Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป (@Apeteks1) header
Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป (@Apeteks1) profile picture

AI training ร— web3, unfiltered ๐Ÿง 

Joined October 2024
Photo by Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป (@Apeteks1)Photo by Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป (@Apeteks1)Photo by Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป (@Apeteks1)Photo by Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป (@Apeteks1)Photo by Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป (@Apeteks1)Photo by Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป (@Apeteks1)Photo by Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป (@Apeteks1)Photo by Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป (@Apeteks1)Photo by Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป (@Apeteks1)Photo by Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป (@Apeteks1)
As a trader, I had always doubted the possibilities for one to learn leverage trading & risk management without the pressure of real monetary loses๐Ÿคท๐Ÿฟโ€โ™‚๏ธ. That mentality changed when I found out about Top trader on Ape Chain ๐Ÿงต๐Ÿ‘‡ @apecoin @Toptrader_xyz @ThankApe
19
8
35
6,889
GM CT โ˜€๏ธ Been away for a while, but weโ€™re back. ๐Ÿ’ช Took some time off, reset, handled a few thingsโ€ฆ now itโ€™s time to lock back in. Let's get the progress started ๐Ÿ’ฏ
1
2
44
Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป retweeted
GM CT โ˜€๏ธ New day, new block, new opportunity. Keep the mind sharp, stay curious, and keep shipping. The timeline will chase narratives, weโ€™ll chase knowledge. ๐Ÿง ๐Ÿ“ˆ Have a productive day, chads. ๐Ÿค
Every mainstream AI chatbot keeps a log of what you asked it, and decides what you're allowed to ask in the first place. Day 21 of my AI & Web3 tool sprint looks at the protocol betting people want the opposite. ๐Ÿ”’๐Ÿค– Tool: Venice AI @VeniceAI โ€“ a privacy-first AI platform founded by Erik Voorhees (ShapeShift) and Jesse Proudman, positioned as a private, unrestricted alternative to ChatGPT. Routes prompts across 200+ open-source and proprietary models for text, image, video, and audio through one interface. The privacy architecture: prompts hit an encrypted proxy that strips identifying data before reaching inference providers, chat data stays in your browser, not their servers. They've since added TEE and end-to-end encryption modes, hardware-attested and independently verifiable, not just a promise in the terms of service. The token layer: VVV isn't a payment token, it's an access key. Stake it, and you get a daily pro-rata share of Venice's total inference capacity through a second token, DIEM, where 1 DIEM equals $1 of API credit, in perpetuity. Heavy users effectively own their compute instead of renting it per request. The scale: raised $65M in July 2026 at a $1B valuation, crossed 3 million users, no presale, no VC token allocation, self-funded until this raise. My Honest Review ๐Ÿ” The privacy engineering here is genuinely serious, not just marketing copy, hardware-attested verification is a real technical commitment most "private AI" products don't bother with. But "uncensored" is a double-edged sword by design, security researchers have flagged Venice as a route to generating content mainstream labs block for real safety reasons. That's not a hypothetical risk, it's the direct tradeoff of the entire premise. My Take ๐Ÿง  Privacy and safety guardrails aren't the same conversation, even though they get bundled together constantly. Venice is proof that you can solve one seriously (verifiable, hardware-attested privacy) while the other stays genuinely unresolved. Worth understanding both halves before deciding which side of that tradeoff you're comfortable with. Freedom and safety were never going to be free of tension. Venice just made that tension impossible to ignore. ๐Ÿ’ฏ
1
1
1
32
GM CT โ˜€๏ธ New day, new block, new opportunity. Keep the mind sharp, stay curious, and keep shipping. The timeline will chase narratives, weโ€™ll chase knowledge. ๐Ÿง ๐Ÿ“ˆ Have a productive day, chads. ๐Ÿค
Every mainstream AI chatbot keeps a log of what you asked it, and decides what you're allowed to ask in the first place. Day 21 of my AI & Web3 tool sprint looks at the protocol betting people want the opposite. ๐Ÿ”’๐Ÿค– Tool: Venice AI @VeniceAI โ€“ a privacy-first AI platform founded by Erik Voorhees (ShapeShift) and Jesse Proudman, positioned as a private, unrestricted alternative to ChatGPT. Routes prompts across 200+ open-source and proprietary models for text, image, video, and audio through one interface. The privacy architecture: prompts hit an encrypted proxy that strips identifying data before reaching inference providers, chat data stays in your browser, not their servers. They've since added TEE and end-to-end encryption modes, hardware-attested and independently verifiable, not just a promise in the terms of service. The token layer: VVV isn't a payment token, it's an access key. Stake it, and you get a daily pro-rata share of Venice's total inference capacity through a second token, DIEM, where 1 DIEM equals $1 of API credit, in perpetuity. Heavy users effectively own their compute instead of renting it per request. The scale: raised $65M in July 2026 at a $1B valuation, crossed 3 million users, no presale, no VC token allocation, self-funded until this raise. My Honest Review ๐Ÿ” The privacy engineering here is genuinely serious, not just marketing copy, hardware-attested verification is a real technical commitment most "private AI" products don't bother with. But "uncensored" is a double-edged sword by design, security researchers have flagged Venice as a route to generating content mainstream labs block for real safety reasons. That's not a hypothetical risk, it's the direct tradeoff of the entire premise. My Take ๐Ÿง  Privacy and safety guardrails aren't the same conversation, even though they get bundled together constantly. Venice is proof that you can solve one seriously (verifiable, hardware-attested privacy) while the other stays genuinely unresolved. Worth understanding both halves before deciding which side of that tradeoff you're comfortable with. Freedom and safety were never going to be free of tension. Venice just made that tension impossible to ignore. ๐Ÿ’ฏ
1
1
1
32
Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป retweeted
Every mainstream AI chatbot keeps a log of what you asked it, and decides what you're allowed to ask in the first place. Day 21 of my AI & Web3 tool sprint looks at the protocol betting people want the opposite. ๐Ÿ”’๐Ÿค– Tool: Venice AI @VeniceAI โ€“ a privacy-first AI platform founded by Erik Voorhees (ShapeShift) and Jesse Proudman, positioned as a private, unrestricted alternative to ChatGPT. Routes prompts across 200+ open-source and proprietary models for text, image, video, and audio through one interface. The privacy architecture: prompts hit an encrypted proxy that strips identifying data before reaching inference providers, chat data stays in your browser, not their servers. They've since added TEE and end-to-end encryption modes, hardware-attested and independently verifiable, not just a promise in the terms of service. The token layer: VVV isn't a payment token, it's an access key. Stake it, and you get a daily pro-rata share of Venice's total inference capacity through a second token, DIEM, where 1 DIEM equals $1 of API credit, in perpetuity. Heavy users effectively own their compute instead of renting it per request. The scale: raised $65M in July 2026 at a $1B valuation, crossed 3 million users, no presale, no VC token allocation, self-funded until this raise. My Honest Review ๐Ÿ” The privacy engineering here is genuinely serious, not just marketing copy, hardware-attested verification is a real technical commitment most "private AI" products don't bother with. But "uncensored" is a double-edged sword by design, security researchers have flagged Venice as a route to generating content mainstream labs block for real safety reasons. That's not a hypothetical risk, it's the direct tradeoff of the entire premise. My Take ๐Ÿง  Privacy and safety guardrails aren't the same conversation, even though they get bundled together constantly. Venice is proof that you can solve one seriously (verifiable, hardware-attested privacy) while the other stays genuinely unresolved. Worth understanding both halves before deciding which side of that tradeoff you're comfortable with. Freedom and safety were never going to be free of tension. Venice just made that tension impossible to ignore. ๐Ÿ’ฏ
Training a frontier AI model has always meant one thing: tens of thousands of GPUs, humming in unison, inside a single data center, owned by a company with a market cap bigger than most countries. Day 20 of my AI & Web3 tool sprint looks at the team trying to break that requirement entirely. ๐Ÿง โš™๏ธ Tool: Prime Intellect @PrimeIntellect ; a decentralized platform that aggregates GPU compute from around the globe and turns it into a unified network for training, evaluating, and deploying large language models. Founded by Vincent Weisser and Johannes Hagemann. What makes this different from just "another compute marketplace": Prime Intellect has actually trained models at meaningful scale, not just theorized about it. INTELLECT-2, a 32B-parameter model, was trained across globally distributed GPUs, no single data center required, and released fully open. The stack: 1๏ธโƒฃ GPU Marketplace: Rent compute on demand, from a single card to 256+ GPU clusters, aggregated across dozens of datacenters worldwide. 2๏ธโƒฃ Environments Hub: 2,500+ open-source reinforcement-learning environments to train and evaluate models against. 3๏ธโƒฃ Hosted Training & Lab: The full post-training pipeline, RL, evaluation, deployment, without needing to own or manage a GPU cluster yourself. The backing: $130M Series A in July 2026 at a $1B valuation, led by Radical Ventures, with Nvidia Ventures, Intel Capital, and Dell Technologies Capital all participating. Over $150M raised in under two years. My Honest Review ๐Ÿ” This is the piece that was missing from everything I've covered so far. Compute access (Day 19) means nothing if you still can't actually train a competitive model on it. Prime Intellect isn't promising decentralized training works, they've already shipped proof, twice. My Take ๐Ÿง  The gap between open-source and closed frontier labs has mostly been a compute-and-coordination gap, not a talent gap. When training infrastructure itself becomes decentralized and accessible, that gap stops being a moat and starts being a temporary head start. The frontier lab moat was never intelligence. It was always infrastructure. That's what's actually up for grabs now.
1
2
1
2
135
Every mainstream AI chatbot keeps a log of what you asked it, and decides what you're allowed to ask in the first place. Day 21 of my AI & Web3 tool sprint looks at the protocol betting people want the opposite. ๐Ÿ”’๐Ÿค– Tool: Venice AI @VeniceAI โ€“ a privacy-first AI platform founded by Erik Voorhees (ShapeShift) and Jesse Proudman, positioned as a private, unrestricted alternative to ChatGPT. Routes prompts across 200+ open-source and proprietary models for text, image, video, and audio through one interface. The privacy architecture: prompts hit an encrypted proxy that strips identifying data before reaching inference providers, chat data stays in your browser, not their servers. They've since added TEE and end-to-end encryption modes, hardware-attested and independently verifiable, not just a promise in the terms of service. The token layer: VVV isn't a payment token, it's an access key. Stake it, and you get a daily pro-rata share of Venice's total inference capacity through a second token, DIEM, where 1 DIEM equals $1 of API credit, in perpetuity. Heavy users effectively own their compute instead of renting it per request. The scale: raised $65M in July 2026 at a $1B valuation, crossed 3 million users, no presale, no VC token allocation, self-funded until this raise. My Honest Review ๐Ÿ” The privacy engineering here is genuinely serious, not just marketing copy, hardware-attested verification is a real technical commitment most "private AI" products don't bother with. But "uncensored" is a double-edged sword by design, security researchers have flagged Venice as a route to generating content mainstream labs block for real safety reasons. That's not a hypothetical risk, it's the direct tradeoff of the entire premise. My Take ๐Ÿง  Privacy and safety guardrails aren't the same conversation, even though they get bundled together constantly. Venice is proof that you can solve one seriously (verifiable, hardware-attested privacy) while the other stays genuinely unresolved. Worth understanding both halves before deciding which side of that tradeoff you're comfortable with. Freedom and safety were never going to be free of tension. Venice just made that tension impossible to ignore. ๐Ÿ’ฏ
Training a frontier AI model has always meant one thing: tens of thousands of GPUs, humming in unison, inside a single data center, owned by a company with a market cap bigger than most countries. Day 20 of my AI & Web3 tool sprint looks at the team trying to break that requirement entirely. ๐Ÿง โš™๏ธ Tool: Prime Intellect @PrimeIntellect ; a decentralized platform that aggregates GPU compute from around the globe and turns it into a unified network for training, evaluating, and deploying large language models. Founded by Vincent Weisser and Johannes Hagemann. What makes this different from just "another compute marketplace": Prime Intellect has actually trained models at meaningful scale, not just theorized about it. INTELLECT-2, a 32B-parameter model, was trained across globally distributed GPUs, no single data center required, and released fully open. The stack: 1๏ธโƒฃ GPU Marketplace: Rent compute on demand, from a single card to 256+ GPU clusters, aggregated across dozens of datacenters worldwide. 2๏ธโƒฃ Environments Hub: 2,500+ open-source reinforcement-learning environments to train and evaluate models against. 3๏ธโƒฃ Hosted Training & Lab: The full post-training pipeline, RL, evaluation, deployment, without needing to own or manage a GPU cluster yourself. The backing: $130M Series A in July 2026 at a $1B valuation, led by Radical Ventures, with Nvidia Ventures, Intel Capital, and Dell Technologies Capital all participating. Over $150M raised in under two years. My Honest Review ๐Ÿ” This is the piece that was missing from everything I've covered so far. Compute access (Day 19) means nothing if you still can't actually train a competitive model on it. Prime Intellect isn't promising decentralized training works, they've already shipped proof, twice. My Take ๐Ÿง  The gap between open-source and closed frontier labs has mostly been a compute-and-coordination gap, not a talent gap. When training infrastructure itself becomes decentralized and accessible, that gap stops being a moat and starts being a temporary head start. The frontier lab moat was never intelligence. It was always infrastructure. That's what's actually up for grabs now.
1
2
1
2
135
Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป retweeted
GM CT โ˜€๏ธ Some of you are grinding charts, some are grinding narratives. Me? Iโ€™m grinding bubbles. ๐Ÿ˜ญ๐Ÿซง Itโ€™s the weekend, the market can wait a little. ๐Ÿ˜‚ Touch grass, recharge and enjoy the day. Have a productive day, chads. ๐Ÿค๐Ÿ“ˆ
1
1
46
GM CT โ˜€๏ธ Some of you are grinding charts, some are grinding narratives. Me? Iโ€™m grinding bubbles. ๐Ÿ˜ญ๐Ÿซง Itโ€™s the weekend, the market can wait a little. ๐Ÿ˜‚ Touch grass, recharge and enjoy the day. Have a productive day, chads. ๐Ÿค๐Ÿ“ˆ
1
1
46
Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป retweeted
GM Chads โ˜€๏ธ It's a new day, and the goal is to lock in, and make the most of the day. No noise. No excuses. Just keep building right. ๐Ÿซก๐Ÿ“ˆ Letโ€™s get it, chads. ๐Ÿค
Training a frontier AI model has always meant one thing: tens of thousands of GPUs, humming in unison, inside a single data center, owned by a company with a market cap bigger than most countries. Day 20 of my AI & Web3 tool sprint looks at the team trying to break that requirement entirely. ๐Ÿง โš™๏ธ Tool: Prime Intellect @PrimeIntellect ; a decentralized platform that aggregates GPU compute from around the globe and turns it into a unified network for training, evaluating, and deploying large language models. Founded by Vincent Weisser and Johannes Hagemann. What makes this different from just "another compute marketplace": Prime Intellect has actually trained models at meaningful scale, not just theorized about it. INTELLECT-2, a 32B-parameter model, was trained across globally distributed GPUs, no single data center required, and released fully open. The stack: 1๏ธโƒฃ GPU Marketplace: Rent compute on demand, from a single card to 256+ GPU clusters, aggregated across dozens of datacenters worldwide. 2๏ธโƒฃ Environments Hub: 2,500+ open-source reinforcement-learning environments to train and evaluate models against. 3๏ธโƒฃ Hosted Training & Lab: The full post-training pipeline, RL, evaluation, deployment, without needing to own or manage a GPU cluster yourself. The backing: $130M Series A in July 2026 at a $1B valuation, led by Radical Ventures, with Nvidia Ventures, Intel Capital, and Dell Technologies Capital all participating. Over $150M raised in under two years. My Honest Review ๐Ÿ” This is the piece that was missing from everything I've covered so far. Compute access (Day 19) means nothing if you still can't actually train a competitive model on it. Prime Intellect isn't promising decentralized training works, they've already shipped proof, twice. My Take ๐Ÿง  The gap between open-source and closed frontier labs has mostly been a compute-and-coordination gap, not a talent gap. When training infrastructure itself becomes decentralized and accessible, that gap stops being a moat and starts being a temporary head start. The frontier lab moat was never intelligence. It was always infrastructure. That's what's actually up for grabs now.
3
1
4
48
GM Chads โ˜€๏ธ It's a new day, and the goal is to lock in, and make the most of the day. No noise. No excuses. Just keep building right. ๐Ÿซก๐Ÿ“ˆ Letโ€™s get it, chads. ๐Ÿค
Training a frontier AI model has always meant one thing: tens of thousands of GPUs, humming in unison, inside a single data center, owned by a company with a market cap bigger than most countries. Day 20 of my AI & Web3 tool sprint looks at the team trying to break that requirement entirely. ๐Ÿง โš™๏ธ Tool: Prime Intellect @PrimeIntellect ; a decentralized platform that aggregates GPU compute from around the globe and turns it into a unified network for training, evaluating, and deploying large language models. Founded by Vincent Weisser and Johannes Hagemann. What makes this different from just "another compute marketplace": Prime Intellect has actually trained models at meaningful scale, not just theorized about it. INTELLECT-2, a 32B-parameter model, was trained across globally distributed GPUs, no single data center required, and released fully open. The stack: 1๏ธโƒฃ GPU Marketplace: Rent compute on demand, from a single card to 256+ GPU clusters, aggregated across dozens of datacenters worldwide. 2๏ธโƒฃ Environments Hub: 2,500+ open-source reinforcement-learning environments to train and evaluate models against. 3๏ธโƒฃ Hosted Training & Lab: The full post-training pipeline, RL, evaluation, deployment, without needing to own or manage a GPU cluster yourself. The backing: $130M Series A in July 2026 at a $1B valuation, led by Radical Ventures, with Nvidia Ventures, Intel Capital, and Dell Technologies Capital all participating. Over $150M raised in under two years. My Honest Review ๐Ÿ” This is the piece that was missing from everything I've covered so far. Compute access (Day 19) means nothing if you still can't actually train a competitive model on it. Prime Intellect isn't promising decentralized training works, they've already shipped proof, twice. My Take ๐Ÿง  The gap between open-source and closed frontier labs has mostly been a compute-and-coordination gap, not a talent gap. When training infrastructure itself becomes decentralized and accessible, that gap stops being a moat and starts being a temporary head start. The frontier lab moat was never intelligence. It was always infrastructure. That's what's actually up for grabs now.
3
1
4
48
Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป retweeted
Training a frontier AI model has always meant one thing: tens of thousands of GPUs, humming in unison, inside a single data center, owned by a company with a market cap bigger than most countries. Day 20 of my AI & Web3 tool sprint looks at the team trying to break that requirement entirely. ๐Ÿง โš™๏ธ Tool: Prime Intellect @PrimeIntellect ; a decentralized platform that aggregates GPU compute from around the globe and turns it into a unified network for training, evaluating, and deploying large language models. Founded by Vincent Weisser and Johannes Hagemann. What makes this different from just "another compute marketplace": Prime Intellect has actually trained models at meaningful scale, not just theorized about it. INTELLECT-2, a 32B-parameter model, was trained across globally distributed GPUs, no single data center required, and released fully open. The stack: 1๏ธโƒฃ GPU Marketplace: Rent compute on demand, from a single card to 256+ GPU clusters, aggregated across dozens of datacenters worldwide. 2๏ธโƒฃ Environments Hub: 2,500+ open-source reinforcement-learning environments to train and evaluate models against. 3๏ธโƒฃ Hosted Training & Lab: The full post-training pipeline, RL, evaluation, deployment, without needing to own or manage a GPU cluster yourself. The backing: $130M Series A in July 2026 at a $1B valuation, led by Radical Ventures, with Nvidia Ventures, Intel Capital, and Dell Technologies Capital all participating. Over $150M raised in under two years. My Honest Review ๐Ÿ” This is the piece that was missing from everything I've covered so far. Compute access (Day 19) means nothing if you still can't actually train a competitive model on it. Prime Intellect isn't promising decentralized training works, they've already shipped proof, twice. My Take ๐Ÿง  The gap between open-source and closed frontier labs has mostly been a compute-and-coordination gap, not a talent gap. When training infrastructure itself becomes decentralized and accessible, that gap stops being a moat and starts being a temporary head start. The frontier lab moat was never intelligence. It was always infrastructure. That's what's actually up for grabs now.
Training or running an AI model costs real money, and right now, almost all of that money flows through three companies: AWS, Google Cloud, and Azure. That's not a market, that's a toll booth. Day 19 of my AI & Web3 tool sprint looks at the protocol trying to break that up. โšก๐Ÿ–ฅ๏ธ Tool: Akash Network @akashnet โ€“ a decentralized cloud computing marketplace built on Cosmos. Anyone with spare GPU, CPU, or storage capacity can become a provider. Anyone who needs compute bids for it through a reverse auction, providers compete on price, not the other way around. How it actually works: 1๏ธโƒฃ You need compute, you post a deployment request describing what you need 2๏ธโƒฃ Providers across the network bid to fulfill it, competing on price in real time 3๏ธโƒฃ You pick a bid, deploy your workload, container, model training job, or inference endpoint, directly onto that provider's hardware 4๏ธโƒฃ AKT settles payment, staking, and governance across the whole marketplace The pitch, backed by real numbers: Akash claims 70-80% cheaper pricing than equivalent AWS, Google Cloud, or Azure workloads, since you're bidding against a global pool of underused hardware instead of paying fixed hyperscaler rates. The scale: Akash sits at roughly $1.1B market cap, and it's not alone, decentralized GPU compute networks (Akash, Render, io.net combined) generated over $200M in annualized protocol revenue in early 2026, real revenue, not just token speculation. My Honest Review ๐Ÿ” This is the layer underneath everything else I've covered so far. Agents need identity (Day 17), agents need to pay each other (Day 16), but none of that matters if running the actual AI model is priced out of reach. Akash is solving the unglamorous, expensive problem sitting underneath the whole agent economy narrative. My Take ๐Ÿง  The AI compute bottleneck is the single biggest structural constraint on this entire industry right now, demand is outpacing supply, and hyperscalers control the tap. A permissionless, price-competitive alternative isn't a nice-to-have, it's the difference between an open AI ecosystem and one gatekept by three companies. Cheaper compute, open to anyone. That's not a narrative, that's infrastructure. ๐Ÿ’ฏ
1
3
1
3
314
Training a frontier AI model has always meant one thing: tens of thousands of GPUs, humming in unison, inside a single data center, owned by a company with a market cap bigger than most countries. Day 20 of my AI & Web3 tool sprint looks at the team trying to break that requirement entirely. ๐Ÿง โš™๏ธ Tool: Prime Intellect @PrimeIntellect ; a decentralized platform that aggregates GPU compute from around the globe and turns it into a unified network for training, evaluating, and deploying large language models. Founded by Vincent Weisser and Johannes Hagemann. What makes this different from just "another compute marketplace": Prime Intellect has actually trained models at meaningful scale, not just theorized about it. INTELLECT-2, a 32B-parameter model, was trained across globally distributed GPUs, no single data center required, and released fully open. The stack: 1๏ธโƒฃ GPU Marketplace: Rent compute on demand, from a single card to 256+ GPU clusters, aggregated across dozens of datacenters worldwide. 2๏ธโƒฃ Environments Hub: 2,500+ open-source reinforcement-learning environments to train and evaluate models against. 3๏ธโƒฃ Hosted Training & Lab: The full post-training pipeline, RL, evaluation, deployment, without needing to own or manage a GPU cluster yourself. The backing: $130M Series A in July 2026 at a $1B valuation, led by Radical Ventures, with Nvidia Ventures, Intel Capital, and Dell Technologies Capital all participating. Over $150M raised in under two years. My Honest Review ๐Ÿ” This is the piece that was missing from everything I've covered so far. Compute access (Day 19) means nothing if you still can't actually train a competitive model on it. Prime Intellect isn't promising decentralized training works, they've already shipped proof, twice. My Take ๐Ÿง  The gap between open-source and closed frontier labs has mostly been a compute-and-coordination gap, not a talent gap. When training infrastructure itself becomes decentralized and accessible, that gap stops being a moat and starts being a temporary head start. The frontier lab moat was never intelligence. It was always infrastructure. That's what's actually up for grabs now.
Training or running an AI model costs real money, and right now, almost all of that money flows through three companies: AWS, Google Cloud, and Azure. That's not a market, that's a toll booth. Day 19 of my AI & Web3 tool sprint looks at the protocol trying to break that up. โšก๐Ÿ–ฅ๏ธ Tool: Akash Network @akashnet โ€“ a decentralized cloud computing marketplace built on Cosmos. Anyone with spare GPU, CPU, or storage capacity can become a provider. Anyone who needs compute bids for it through a reverse auction, providers compete on price, not the other way around. How it actually works: 1๏ธโƒฃ You need compute, you post a deployment request describing what you need 2๏ธโƒฃ Providers across the network bid to fulfill it, competing on price in real time 3๏ธโƒฃ You pick a bid, deploy your workload, container, model training job, or inference endpoint, directly onto that provider's hardware 4๏ธโƒฃ AKT settles payment, staking, and governance across the whole marketplace The pitch, backed by real numbers: Akash claims 70-80% cheaper pricing than equivalent AWS, Google Cloud, or Azure workloads, since you're bidding against a global pool of underused hardware instead of paying fixed hyperscaler rates. The scale: Akash sits at roughly $1.1B market cap, and it's not alone, decentralized GPU compute networks (Akash, Render, io.net combined) generated over $200M in annualized protocol revenue in early 2026, real revenue, not just token speculation. My Honest Review ๐Ÿ” This is the layer underneath everything else I've covered so far. Agents need identity (Day 17), agents need to pay each other (Day 16), but none of that matters if running the actual AI model is priced out of reach. Akash is solving the unglamorous, expensive problem sitting underneath the whole agent economy narrative. My Take ๐Ÿง  The AI compute bottleneck is the single biggest structural constraint on this entire industry right now, demand is outpacing supply, and hyperscalers control the tap. A permissionless, price-competitive alternative isn't a nice-to-have, it's the difference between an open AI ecosystem and one gatekept by three companies. Cheaper compute, open to anyone. That's not a narrative, that's infrastructure. ๐Ÿ’ฏ
1
3
1
3
314
GM CT. โ˜€๏ธ Happy Taco Tuesday. ๐ŸŒฎ โ€‹Stepping out in full @zippRepublic drip today. ๐Ÿ”ด๐Ÿงข โ€‹โ€‹Huge shoutout to @PeterPsquare , the brain behind this brand. The streetwear drip is top tier! ๐Ÿš€ Letโ€™s build, ship, and make the numbers move. โ€‹Have a great day, family. Letโ€™s get it! ๐Ÿ’ผ๐Ÿ’ซ
Training or running an AI model costs real money, and right now, almost all of that money flows through three companies: AWS, Google Cloud, and Azure. That's not a market, that's a toll booth. Day 19 of my AI & Web3 tool sprint looks at the protocol trying to break that up. โšก๐Ÿ–ฅ๏ธ Tool: Akash Network @akashnet โ€“ a decentralized cloud computing marketplace built on Cosmos. Anyone with spare GPU, CPU, or storage capacity can become a provider. Anyone who needs compute bids for it through a reverse auction, providers compete on price, not the other way around. How it actually works: 1๏ธโƒฃ You need compute, you post a deployment request describing what you need 2๏ธโƒฃ Providers across the network bid to fulfill it, competing on price in real time 3๏ธโƒฃ You pick a bid, deploy your workload, container, model training job, or inference endpoint, directly onto that provider's hardware 4๏ธโƒฃ AKT settles payment, staking, and governance across the whole marketplace The pitch, backed by real numbers: Akash claims 70-80% cheaper pricing than equivalent AWS, Google Cloud, or Azure workloads, since you're bidding against a global pool of underused hardware instead of paying fixed hyperscaler rates. The scale: Akash sits at roughly $1.1B market cap, and it's not alone, decentralized GPU compute networks (Akash, Render, io.net combined) generated over $200M in annualized protocol revenue in early 2026, real revenue, not just token speculation. My Honest Review ๐Ÿ” This is the layer underneath everything else I've covered so far. Agents need identity (Day 17), agents need to pay each other (Day 16), but none of that matters if running the actual AI model is priced out of reach. Akash is solving the unglamorous, expensive problem sitting underneath the whole agent economy narrative. My Take ๐Ÿง  The AI compute bottleneck is the single biggest structural constraint on this entire industry right now, demand is outpacing supply, and hyperscalers control the tap. A permissionless, price-competitive alternative isn't a nice-to-have, it's the difference between an open AI ecosystem and one gatekept by three companies. Cheaper compute, open to anyone. That's not a narrative, that's infrastructure. ๐Ÿ’ฏ
2
2
54
Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป retweeted
Training or running an AI model costs real money, and right now, almost all of that money flows through three companies: AWS, Google Cloud, and Azure. That's not a market, that's a toll booth. Day 19 of my AI & Web3 tool sprint looks at the protocol trying to break that up. โšก๐Ÿ–ฅ๏ธ Tool: Akash Network @akashnet โ€“ a decentralized cloud computing marketplace built on Cosmos. Anyone with spare GPU, CPU, or storage capacity can become a provider. Anyone who needs compute bids for it through a reverse auction, providers compete on price, not the other way around. How it actually works: 1๏ธโƒฃ You need compute, you post a deployment request describing what you need 2๏ธโƒฃ Providers across the network bid to fulfill it, competing on price in real time 3๏ธโƒฃ You pick a bid, deploy your workload, container, model training job, or inference endpoint, directly onto that provider's hardware 4๏ธโƒฃ AKT settles payment, staking, and governance across the whole marketplace The pitch, backed by real numbers: Akash claims 70-80% cheaper pricing than equivalent AWS, Google Cloud, or Azure workloads, since you're bidding against a global pool of underused hardware instead of paying fixed hyperscaler rates. The scale: Akash sits at roughly $1.1B market cap, and it's not alone, decentralized GPU compute networks (Akash, Render, io.net combined) generated over $200M in annualized protocol revenue in early 2026, real revenue, not just token speculation. My Honest Review ๐Ÿ” This is the layer underneath everything else I've covered so far. Agents need identity (Day 17), agents need to pay each other (Day 16), but none of that matters if running the actual AI model is priced out of reach. Akash is solving the unglamorous, expensive problem sitting underneath the whole agent economy narrative. My Take ๐Ÿง  The AI compute bottleneck is the single biggest structural constraint on this entire industry right now, demand is outpacing supply, and hyperscalers control the tap. A permissionless, price-competitive alternative isn't a nice-to-have, it's the difference between an open AI ecosystem and one gatekept by three companies. Cheaper compute, open to anyone. That's not a narrative, that's infrastructure. ๐Ÿ’ฏ
Every DeFi hack story has the same plot twist: someone clicked "Sign" without reading what they were actually signing. Day 18 of my AI & Web3 tool sprint dives into the tech behind that one click. ๐Ÿ”โš ๏ธ First, the concept most people never learn: A signature and a transaction aren't the same thing. Signing costs no gas and doesn't move funds by itself, but it can authorize something that does, later, without you seeing it happen. That gap is exactly what wallet drainers exploit: trick you into signing an approval or a permit, and the attacker moves your tokens whenever they want, no need to ever touch your private key. Older signing methods (like raw eth_sign) show you meaningless hex data, basically asking you to trust blindly. EIP-712 fixed part of this by making signatures human-readable, structured data instead of gibberish. But readable isn't the same as understood, most people still don't know what they're approving. Tool: Rabby Wallet @Rabby_io โ€“ an EVM wallet built by the DeBank team that decodes what you're about to sign before you sign it. Instead of raw contract calls, it shows you plain language: "Swapped 100 USDC for 0.08 ETH," or a clear warning if your balance is about to drop with nothing coming back. What it actually does: 1๏ธโƒฃ Transaction simulation: Previews exactly what will happen before you approve, catching malicious contracts before they execute. 2๏ธโƒฃ Risk scanning: Flags known phishing sites and suspicious contracts using a maintained threat database 3๏ธโƒฃ Approval management: Shows and lets you revoke old, unlimited token approvals, the single most common vector behind DeFi exploits My Honest Review ๐Ÿ” With over $2.3 billion lost to crypto exploits in 2025 alone, most of it from consent mistakes rather than actual protocol breaks, a wallet that forces you to slow down at the exact moment of risk isn't a luxury feature. It's the difference between losing your funds and not. My Take ๐Ÿง  The real vulnerability was never the blockchain, it's the human clicking "Sign" without understanding what that means. Tools like Rabby don't remove that risk entirely, but they finally make the moment of signing something you can actually see through instead of trust blindly. The safest habit to adopt in DeFi, isn't depending on a particular tool. It's slowing down before you click "Sign." โœ๏ธ
2
1
2
312
Training or running an AI model costs real money, and right now, almost all of that money flows through three companies: AWS, Google Cloud, and Azure. That's not a market, that's a toll booth. Day 19 of my AI & Web3 tool sprint looks at the protocol trying to break that up. โšก๐Ÿ–ฅ๏ธ Tool: Akash Network @akashnet โ€“ a decentralized cloud computing marketplace built on Cosmos. Anyone with spare GPU, CPU, or storage capacity can become a provider. Anyone who needs compute bids for it through a reverse auction, providers compete on price, not the other way around. How it actually works: 1๏ธโƒฃ You need compute, you post a deployment request describing what you need 2๏ธโƒฃ Providers across the network bid to fulfill it, competing on price in real time 3๏ธโƒฃ You pick a bid, deploy your workload, container, model training job, or inference endpoint, directly onto that provider's hardware 4๏ธโƒฃ AKT settles payment, staking, and governance across the whole marketplace The pitch, backed by real numbers: Akash claims 70-80% cheaper pricing than equivalent AWS, Google Cloud, or Azure workloads, since you're bidding against a global pool of underused hardware instead of paying fixed hyperscaler rates. The scale: Akash sits at roughly $1.1B market cap, and it's not alone, decentralized GPU compute networks (Akash, Render, io.net combined) generated over $200M in annualized protocol revenue in early 2026, real revenue, not just token speculation. My Honest Review ๐Ÿ” This is the layer underneath everything else I've covered so far. Agents need identity (Day 17), agents need to pay each other (Day 16), but none of that matters if running the actual AI model is priced out of reach. Akash is solving the unglamorous, expensive problem sitting underneath the whole agent economy narrative. My Take ๐Ÿง  The AI compute bottleneck is the single biggest structural constraint on this entire industry right now, demand is outpacing supply, and hyperscalers control the tap. A permissionless, price-competitive alternative isn't a nice-to-have, it's the difference between an open AI ecosystem and one gatekept by three companies. Cheaper compute, open to anyone. That's not a narrative, that's infrastructure. ๐Ÿ’ฏ
Every DeFi hack story has the same plot twist: someone clicked "Sign" without reading what they were actually signing. Day 18 of my AI & Web3 tool sprint dives into the tech behind that one click. ๐Ÿ”โš ๏ธ First, the concept most people never learn: A signature and a transaction aren't the same thing. Signing costs no gas and doesn't move funds by itself, but it can authorize something that does, later, without you seeing it happen. That gap is exactly what wallet drainers exploit: trick you into signing an approval or a permit, and the attacker moves your tokens whenever they want, no need to ever touch your private key. Older signing methods (like raw eth_sign) show you meaningless hex data, basically asking you to trust blindly. EIP-712 fixed part of this by making signatures human-readable, structured data instead of gibberish. But readable isn't the same as understood, most people still don't know what they're approving. Tool: Rabby Wallet @Rabby_io โ€“ an EVM wallet built by the DeBank team that decodes what you're about to sign before you sign it. Instead of raw contract calls, it shows you plain language: "Swapped 100 USDC for 0.08 ETH," or a clear warning if your balance is about to drop with nothing coming back. What it actually does: 1๏ธโƒฃ Transaction simulation: Previews exactly what will happen before you approve, catching malicious contracts before they execute. 2๏ธโƒฃ Risk scanning: Flags known phishing sites and suspicious contracts using a maintained threat database 3๏ธโƒฃ Approval management: Shows and lets you revoke old, unlimited token approvals, the single most common vector behind DeFi exploits My Honest Review ๐Ÿ” With over $2.3 billion lost to crypto exploits in 2025 alone, most of it from consent mistakes rather than actual protocol breaks, a wallet that forces you to slow down at the exact moment of risk isn't a luxury feature. It's the difference between losing your funds and not. My Take ๐Ÿง  The real vulnerability was never the blockchain, it's the human clicking "Sign" without understanding what that means. Tools like Rabby don't remove that risk entirely, but they finally make the moment of signing something you can actually see through instead of trust blindly. The safest habit to adopt in DeFi, isn't depending on a particular tool. It's slowing down before you click "Sign." โœ๏ธ
2
1
2
312
Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป retweeted
GM CT โ˜€๏ธ Saturday is for touching grass before touching charts. ๐ŸŒฑ๐Ÿ™‚โ€โ†”๏ธ Taking a little time to breathe, reset, and enjoy the weekend before getting back to the grind. Clear mind. Fresh energy. Then we go chase those green candles. ๐Ÿ“ˆ๐Ÿค Have a blessed Weekend, chads. ๐Ÿซก
Every DeFi hack story has the same plot twist: someone clicked "Sign" without reading what they were actually signing. Day 18 of my AI & Web3 tool sprint dives into the tech behind that one click. ๐Ÿ”โš ๏ธ First, the concept most people never learn: A signature and a transaction aren't the same thing. Signing costs no gas and doesn't move funds by itself, but it can authorize something that does, later, without you seeing it happen. That gap is exactly what wallet drainers exploit: trick you into signing an approval or a permit, and the attacker moves your tokens whenever they want, no need to ever touch your private key. Older signing methods (like raw eth_sign) show you meaningless hex data, basically asking you to trust blindly. EIP-712 fixed part of this by making signatures human-readable, structured data instead of gibberish. But readable isn't the same as understood, most people still don't know what they're approving. Tool: Rabby Wallet @Rabby_io โ€“ an EVM wallet built by the DeBank team that decodes what you're about to sign before you sign it. Instead of raw contract calls, it shows you plain language: "Swapped 100 USDC for 0.08 ETH," or a clear warning if your balance is about to drop with nothing coming back. What it actually does: 1๏ธโƒฃ Transaction simulation: Previews exactly what will happen before you approve, catching malicious contracts before they execute. 2๏ธโƒฃ Risk scanning: Flags known phishing sites and suspicious contracts using a maintained threat database 3๏ธโƒฃ Approval management: Shows and lets you revoke old, unlimited token approvals, the single most common vector behind DeFi exploits My Honest Review ๐Ÿ” With over $2.3 billion lost to crypto exploits in 2025 alone, most of it from consent mistakes rather than actual protocol breaks, a wallet that forces you to slow down at the exact moment of risk isn't a luxury feature. It's the difference between losing your funds and not. My Take ๐Ÿง  The real vulnerability was never the blockchain, it's the human clicking "Sign" without understanding what that means. Tools like Rabby don't remove that risk entirely, but they finally make the moment of signing something you can actually see through instead of trust blindly. The safest habit to adopt in DeFi, isn't depending on a particular tool. It's slowing down before you click "Sign." โœ๏ธ
2
1
4
49
GM CT โ˜€๏ธ Saturday is for touching grass before touching charts. ๐ŸŒฑ๐Ÿ™‚โ€โ†”๏ธ Taking a little time to breathe, reset, and enjoy the weekend before getting back to the grind. Clear mind. Fresh energy. Then we go chase those green candles. ๐Ÿ“ˆ๐Ÿค Have a blessed Weekend, chads. ๐Ÿซก
Every DeFi hack story has the same plot twist: someone clicked "Sign" without reading what they were actually signing. Day 18 of my AI & Web3 tool sprint dives into the tech behind that one click. ๐Ÿ”โš ๏ธ First, the concept most people never learn: A signature and a transaction aren't the same thing. Signing costs no gas and doesn't move funds by itself, but it can authorize something that does, later, without you seeing it happen. That gap is exactly what wallet drainers exploit: trick you into signing an approval or a permit, and the attacker moves your tokens whenever they want, no need to ever touch your private key. Older signing methods (like raw eth_sign) show you meaningless hex data, basically asking you to trust blindly. EIP-712 fixed part of this by making signatures human-readable, structured data instead of gibberish. But readable isn't the same as understood, most people still don't know what they're approving. Tool: Rabby Wallet @Rabby_io โ€“ an EVM wallet built by the DeBank team that decodes what you're about to sign before you sign it. Instead of raw contract calls, it shows you plain language: "Swapped 100 USDC for 0.08 ETH," or a clear warning if your balance is about to drop with nothing coming back. What it actually does: 1๏ธโƒฃ Transaction simulation: Previews exactly what will happen before you approve, catching malicious contracts before they execute. 2๏ธโƒฃ Risk scanning: Flags known phishing sites and suspicious contracts using a maintained threat database 3๏ธโƒฃ Approval management: Shows and lets you revoke old, unlimited token approvals, the single most common vector behind DeFi exploits My Honest Review ๐Ÿ” With over $2.3 billion lost to crypto exploits in 2025 alone, most of it from consent mistakes rather than actual protocol breaks, a wallet that forces you to slow down at the exact moment of risk isn't a luxury feature. It's the difference between losing your funds and not. My Take ๐Ÿง  The real vulnerability was never the blockchain, it's the human clicking "Sign" without understanding what that means. Tools like Rabby don't remove that risk entirely, but they finally make the moment of signing something you can actually see through instead of trust blindly. The safest habit to adopt in DeFi, isn't depending on a particular tool. It's slowing down before you click "Sign." โœ๏ธ
2
1
4
49
Apeteks ๐Ÿต๐Ÿ‘จโ€๐Ÿ’ป retweeted
Every DeFi hack story has the same plot twist: someone clicked "Sign" without reading what they were actually signing. Day 18 of my AI & Web3 tool sprint dives into the tech behind that one click. ๐Ÿ”โš ๏ธ First, the concept most people never learn: A signature and a transaction aren't the same thing. Signing costs no gas and doesn't move funds by itself, but it can authorize something that does, later, without you seeing it happen. That gap is exactly what wallet drainers exploit: trick you into signing an approval or a permit, and the attacker moves your tokens whenever they want, no need to ever touch your private key. Older signing methods (like raw eth_sign) show you meaningless hex data, basically asking you to trust blindly. EIP-712 fixed part of this by making signatures human-readable, structured data instead of gibberish. But readable isn't the same as understood, most people still don't know what they're approving. Tool: Rabby Wallet @Rabby_io โ€“ an EVM wallet built by the DeBank team that decodes what you're about to sign before you sign it. Instead of raw contract calls, it shows you plain language: "Swapped 100 USDC for 0.08 ETH," or a clear warning if your balance is about to drop with nothing coming back. What it actually does: 1๏ธโƒฃ Transaction simulation: Previews exactly what will happen before you approve, catching malicious contracts before they execute. 2๏ธโƒฃ Risk scanning: Flags known phishing sites and suspicious contracts using a maintained threat database 3๏ธโƒฃ Approval management: Shows and lets you revoke old, unlimited token approvals, the single most common vector behind DeFi exploits My Honest Review ๐Ÿ” With over $2.3 billion lost to crypto exploits in 2025 alone, most of it from consent mistakes rather than actual protocol breaks, a wallet that forces you to slow down at the exact moment of risk isn't a luxury feature. It's the difference between losing your funds and not. My Take ๐Ÿง  The real vulnerability was never the blockchain, it's the human clicking "Sign" without understanding what that means. Tools like Rabby don't remove that risk entirely, but they finally make the moment of signing something you can actually see through instead of trust blindly. The safest habit to adopt in DeFi, isn't depending on a particular tool. It's slowing down before you click "Sign." โœ๏ธ
4
3
1
4
250
Every DeFi hack story has the same plot twist: someone clicked "Sign" without reading what they were actually signing. Day 18 of my AI & Web3 tool sprint dives into the tech behind that one click. ๐Ÿ”โš ๏ธ First, the concept most people never learn: A signature and a transaction aren't the same thing. Signing costs no gas and doesn't move funds by itself, but it can authorize something that does, later, without you seeing it happen. That gap is exactly what wallet drainers exploit: trick you into signing an approval or a permit, and the attacker moves your tokens whenever they want, no need to ever touch your private key. Older signing methods (like raw eth_sign) show you meaningless hex data, basically asking you to trust blindly. EIP-712 fixed part of this by making signatures human-readable, structured data instead of gibberish. But readable isn't the same as understood, most people still don't know what they're approving. Tool: Rabby Wallet @Rabby_io โ€“ an EVM wallet built by the DeBank team that decodes what you're about to sign before you sign it. Instead of raw contract calls, it shows you plain language: "Swapped 100 USDC for 0.08 ETH," or a clear warning if your balance is about to drop with nothing coming back. What it actually does: 1๏ธโƒฃ Transaction simulation: Previews exactly what will happen before you approve, catching malicious contracts before they execute. 2๏ธโƒฃ Risk scanning: Flags known phishing sites and suspicious contracts using a maintained threat database 3๏ธโƒฃ Approval management: Shows and lets you revoke old, unlimited token approvals, the single most common vector behind DeFi exploits My Honest Review ๐Ÿ” With over $2.3 billion lost to crypto exploits in 2025 alone, most of it from consent mistakes rather than actual protocol breaks, a wallet that forces you to slow down at the exact moment of risk isn't a luxury feature. It's the difference between losing your funds and not. My Take ๐Ÿง  The real vulnerability was never the blockchain, it's the human clicking "Sign" without understanding what that means. Tools like Rabby don't remove that risk entirely, but they finally make the moment of signing something you can actually see through instead of trust blindly. The safest habit to adopt in DeFi, isn't depending on a particular tool. It's slowing down before you click "Sign." โœ๏ธ
4
3
1
4
250
My dad has been having fun exploring ai tools, in his free time since I introduced him to some of them. He showed me this cool flyer he made using @ChatGPT ๐Ÿ˜Ž Good tek fr. ๐Ÿ‘Œ Anyways, these are the kind of conversations we usually have.
1
1
1
2
195