2026-05-17 Domingo

Centro de Notícias - Página 7

Obtém notícias cripto em tempo real e tendências de mercado com o Centro de Notícias da HTX.

Anthropic Has Taught Models to Understand Morality and Opened a New Path for Distillation

Anthropic's research "Teaching Claude Why" reveals a new, data-efficient method for AI alignment. Instead of relying on massive reinforcement learning with punishment (RLHF), which only teaches models to mimic safe answers without true ethical understanding, they used a small dataset (3 million tokens) of "difficult advice." This data consisted of detailed moral deliberations, reasoning, and debates, teaching the model the *why* behind decisions. The key was "deliberation-enhanced" Supervised Fine-Tuning (SFT). The model was trained on responses that included a "chain of thought" (CoT) process based on a constitutional framework. This framework included top-level principles, practical heuristics (like the "1000-user test"), and an 8-factor utility calculator (evaluating harm probability, reversibility, consent, etc.) for weighing complex trade-offs. This approach dropped model misalignment rates from 22% to 3% and showed strong generalization to unseen scenarios. The success challenges the old belief that "SFT memorizes, RL generalizes." It shows that SFT can generalize powerfully if the training data has two features: 1) high prompt diversity (many different scenario types) and 2) CoT supervision (showing the reasoning steps, not just the final answer). The model learns the underlying *thinking framework*, not just surface-level behaviors. This method points to a new paradigm for training AI in "non-RLVR" domains—areas like ethics, creative writing, or strategy where there's no single verifiable answer. The formula is: Domain Constitution + Heuristics + Multi-Factor Deliberation Framework + Diverse Deliberative CoT Data = Generalized capability. It represents a new form of "distillation," moving competition from pure compute towards who can best structure expert knowledge into high-quality reasoning datasets.

marsbit2 dias atrás 10:55

Anthropic Has Taught Models to Understand Morality and Opened a New Path for Distillation

marsbit2 dias atrás 10:55

CLARITY Bill Still Unsettled, Caught Up in U.S. Bipartisan Political Game

The U.S. Senate Banking Committee voted 15-9 to advance the CLARITY Act for a full Senate vote, overcoming a key industry compromise on stablecoin rewards. However, the legislation faces significant political hurdles as it moves forward. The vote fell largely along party lines, with all 13 Republican committee members voting in favor and only two Democrats—Ruben Gallego and Angela Alsobrooks—joining them. Both Democrats indicated their support was conditional to keep debate alive and could be withdrawn later. The path to final passage in the full Senate remains difficult. The bill requires 60 votes to pass. With a current 53-47 Republican majority, this means at least 7 Democratic votes are needed, assuming no Republican defections. The narrow, partisan committee vote suggests securing that level of bipartisan support will be challenging. During the committee process, Democrats pushed for numerous amendments, including a key ethics provision to restrict top government officials' financial ties to the crypto industry. Republicans used their majority to block these amendments, leaving core Democratic concerns unaddressed. These issues, particularly around anti-money laundering and ethics, are expected to resurface as major points of contention during the full Senate debate. Republicans are pushing for a swift Senate vote, ideally before the July 4 recess and the upcoming election season, fearing that a shift in congressional power after the next election could jeopardize the bill's future. However, given the current political stalemate and Democratic demands for concessions, the act's prospects for becoming law in 2026 are uncertain.

Odaily星球日报2 dias atrás 10:48

CLARITY Bill Still Unsettled, Caught Up in U.S. Bipartisan Political Game

Odaily星球日报2 dias atrás 10:48

Listed and Halted, Surge Over 108% in a Single Day, Is Cerebras Really the 'Next Nvidia'?

Cerebras Systems (CBRS), labeled the "next Nvidia," debuted on the NASDAQ on May 14th, 2025. Its stock price surged over 108% from its $185 IPO price, briefly touching $385 before settling around $311. CEO Andrew Feldman claimed the company's wafer-scale AI chips are "58 times larger and 15-20 times faster" than competitors like Nvidia. The company's core innovation is the Wafer Scale Engine (WSE), a massive, dinner-plate-sized chip designed to avoid the bottlenecks of interconnecting multiple GPUs. Its latest system, the CS-3, offers high-performance computing for AI training and inference. While still a niche player with $5.1 billion in 2025 revenue, Cerebras has secured major contracts, most notably a multi-year, over $20 billion computing deal with OpenAI. This partnership is deep: OpenAI is a major customer, a creditor via a $1 billion loan, and holds warrants that could make it a 10-11% shareholder in Cerebras. Despite the hype, the article argues Cerebras is unlikely to dethrone Nvidia soon. Nvidia's ecosystem (CUDA), vast scale, manufacturing efficiency, and diversified product line present a formidable moat. Cerebras faces high costs, production challenges with its giant chips, and competition from AMD, Google, and others. However, strong demand for AI inference and its key partnerships could support its stock price in the short to medium term. In conclusion, Cerebras is positioned as a high-speed specialist in the AI hardware market, not a broad-market replacement for the current industry leader.

Odaily星球日报2 dias atrás 10:34

Listed and Halted, Surge Over 108% in a Single Day, Is Cerebras Really the 'Next Nvidia'?

Odaily星球日报2 dias atrás 10:34

BIT Research: If It Followed Nasdaq, Bitcoin Should Be Close to $140,000

BIT Research: Bitcoin Price Analysis Under Inflation Re-pricing The market is currently undergoing a macro adjustment phase dominated by inflation re-pricing. Analysis suggests that if Bitcoin had continued to follow Nasdaq's trajectory, its theoretical price would be near $140,000. However, a significant divergence between the two assets has emerged since October 2025. The core reason is the resurgence of US inflation, which has led to a reversal in market expectations for the Federal Reserve's rate-cut path. Recent data shows US CPI rising to 3.8% and PPI to 6.0%, prompting markets to scale back expectations for 2026 rate cuts. For Bitcoin, the previous supportive narrative of anticipated loose liquidity is weakening. Concurrently, escalating tensions involving Iran have driven oil prices up approximately 40% since late February 2026, heightening inflation concerns through rising energy costs. While the market currently views this inflation surge as a temporary pressure point, the interplay between energy, interest rates, and risk appetite is prompting a reassessment of the potential for a prolonged high-rate environment. In this context, Bitcoin has begun to underperform tech stocks, which can benefit from nominal inflation. The divergence stems from a key distinction: Bitcoin's past rallies were driven by loose liquidity and rate-cut expectations, not inflation itself. As a long-duration asset, Bitcoin is highly sensitive to interest rate paths. When expectations for rate cuts are withdrawn, its valuation faces pressure. Unlike equities, which can benefit from increased nominal revenues and reduced real debt burdens during inflation, Bitcoin possesses neither debt that inflates away nor cash flows that expand with inflation, offering no direct structural benefit from rising prices. Looking ahead, the critical question is whether high inflation will force the Fed to maintain elevated rates for longer. The BIT model anticipates US CPI could potentially rise further to 6.0%. Additionally, factors like AI infrastructure expansion—driving data center construction and power demand—may sustain energy price pressures and extend the period of above-target inflation. In such an environment, tech stocks gain from order growth and improved earnings expectations, while Bitcoin remains susceptible to high-rate pressure. In summary, the current shift does not invalidate Bitcoin's long-term thesis but reflects a market re-evaluation of interest rate and liquidity paths amid resurgent inflation. In the short term, a high-inflation environment may continue to suppress Bitcoin's performance relative to Nasdaq. This represents a slowdown in its upward momentum rather than a bearish turn. Bitcoin could regain support once markets begin to reprice expectations for future liquidity easing.

marsbit2 dias atrás 10:07

BIT Research: If It Followed Nasdaq, Bitcoin Should Be Close to $140,000

marsbit2 dias atrás 10:07

Bankless Interview: Private Equity Insiders Reveal the Inside Story of Anthropic's Primary Market Trading

**Bankless Interview: A Private Equity Veteran Exposes the Dark Side of Anthropic's Pre-IPO Trading** In a Bankless podcast, Patagon founder Dio Casares reveals the opaque inner workings of the massive secondary market for shares in pre-IPO giants like Anthropic. The market, driven by private SPVs (special purpose vehicles), brokers, and even informal networks, sees hundreds of billions in notional value changing hands, with single-deal fees as high as 10%. However, an estimated 10-20% of transactions involve fraud or fabricated share certificates. Intermediaries often profit more from these deals than from their core investment businesses. Two types of "secondary" exist: company-sanctioned trades (like employee tender offers) that bring new money to the company, and disruptive "gray market" trades on platforms like Hive or Forge, which companies like Anthropic actively fight. The latter creates pricing chaos and complicates primary fundraising. A major risk involves multi-layered, nested SPV structures. When a company like Anthropic finally IPOs, delays in distributing shares down these chains, combined with discretionary powers of fund managers (GPs) to hold or sell, could trigger a wave of lawsuits and settlement nightmares lasting years. For small investors in "tokenized" versions of these assets, transparency is minimal, and due diligence is often impossible. Casares advises extreme caution, suggesting investors trust their gut and exit if something feels wrong. He warns that the post-IPO period will be a major "reckoning" for this wild and largely unregulated market.

marsbit2 dias atrás 09:44

Bankless Interview: Private Equity Insiders Reveal the Inside Story of Anthropic's Primary Market Trading

marsbit2 dias atrás 09:44

Is Elon Musk Actually the Victim?

"Victim or Vindicator? Inside the OpenAI Trial That Shattered the Myth." In May 2026, the federal court in Oakland became the stage for deconstructing the carefully curated narrative of OpenAI. The trial revealed a complex reality far removed from its founding ideals. The core dispute centered on whether OpenAI, founded in 2015 as a non-profit dedicated to benefiting "all of humanity," had betrayed its mission by shifting towards a lucrative commercial structure, particularly after its 2019 capped-profit affiliate (OpenAI LP) was established and Microsoft invested $13 billion. Elon Musk, a co-founder and early funder, sued, claiming the organization was "stolen" and turned into a de facto Microsoft subsidiary for private gain. OpenAI countered that Musk's funds were unconditional donations and his lawsuit was driven by a desire for control and regret after leaving to found his own AI venture, xAI. The trial exposed early fractures. Evidence from 2017, years before ChatGPT's success, showed the founders were already grappling with the immense financial demands of pursuing Artificial General Intelligence (AGI). Musk himself had proposed having Tesla fund OpenAI. The court scrutinized whether the founders knowingly crossed a moral line. Greg Brockman's personal diary, entered as evidence, contained entries about wealth goals and anxieties over the company's revenue path, alongside self-reminders about the moral bankruptcy of "stealing" the non-profit. Brockman later testified his OpenAI stake was worth nearly $30 billion. The character of CEO Sam Altman was a key battleground. Musk's legal team cited five individuals, including co-founder Ilya Sutskever and former board members, who had described Altman as dishonest. This highlighted a recurring "trust debt" within OpenAI's leadership, exemplified by the chaotic 2023 boardroom coup and subsequent reinstatement. Altman defended his position, arguing Musk sought to absorb OpenAI into Tesla and that commercial success amplified OpenAI's charitable impact. Testimony from Microsoft CEO Satya Nadella underscored how commercial realities now dominated. While framing Microsoft's massive investment as a way to enlarge the non-profit's funding "pie," texts revealed Nadella pressuring Altman to launch ChatGPT's paid version quickly. Nadella also revealed that during the 2023 crisis, Microsoft was prepared to hire Altman and his team, showcasing the board's diminished power against the gravity of capital, talent, and infrastructure. Ultimately, the trial depicted OpenAI not as a singular act of betrayal but as a gradual, systemic transformation. Its grand AGI mission required a "heavier machine" to sustain it—a machine of computing power (largely from Microsoft), capital, and commercial obligations that inevitably reshaped its priorities. The non-profit board, tasked with guarding the mission, found itself unable to control the commercial juggernaut it had enabled. For the public, the proceedings served as a sobering window into the making of a foundational technology. The AI tools increasingly integrated into daily life—from writing and coding to customer service—are not born from a transparent, purely altruistic process. They emerge from a tangled web of personal ambitions, private negotiations, control struggles, and cloud computing bills. The trial's legacy is the stark realization that as AI becomes societal infrastructure, its steering wheel remains in very few, and very human, hands.

marsbit2 dias atrás 09:06

Is Elon Musk Actually the Victim?

marsbit2 dias atrás 09:06

Pump.fun with Built-in Leverage? A Deep Dive into Hyperliquid's New Project alt.fun

"Hyperliquid's 'Pump.fun'? Decoding alt.fun, a New Project with Built-in Leverage" alt.fun is a new dApp launchpad on HyperEVM, described as Hyperliquid's version of "Pump.fun." It allows anyone to permissionlessly launch meme/altcoins backed by leveraged perpetual contracts (perp-backed altcoins). Instead of a standard bonding curve with spot reserves, each token is directly tied to a leveraged token (LT) representing a Hyperliquid perp position. The token price is driven by both trading activity and the underlying perp's leveraged P&L. Key mechanics: Creators choose an underlying asset (HYPE, BTC, ETH, etc.), direction (long/short), and leverage (2x, 5x). All tokens launch with a fixed $4000 market cap. The bonding curve uses BounceTech's LTs as reserve assets. A "graduation" mechanism automatically migrates tokens to HyperSwap AMM once a USD threshold is met or the curve sells out, with LPs permanently locked. The platform token is $ALT. The project saw over $1M volume in its first hour and brought 300+ new users to HyperEVM. The whitepaper highlights significant risks: leverage exposure (LTs can go to zero), volatility decay for high leverage in sideways markets, and reliance on Hyperliquid/BounceTech infrastructure. It represents an innovative fusion of meme coin virality and real leveraged financial exposure, positioned as an "altcoin factory" and potential traffic gateway for the HyperEVM ecosystem.

marsbit2 dias atrás 09:00

Pump.fun with Built-in Leverage? A Deep Dive into Hyperliquid's New Project alt.fun

marsbit2 dias atrás 09:00

活动图片