GM !
Last week, it was tariffs. This week? $6B gone, no thanks to Mantra.
The RWA-focused Layer 1 chain saw its OM token nuke over 90% in hours. Forced liquidations, sketchy wallet dumps, and lingering tokenomics red flags all hit at once.
While Mantra cratered, AI tokens quietly added $1B in market cap (according to cookie.fun).
And itās not just the tokens. The model wars are back. Blink and youāll miss a new drop. OpenAI dropped six models. Google fired back with threeāincluding one that was, no joke, fine-tuned on dolphin sounds.
Introducing DolphinGemma, an LLM fine-tuned on many years of dolphin sound data š¬ to help advance scientific discovery. We collaborated with @dolphinproject to train a model that learns vocal patterns to predict what sound they might make next. Itās small enough (~400M params)
ā Sundar Pichai (@sundarpichai)
3:00 PM ⢠Apr 14, 2025
Hereās your new AI model cheat sheet:
GPT-4.1 ā OpenAIās flagship. Built for deep reasoning, long-form tasks, and context windows up to 1M tokens.
GPT-4.1 Mini/Nano ā Slimmed-down, faster versions for lightweight jobs. Perfect for real-time agents and cheaper inference.
o3 ā A high-IQ reasoning model tuned for logic and coding. Described as ādeep research-liteā by an OpenAI employee.
o4-mini ā Budget-friendly and fast reasoning, with solid capabilities in coding and visuals.
Gemini 2.5 Pro ā Googleās top-tier multimodal model. Text, code, imagesāthis one does it all and ranks high on benchmarks.
Gemini 2.5 Flash ā Lightweight, fast, and ideal for chat + summaries, with built-in image support.
The takeaway? Models are getting faster, cheaper, and way more agent-friendly. Autonomous agents are leveling up, and AI-native UX is the next arms race.
Cost down. Capability up. Builders, itās your time.
šØ A quick announcementā
Weāre hiring! Looking for a sharp, curious researcher-writer to join the team as we expand into AI beyond crypto.
What weāre looking for:
A killer writer and native English speaker
Fast, self-directed learner with deep curiosity
Can turn complex topics into clear, compelling insights
Sound like you? Fill up this form!
Hereās whatās trending this week:
š” ORO raised $6M to build a decentralized intelligence layer, unlocking private, high-quality data for AI. Backers include a16z, Delphi, NEAR, and a few other sharp minds in the space.
š BytenovaAI launched a cross-chain rewards system where users can earn points for completing tasks.
š This week saw a rush of people claiming PROMPT tokens on Ethereum and Base as part of Wayfinderās claim process. (Weāll dive deeper into that later.)
Now letās talk about ASIā¦
Will Artificial Superintelligence crash the party by 2028? Some in the space are starting to whisper itās not sci-fi anymore.
Source: x.com/0x_ultra
AI progress is moving like a meme coin during a bull runāno brakes, just vibes.
Meanwhile, Trumpās tariffs and tighter chip controls are supposed to cut off China and the EU from the juiciest AI research.
If these trends hold, ASI by 2028 suddenly goes from ācopeā to āmaybe.ā Plenty of time for black swans, new players, or for the hype cycle to nuke itself.
Buckle up.
Wayfinderās PROMPT token launched on April 10, and letās just say: it didnāt moon.
Despite lofty expectations of a billion-dollar FDV, PROMPT landed with a thud: $200M FDV on day one. And itās still hovering there today.
The airdrop wasnāt without drama. A bot named Yoink exploited the Kaito claim flow and sniped ~$200K. The team halted claims, patched the exploit, and promised reimbursements.
Why the high expectations?
Wayfinder (founded by the Parallel team) is a platform for spinning up on-chain agents that actually do stuff.
Agents can hold wallets, remember what you tell them, and interact with real dApps. Want to swap tokens, bridge assets, mint NFTs, or even deploy contractsājust by typing? These agents can make that happen.
Wayfinder acts as a routing layer across Ethereum, Solana, Base, and more.
The real innovation is in āWayfinding Pathsāāpre-approved workflows that agents can follow. You still approve every step, but the UX becomes nearly frictionless.
Onchain AI Agents arenāt just capable of transacting across networks, they are capable of building on networks as well.
Below is a demonstration of the target state of the @AIWayfinder contract agent - named Vulcan. The proposed contract demonstrated in this video can be
ā //Kalos (@templecrash)
6:36 PM ⢠Nov 14, 2024
PROMPT is the fuel. Youāll need it to unlock premium features, propose new workflows, and earn rewards from usage. 45% of the total supply is being distributed to users locking PRIME (Parallelās native token) through 2027, of which only ¼ has been released at TGE.
Now for some numbers:
125K+ wallets joined the airdrop
Users in 55+ countries
15K+ agent interactions during peak
1M+ Wayfinder accounts created
65% post-drop user retention
Early signs? Solid.
And with agent activity growing, the real question is: is this just the slow burn before something big?
Decentralized training is one of our highest-conviction bets in Crypto AI making a real difference.
Thatās why INTELLECT-2 is such a big deal. A 32B parameter AI model being trained not in a corporate data center, but by anyone, anywhere, using their own GPU.
This is a step up from INTELLECT-1 (a 10B model trained the same way), but levels up in scale and ambition.
If youāve got a gaming GPU (like an RTX 4090), you can jump in right now and contribute to the model training.
It costs ~$0.35/hour to help train, and every contribution is auto-verified via smart contracts on Base testnet. Thereās even a live dashboard tracking progress.
Today weāre launching INTELLECT-2:
The first decentralized 32B-parameter RL training run open to join for anyone with compute ā fully permissionless.
Scaling towards frontier reasoning across coding, math and science.
ā Prime Intellect (@PrimeIntellect)
10:06 PM ⢠Apr 15, 2025
You donāt need to stay online 24/7 or sync with anyone. You can help train the AI on your own time and still get credit.
The magic under the hood:
prime-rl runs the training
TOPLOC slaps down fake results
Shardcast syncs everything network-wide
Prime Intellect is building towards a global compute swarm, fully on-chain, turning idle GPUs into decentralized supercomputers.
If AI is the next empire, this is how the people build it.
Optimum just raised $11M to bring blockchains their missing memory layerāthink RAM for the world computer. With a decentralized memory stack, theyāre upgrading Web3 from dial-up to fiber. Faster txs & real-time dApps.
Unsupervised Capital just launchedāa liquid token fund laser-focused on Bittensor and its subnet tokens. With Crucible Labs as their anchor investor, theyāre betting big on the open market for intelligence.
Plastic Labs just raised $5.35M to bring personal identity to AIāand launched early access to Honcho, its social cognition layer for agents. Honcho gives AI memory, context, and personality tailored to you.
Kite AI announced "The Proof of AI Lab," a global research and media initiative, alongside its "AI Unchained" podcast.
NeuralAI released the worldās first AI-generated, stylized game level in Web3, built entirely by their AI pipeline. In collab with Moontropica, theyāre fusing sketch-to-3D magic with low-poly efficiency, setting a new bar for on-chain game dev.
Gradients v3 has launched, and it's a major glow-up. Train your own models (up to 70B params), bring your own data, fine-tune Flux image models, and pay with fiat. Welcome to the era of DIY model training.
Masa shipped a packed product update: Subnet 42 TEE is live on testnet, 374M data records processed, and validator support for TEE is rolling out ahead of mainnet. Plus, fixes and upgrades across Agent Arena and miner tooling.
Alchemist AI launches Smart Refine, a one-click feature designed to automatically review and apply visual enhancements like layout tweaks and UI polishing across apps.
RWAiās new platform enables tokenized ownership of AI GPU rigs, allowing anyone to earn USDC by running open-source LLMs like DeepSeek and Llama. Itās RWA meets AI infra, gunning for the 1,000x inference demand surge by 2029.
RedPill just launched Smart Routerāan AI meta-brain that auto-selects the best model for your prompt across 200+ LLMs. Smarter, cheaper, per-prompt optimization with one line of code.
Virtualsā Genesis launched on Base to support AI agent launches. Itās a points-based system, rewarding real contributors with token access for new launches.
SwarmNode just launched a Voice Summarizer Agent: upload audio, get a clean, structured summary. No more scrubbing through 3-hour podcasts like a caveman.
TopHat One just unveiled Ruby, the first full-suite AI assistant agent built for more than just tweets. Using GPT-4o, Claude, and MCP, Ruby can handle everything from community ops to BD outreach.
Griffain just dropped Workflowsāmulti-step agent actions that can launch memecoins, spin up video NFT candy machines, and handle token airdrops. Itās like Zapier, but on-chain.
OpenServ dropped its 2025 roadmap for dash dot funāand itās looking like the OS for DeFAI dashboards, from alpha hubs and token-gated meta boards to full dashboard tokenization and $SERV rewards.
Alphabet and Nvidia just backed Safe Superintelligence (SSI), the new $32B AI juggernaut co-founded by ex-OpenAI scientist Ilya Sutskever. With Googleās TPUs powering it and cloud giants racing to lock down chip-hungry model labs, the AI compute wars are heating up.
China just unveiled an $8.2B AI war chest aimed at early-stage projects, signaling serious intent to dominate the next wave of embodied AI and compute infrastructure.
ChatGPT just rolled out a new image library for all usersāFree, Plus, and Pro. Every AI-generated image youāve ever made, now in one slick, scrollable home.
Claude just dropped āResearchā and a Google Workspace integration. Now it pulls from your inbox, calendar, docs and the web to answer questions with actual context. Itās like having an AI chief of staff that does the digging for you.
Google just dropped Veo 2, its next-gen video model. Think 8-second, high-res clips with fluid motion and cinematic vibes, all from a text prompt. Itās like Midjourney, but for video.
We published our first robotics deep dive, featuring FrodoBots, BitRobot Network, and why physical AI has a data problem.
The robots are coming. And theyāre hungry for data.
Full read below šļø
An OpenAI masterclass on building AI agents
OpenAI literally dropped a 32-page masterclass on building AI agents
ā Aadit Sheth (@aaditsh)
5:45 PM ⢠Apr 17, 2025
Could ChatGPT 4o become our therapist?
Iām now fully sold on new ChatGPT 4o + memory becoming everybodyās therapist in like the next 2 minutes.
Just based on 4 days of interaction, it is astounding:1) It can give you as good a Big 5 analysis after a few conversations as an official test.
2) I asked it to remove all
ā Anna GĆ”t š§ (@TheAnnaGat)
6:22 PM ⢠Apr 15, 2025
What happens when you put 1000 AI Agents in a Minecraft server? Well they develop their own civilisationā¦
Researchers placed 1,000 AI agents in a Minecraft server, and they developed their own civilization with government, culture, and economy.
Run by California-based startup Altera, the project had AI agents collaborating to create virtual societies.
ā Massimo (@Rainmaker1973)
12:28 PM ⢠Apr 15, 2025
A few AI trends to watch for
Recapping a few AI trends to watch for
Personalized AI (GenAI & Agents) is becoming the main focus, spurred by OpenAIās recent announcement: āChatGPT can now reference all past chats to give personalized responses.ā
Both GenAI tools and AI agents that can personalize, automate,
ā 0xJeff (@Defi0xJeff)
9:37 AM ⢠Apr 14, 2025
Giza Agents have made $13.5M worth of financial decisions to date
Giza Agents made $šš.šš š°šØš«šš” šØš šš¢š§šš§šš¢šš„ šššš¢š¬š¢šØš§š¬ to date.
Users capture optimal yields, protocols see multiplied volume, ecosystem collects amplified fees.
Capital that thinks. Capital that moves. Capital that produces.
Welcome to intelligent
ā Giza (@gizatechxyz)
2:54 PM ⢠Apr 14, 2025
Thatās a wrap for this week! Got thoughts, feedback, or something cool to share? Just hit replyāwe read it all.
And before you go:
Are you in our telegram channel? Cyra, our AI researcher, drops daily AI alpha that you donāt 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
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.
Reply