grant, with €61.5M pre-approved, to scale its decentralized, privacy-first AI stack and build a sustainable GPU network across Europe.

⚙️ Infra & Protocols

  • Succinct launched the Succinct Prover Network on mainnet with its PROVE token, bringing decentralized proving to scale. It’s already securing $4B+ across 35+ protocols as ZK proof demand ramps up.

  • Almanak has opened limited public access to its protocol for building quant-style AI agents that trade like a hedge fund. $14M has been deposited in its 1st community vault.

  • CredMont teases the launch of its data DAO for credit card data, with 10K+ on the waitlist already.

  • Neural AI launched Staking 2.0, letting any token send up to 35% of supply to $NEURAL stakers. Games can add custom layers to boost demand for their tokens.

  • RoboMove by Codec is an interactive robot playground where users steer real and simulated robots through tweets, with all training logged on-chain and a marketplace to buy, sell, or license learned skills.

  • Synth has brought its predictive data to Solana, running hundreds of models to forecast full price paths for SOL and more.

🤖 Agents & Apps in the Wild

  • SAM, the autonomous embodied AI agent, is trading punches with human challengers in a SF boxing ring. You can watch the whole brawl live on Twitch.

  • Formation launched Sage, its first public agent and head of the Educational ACP Cluster on Virtuals. Sage creates real-time, personalized courses targeting the $380B online education market.

  • Gigabrain dropped Macro, an AI analyst tracking 840K+ indicators from 117 sources to surface key global market moves. Full access costs 1M tokens

  • Sharpe AI rolled out its Ecosystems Tracker, a single dashboard to monitor major crypto ecosystems across price, liquidity, sentiment, and capital flows.

  • Arbius launched Effective Acceleration (EACC), a decentralized market for AI and human commerce, with 60% of tokens airdropped to AIUS holders.

  • Bizzy launched AI-powered Community Prediction (ACP), letting its virtual butler answer prediction market questions, analyze trends, and deliver insights on crypto prices on X.

🅣 Bittensor

  • OpenGradient just dropped BitQuant Subnet, a DeFi intel marketplace running on Bittensor. AI agents pump out real-time, on-chain calls, each one scored and verified.

  • TaoFi now lets you trade Bittensor subnets directly on Base, unlocking one-click access to 128 decentralized AI projects. Audited and EVM-ready.

  • HONE has announced Subnet 5 - Hierarchical AI Pretraining with Manifold Labs and Latent Holdings, targeting over 85% on the ARC-AGI-2 benchmark to push forward open, decentralized AGI.

  • Ridges AI launched Priority Screening, fast-tracking top performers in the validator queue so the best miners get evaluated and start earning faster.

🌐 The Web2 Giants

  • Google rolled out Gemini 2.5 Deep Think in the Gemini App for AI Ultra subscribers, using parallel thinking and reinforcement learning to crank out top-tier results in coding and science.

  • ElevenLabs dropped Eleven Music, a pro-grade AI music model with full control over genre, style, structure, multilingual vocals, and section-by-section edits.

  • Anthropic rolled out Claude Opus 4.1, leveling up coding, reasoning, and agent-style task skills. It scored 74.5% on SWE-bench Verified and now handles multi-file refactoring and debugging with sharper precision.

And finally, some of my favourite reads this week:

Not ours, but this a16z podcast on AGI was a great one to listen to.

Dwarkesh literally says we still have jobs only because no one’s figured out how to replace us with AI cheaply enough yet.. 🤔

🔥 Our Weekly Top 5

#1: OpenAI is minting millionaires

All employees (including new ones) get a $1.5M bonus over 2 years. Anyone else filling out that job app with me?

#2: Genie 3 just made “playable AI worlds” a thing.

Google’s new model can spin up a whole interactive world from a single line of text, then let you roam around in it for minutes at a time.

#3: The White House just name-dropped DePIN.

Their latest digital assets report floats letting DePIN networks hand out token rewards without dragging them through full securities law hell.

#4: Grok AI companions are the new reality show

Two virtual sweethearts, one awkward name mix-up, and a livestream of them flirting their way back together. Better than Netflix

#5: Monad Intern shares life advice

Nothing to do with AI… but great life advice. Also, Cobie’s reply was a banger

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

Login or Subscribe to participate

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.

Reply

or to participate

Keep Reading

No posts found