avatar
고로치
28,460 Twitter followers
Follow
지혜로움에 관심이 많고 solana와 jupiter 그리고 ai에 관심 많습니다. 생각을 만드는 글을 씁니다. virtual referral - https://t.co/LPnlJcLsfK
Posts
avatar
고로치
Thread
This graph demonstrates the rapid advancement of large-scale language models (LLMs) and their rapidly expanding capabilities. In summary, AI, while previously capable of simple tasks that humans could perform in minutes, is now rapidly evolving to the point where it can perform complex, specialized tasks that would require four to five hours of human concentration. - Graph Axes X-axis: LLM release year (2020–2026) Represents the emergence of AI models over time. Y-axis: Task duration for humans The time it takes a human expert to perform a given task is used as a measure of difficulty. The higher the value, the more complex and specialized the task requires. - Early Stages (2020–2023) During the GPT-3 era, AI had a high success rate only with very simple tasks that humans could complete in seconds or minutes, such as sentence completion or simple information retrieval. - Present and Near Future (2024-2026) The graph suddenly spikes upward, suggesting that technological advancement is occurring exponentially, not linearly. With the emergence of models like o4-mini, gpt-5, and Claude Opus 4.5, the difficulty of solving tasks has increased dramatically. - Progress through Specific Task Examples The tasks listed on the left side of the graph demonstrate just how intelligent AI is becoming. Fix bugs in small Python libraries (approximately 1.5 hours): Beyond simple coding, AI has begun to develop debugging skills to find and fix bugs in libraries. Exploit a buffer overflow (approximately 2 hours): This task requires advanced hacking and security knowledge to identify and exploit security vulnerabilities. Train adversarially robust image model (approximately 4 hours): This is an AI engineering task that goes beyond simply creating images to train robust AI models that can withstand adversarial attacks. - Conclusion and Implications The core message of this graph is that AI has now moved beyond mere assistance and entered the realm of expert expertise. If AI can complete tasks that would take humans half a day in seconds or minutes, it will revolutionize productivity across industries. This demonstrates the potential for AI to perform tasks previously reserved for highly skilled knowledge workers, such as coding, security analysis, and modeling. The chart shows that the y-axis doubles every four months, suggesting that the day when AI will perform a full day's worth of work (24 hours) or even longer seems imminent. twitter.com/gorochi0315/status...
avatar
고로치
During the gold standard era, people didn't carry heavy gold bars in their pockets to make purchases. Instead, they safely deposited their gold in bank vaults. Banks issued paper certificates as proof of their gold deposits, which were the original "paper money (dollars)." At the time, a $10 bill was like a "gold depository receipt" that promised, "I'll exchange it for $10 worth of gold at any bank." People kept the physical gold in the vault, exchanging only the lightweight depository receipts (bills) for convenience. Surprisingly, today's stock market operates exactly like this old gold standard system. Suppose we bought Samsung Electronics or Tesla stock through a brokerage app (MTS). Would our names be directly listed on the company's shareholder register? No. Just like gold in a bank vault, all physical stocks (either physical or electronically registered originals) are stored in a centralized vault like the DTCC (Depository and Clearing Corporation of America). Through a securities firm, we simply hold a digital record—a stock depository receipt—that states, "I have the right to one share of the stock held in that vault." So, rather than exchanging physical shares, we're merely transferring ownership of the "stock depository receipt" between us over a network. - With the advent of blockchain technology, a movement to revolutionize this system, namely "tokenization," has begun. A crucial fork in the road emerges here. The meaning changes slightly depending on the question of "what to tokenize." 1/ Model A: Keep the stocks in the depository as before, and tokenize the "stock depository receipt (rights)". 2/ Model B: Bypass the depository and tokenize the "stock itself." There are two methods... - Of these, Model A, tokenizing the "deposit receipt," may seem similar on the surface, but it makes a significant difference in speed and cost. Here's why: Currently, financial institutions maintain their own separate ledgers (Excel files). When Bank A sends stocks to Securities Company B, the "reconciliation" process alone—verifying that their accounts match—takes two to three days and costs hundreds of billions of won. However, if deposit receipts are tokenized and placed on the blockchain, all financial institutions will be able to view a single, shared ledger in real time. This eliminates the need for arguments like, "My ledger is different from yours!" and shifts stock transaction settlement from t+2 days (a three-day process) to t+0 (immediate settlement). Furthermore, when settlement took three days, a significant amount of "margin (deposit)" had to be tied up to protect against unforeseen events. With immediate settlement, this money can be freed up and invested elsewhere. To use an analogy, it's like replacing the traditional process of exchanging and stamping contracts by mail with a real-time, shared document like Google Docs, allowing for simultaneous work. In other words, work processing speeds are dramatically increased. - And if the "stock itself" is tokenized, as in Model B, bypassing a central vault like the DTCC, tokens held in personal digital wallets become stocks, transferring complete ownership of the stocks back to the individual. As a result, stocks can be traded peer-to-peer (P2P) with anyone around the world, even on nights or weekends when the stock market is closed. You can use the "Apple stock tokens" in your wallet as collateral and instantly borrow dollars (stablecoins) from DeFi protocols without bank review. Stocks go beyond mere investment assets; they can be freely combined with various financial products, like Lego blocks. We can say we own something when we can use it as we wish. Currently, the financial system doesn't fully own each stock, so we can't freely lend or use it. But if tokenization makes those stocks our own, we can freely lend them out and use them as collateral for loans. This is through blockchain DeFi services. - To summarize, Just as we used gold warrants (dollars) until now, we are now trading stock warrants (rights to receive payments). However, tokenizing 'warrants' can maximize efficiency by eliminating complex paperwork and waiting times between financial institutions. And tokenizing 'stocks' themselves will give individuals direct control over their stocks, enabling 24-hour trading and free financial use (DeFi). twitter.com/gorochi0315/status...
BTC
2.6%
avatar
고로치
Thread
I just finished reading Google's latest AGI (Artificial General Intelligence) paper, and its perspective is so bold and crypto-native that I almost felt like I was reading a cryptocurrency project's white paper. Let me summarize some of the key points: 1. AGI will ultimately be a decentralized autonomous organization (DAO), not a CEO. We often fantasize about waking up one morning to an omnipotent deity like GPT-10. However, the paper points out that the future AGI is likely to be decentralized. Just as no one person in a company excels at everything, AGI will be a network of numerous, complementary "specialized agents." This network has no single central point, and superintelligence emerges from the intense trading and collaboration among the agents. In other words, AGI is not a single entity, but a kind of "market state." 2. AGI's governance should rely on "smart contracts," not "law." As models evolve from a single structure to a market structure, the safety paradigm must also shift from "psychology" to "governance studies." Previously, AI safety was a matter of aligning a single, massive brain. However, human oversight is powerless in high-frequency interactions occurring hundreds of millions of times per second. Therefore, the introduction of smart contracts is essential. When an agent completes a task, an oracle verifies the result and automatically executes payment. "Code becomes law," and if safety constraints are not met, the flow of funds is blocked. 3. Introduction of "Staking" and "Slashing" How can we prevent malicious agents? The research team surprisingly replicated the Proof-of-Stake (PoS) mechanism. If an agent wants to place a large order, it must first stake its assets. If malicious activity is detected during the audit process, the smart contract immediately slashes the collateral assets. This trust based on economic collateral is far more effective than simple code review. 4. On-Chain Identity Authentication (DID) and Gas Fee Regulation DID: To prevent Sybil attacks, each agent must have a unique identity based on public key cryptography, linked to a legal entity. Dynamic Gas Fee: To prevent spam data contamination, we propose charging a dynamic fee for agent operations. This closely mirrors the gas fee regulation mechanism used on the Ethereum network during congestion. 5. On-Chain Records: All decisions and transaction histories must be recorded in a cryptographically secure, immutable, append-only ledger. This enables forensic analysis in the event of a system failure and ensures that no one can evade accountability. This paper represents a paradigm shift in AI security. In other words, it extends beyond simple computer science and value alignment to the realms of economics and game theory. The cryptocurrency industry's intense exploration of DIDs, smart contracts, oracles, economic models, and governance mechanisms over the past decade, regardless of their maturity, represents at least a few steps forward. And these steps may even lay the foundation for a future, massive, decentralized, silicon-based life form. Future AGI security experts will likely be closer to "AI economists" who understand game theory, market design, and decentralized governance than code-writing engineers. Our fundamental task is not simply to modify the neurons of this massive model, but to design the consensus, incentive, and governance models of this new species. Due to the extensive amount of information in the paper, the above is only a partial excerpt. Those interested are encouraged to read the original text.
Chao
@chaowxyz
刚读完Google最新的AGI论文,论文的观点大胆且非常crypto native,我一度以为是在看加密项目的白皮书。 几个核心观点: 1. AGI终将是一个 DAO,而不是一个 CEO。 我们总是幻想某天一觉醒来,诞生了 GPT-10 这样的全知全能神。论文指出未来的AGI大概率是分布式的
loading indicator
Loading..