Technology Trends

不只会后空翻，波士顿动力的人形机器人Atlas开始“打工”了

marsbitHace 45 min(s)

黄仁勋、Marvell CEO同台对谈：未来AI拼的不是算力是「连接」

marsbitHace 45 min(s)

Can DeepSeek Save China One Trillion Dollars?

"DeepSeek and the $1 Trillion Infrastructure Question" The article examines whether DeepSeek's AI optimization breakthroughs could potentially save China $1 trillion in future AI infrastructure costs. The analysis begins with Nvidia's upcoming Vera Rubin AI platform, costing ~$7.8 million, where memory (HBM4/LPDDR5X) constitutes $2 million—a 435% cost increase in one year, highlighting how AI hardware spending is shifting toward expensive memory components. DeepSeek's approach works in the opposite direction. Through three key technical innovations showcased in DeepSeek V4, the company dramatically improves hardware efficiency: 1. **Memory Compression (MLA)**: Re-engineers the attention mechanism to compress long-context memory (KV Cache) by over 90%, drastically reducing expensive HBM usage. 2. **Selective Activation (MoE)**: Employs Mixture-of-Experts architecture where only a small fraction of parameters (e.g., 49B out of 1.6T in V4-Pro) are activated per token, allowing most parameters to reside in cheaper memory/SSD. 3. **Computation Caching**: Reuses previously computed results via cache hits, replacing expensive GPU computations with cheap memory reads. Combined, these optimizations allow the same hardware to produce approximately 4x more tokens, effectively reducing required hardware investment by 75%. DeepSeek's pricing reflects this: a 10-billion token workload costs ~$522 monthly versus ~$9,000-$10,000 for competitors. The $1 trillion savings projection stems from McKinsey's estimate that global AI infrastructure will require ~$5.2 trillion investment by 2030. As China's daily token consumption grows toward quadrillions, even marginal efficiency gains scale massively. With a conservative 4x throughput improvement, China could avoid building tens of thousands of AI data centers equivalent to ~7 trillion RMB ($1 trillion) in saved investment. Critically, this strategy shifts dependency from scarce, expensive GPU/HBM—where China lags—toward more accessible storage, caching, and systems engineering where domestic suppliers like CXMT are gaining strength. Rather than "replacing Nvidia," DeepSeek rebalances AI's value chain away from monolithic hardware dependency. Ultimately, DeepSeek's technical breakthroughs could lower the barrier to AI adoption across Chinese industries by making advanced capabilities affordable at scale—transforming who can access next-generation AI.

marsbitHace 4 hora(s)

Overturning the Mainstream Approach to Hallucinations: Metacognition is the New Solution for Large Models to Break the Hallucination Barrier

This paper, "Hallucinations Undermine Trust; Metacognition is a Way Forward," proposes a paradigm shift in combating AI hallucination. It argues that the current mainstream approaches—striving for omniscience by scaling data/models or having AI abstain from uncertain answers—are fundamentally flawed. The former has inevitable knowledge gaps, while the latter imposes a crippling "utility tax," requiring the rejection of many correct answers to achieve high accuracy, due to models' poor "discrimination" (the ability to distinguish correct from incorrect answers internally). The core contribution is redefining hallucination not as "being wrong," but as "expressing false information with unwarranted certainty." The proposed solution is **Faithful Uncertainty** or **Metacognition**: enabling AI to accurately perceive its internal uncertainty and honestly express it in its language (e.g., using hedging phrases when unsure). This creates a more reliable assistant that provides useful information while signaling its confidence, minimizing harm from errors. The paper emphasizes that metacognition is critical for the era of AI Agents. Without it, Agents cannot intelligently decide when to use tools like search engines, leading to inefficiency and misuse. Key implementation challenges are highlighted: the "bootstrapping paradox" of training with static uncertainty data, the "alignment distortion signal" where human preference training suppresses internal uncertainty cues, and the difficulty of causally evaluating true metacognition vs. its superficial imitation. The paper concludes that the goal should not be an infallible AI, but one that is honest about the limits of its knowledge, thereby building user trust through transparent communication of its certainty.

marsbitHace 4 hora(s)

After 540,000 Lines of Code, Garry Tan Realizes the Old Game of AI Programming Is Over

YC President Garry Tan's project "Garry's List" involved over 540,000 lines of Rails code, but he concluded this approach is outdated. He argues the software industry is stuck in a "Foxconn factory" mindset: building excessive tests, validators, and control logic to constrain LLMs, which have become cheap and capable enough to work autonomously. The old paradigm treated LLM calls as expensive and code as cheap, leading to complex systems to "manage" the AI. This has now reversed. The future lies in "just-in-time software"—using natural language instructions (Markdown-based "skill packs") and minimal code, letting the AI handle the work. Tan advocates for "skillifying" workflows: after an agent completes a task, it packages the process into a reusable, tested skill pack. This shifts value from writing code to designing capabilities. For example, his agent reviewed 85 hackathon submissions in 30 minutes, a task that previously took days. He emphasizes "tokenmaxxing"—willingly spending on LLM tokens to gain a years-long competitive advantage, as costs are plummeting. The core shift is from measuring output in lines of code to focusing on clarity, taste, and judgment. The best future engineers won't write the most code, but will know what to build and how to unlock the most intelligence with the least code.

marsbitHace 7 hora(s)

Pantera Partner: In the Age of Agents, Blockchain is the Inevitable Answer for AI

Summary: AI and blockchain are converging around four key pillars: payment settlement, identity systems, open systems, and resource aggregation, with commercial projects already emerging in each area. The two technologies are fundamentally complementary: AI enables infinite supply (content, agents), while blockchain establishes scarcity and verifiable ownership. AI agents generate content and services, and blockchain handles the verification and value settlement. A significant valuation mismatch exists, with leading AI companies historically overvalued compared to crypto assets, despite their deep underlying integration. The emergence of autonomous AI agents—which require assets, value transfer, and large-scale coordination—creates a need for a non-human-centric financial infrastructure. Blockchain, with its programmability, 24/7 access, and low-trust settlement, is the only suitable foundation. AI agents will not use traditional bank accounts or payment rails; they will transact using stablecoins and on-chain systems. Examples include OpenFX, which settles hundreds of billions in forex trades on-chain for AI agents, and Alchemy, a core development platform. For human identity verification in an age of AI-generated content, projects like World (Worldcoin) use blockchain-based biometric verification, while TransCrypts focuses on self-sovereign identity and verifiable credentials. The current divergence presents a unique investment opportunity. AI valuations are highly elevated, while crypto assets trade at a significant discount, even though the future smart agent economy will be built on blockchain infrastructure. The fusion of AI and blockchain is not a future trend but an ongoing reality, creating a prime environment for entrepreneurs in areas like agent-native finance, decentralized identity, and on-chain AI coordination.

marsbitHace 15 hora(s)

Chatbot has been burning money for three years, is it still the 'New Continent' of the AI era?

For years, the AI industry has been guided by a singular "map" — the belief that the AI era's "new continent" would be found in the Chatbot, a super-app akin to the mobile internet's super-apps. This belief was fueled by ChatGPT's explosive 2022 debut. However, three years of heavy investment reveal a different reality: the Chatbot-as-ultimate-entry-point model is struggling. The core issue is economic. Chatbots defy traditional internet economics. Unlike apps with near-zero marginal cost, each AI query consumes significant, expensive compute. More users mean higher costs, not profits. OpenAI, despite ~900M weekly active users, reportedly loses money. The expected network effects and data flywheels that power internet giants are weak in Chatbots, as one user's interactions don't improve another's experience. Monetization is a major hurdle. The subscription model faces low conversion rates, especially in China where users expect AI to be free. The "free + ads" model also struggles. Chatbot interactions often lack commercial intent, and inserting ads compromises the trust essential for an answer engine. Perplexity's minimal ad revenue and subsequent pivot away from ads highlight this difficulty. Switching between Chatbots is easy, making user loyalty low and competition a potential race to the bottom on price. Data suggests the standalone Chatbot's growth is slowing, and user engagement (avg. ~6 mins/day) pales compared to apps like TikTok. The product form itself is limiting; studies show nearly half of interactions are simple Q&A, trapping AI's potential in a passive, single-turn "cage." A contrasting, more successful path is emerging, exemplified by Anthropic. With over 85% of its ~$30B annualized revenue from enterprises, it focuses on AI as a productivity tool, not a companion. The rise of AI Agents (like OpenClaw) and the integration of AI into existing workflows (e.g., Google's AI Overviews, Apple Intelligence in OS) signal a shift. The future may not be a dominant Chatbot app, but AI embedded seamlessly into social apps, operating systems, and hardware — a capability-layer revolution, not a new distribution container. The conclusion is clear: the old "map" centered on a standalone Chatbot super-app is leading to a dead end. To find the true valuable "continent" of the AI era, the industry must update its navigation to prioritize deep integration, practical utility, and sustainable economics over a generic conversation window.

marsbitHace 18 hora(s)

The Death of the Three-Act Play: AI Ushers Enterprise Software Startups into the ‘Speedrun Era’

The Death of the Three-Act Play: How AI is Ushering in a 'Speedrun Era' for Enterprise Software Startups The traditional three-act play for building an enterprise software company—first, a niche wedge product; second, an expanded suite; third, a dominant platform—is becoming obsolete in the AI era. Previously, startups would spend 3-5 years perfecting a single-point solution to reach tens of millions in ARR (Act 1: The Wedge). Then, over another few years, they'd build adjacent products to form a suite and cross the $100M ARR threshold (Act 2: The Suite). Finally, with scale and user engagement, they could aim to become a foundational platform themselves (Act 3: The Platform). This model assumed a timeline measured in years. However, AI-driven tools have dramatically compressed software development costs and timelines. Companies like Cursor, Clay, and Harvey have scaled from near zero to approaching or surpassing $100M ARR in remarkably short periods, demonstrating a new competitive pace. The core argument is that in this rapidly changing market, relying on a small, "safe" wedge as a protective harbor may now be a conservative, even risky, strategy. The plummeting cost of building software means the time required for Acts 1 and 2 is approaching zero. Consequently, rational strategy now favors planning to build the entire vision from the outset. This shift changes the calculus for early-stage investment. The emphasis is moving from finding a defensible niche to backing founders with "unreasonable, relentless ambition" to reimagine entire workflows or replace incumbent platforms from day one. The age of gradual expansion is giving way to an era of immediate, full-scale ambition.

marsbitHace 20 hora(s)

Issued Two Work Badges to Unitree

At the keynote of his speech at the Taipei Music Center, Jensen Huang introduced a humanoid robot named Isaac GR00T. This robot, described as a 'reference design,' is a collaboration: its body comes from Unitree Robotics' H2 Plus, its hands from Singapore's Sharpa, and its 'brain'—the chip and full software stack—is from Nvidia, powered by the Jetson Thor. Huang positioned it as a turnkey solution for universities and researchers, aimed at drastically reducing setup time for experiments. On the same day as this reveal, Unitree Robotics passed its IPO review in Shanghai, seeking to raise 4.2 billion yuan, with a significant portion earmarked for developing its own embodied AI model—its own 'brain.' The article draws a parallel to the smartphone industry, where Qualcomm's 'reference design' led to homogenized hardware and concentrated profits in chips and software. It suggests Nvidia's GR00T initiative follows a similar playbook: by open-sourcing the model and framework, it aims to establish the industry standard, potentially relegating hardware makers to low-margin roles. While currently a body supplier for Nvidia's project, Unitree is actively pursuing its own AI brain, having open-sourced initial models and tested a more advanced one. The company faces a critical window to develop a competitive proprietary system before GR00T becomes the default. The article contrasts this with Tesla's vertically integrated approach for its Optimus robot, which uses in-house chips and benefits from its automotive data and manufacturing scale. It concludes that while the robot body still holds technical value and differentiation, the race for the 'brain' will ultimately define the industry's profit centers and power dynamics.

marsbitHace 23 hora(s)

Silicon Valley's New FDE Role Gains Popularity: What Kind of AI Talent Do Enterprises Need?

The Silicon Valley job market is seeing a resurgence of the Forward Deployed Engineer (FDE) role, originally pioneered by Palantir, now focused on AI. Companies like OpenAI and Anthropic are deploying AI FDEs to client sites to customize general-purpose large language models into specific Agent workflows for business processes. However, the article argues that the broader and more significant trend is the rise of the in-house "AI Engineer." These professionals, needed in far greater numbers than external FDEs, integrate AI capabilities—like prompt engineering, Agent frameworks, and evaluation systems—directly into a company's own software and operations, often using AI coding tools. This evolution suggests AI's impact on employment is not simply about replacement but about creating new, generalist roles that will later specialize, much like software engineering split into front-end, back-end, and DevOps. The future is likely to see a proliferation of specialized AI engineering roles (e.g., LLMOps, Evals Engineer), but the current and critical demand is for engineers who blend technical skill with business understanding to embed AI effectively, ensuring companies retain flexibility in their technology choices.

marsbitAyer 04:40