독서리서치레이더투자 프레임워크
로그인 / 회원가입
전체기술제품투자상장시장기타

독서

2026-11-0626 게시물

새로고침

Mistral Science VP Pierre Stock on ML4 Training: Reinforcement Learning, Ten Megawatts of Power, and European Sovereignty

중대
2026-11-06 22:00 GMT+8

Pierre Stock, Vice President of Science at Mistral, provided a detailed explanation in the heise KI-Update podcast on how this French AI laboratory trains frontier models under extreme compute constraints, and why he considers open weights to be an issue of sovereignty for Europe rather than a business model.

Stock identifies the core of ML4's training as the art of difficulty calibration within reinforcement learning. He uses an exam analogy: the model solves problems independently and receives rewards for complete or partial solutions. The entire art lies in setting the correct difficulty level—too simple, and no learning occurs; too difficult, and there is no signal. This RL approach enabled Mistral to achieve significant progress between the preview version released on October 6, 2026, and the open-weight version on October 29: the final version consumed 50% fewer tokens for equivalent intelligence, while the agent task resolution rate jumped from approximately 6% in the previous generation to nearly 60%.

This training is supported by data centers built by Mistral in France, though Stock admits their scale: ML4 was trained using only ten megawatts of power, which he describes as laughably small compared to competitors. On October 29, 2026, Mistral released the weights for ML4—a mixture-of-experts model with one trillion parameters and 49 billion active parameters per token, internally codenamed Le Chonk. Stock notes that ML4 is three times smaller than Moonshot AI's Kimi K3; quantized in FP4, it can run on four GB300 GPUs, making self-deployment possible for universities and research laboratories.

Regarding performance, Stock claims ML4 is currently the best open-weight model for coding and cybersecurity. Benchmarks published by Mistral partially support this claim: ML4 scored 48% in the Artificial Analysis Coding Agent Index, essentially tying with Zhipu's GLM-5.3; it scored 79% in the CyberGym end-to-end benchmark, comparable to Xiaomi's MiMo-V2.6-Pro. However, Stock acknowledges that ML4 still lags behind the strongest models in coding: other companies currently possess one hundred times more resources than we do. He simultaneously claims that ML4 outperforms Anthropic's Opus 4.8, released in late May, believing Mistral is only months behind its competitors.

Stock positions open weights as the technical foundation for European sovereignty. His argument equates access to intelligence (i.e., tokens) with access to electricity: without local power production, the Industrial Revolution would not have succeeded. He states bluntly that if we abandon frontier models, we are essentially saying goodbye to the idea of Europe having a say as a superpower in the future. This stance creates tension with recommendations from European Commission President Ursula von der Leyen, who tends to favor focusing Europe on AI applications rather than developing frontier models herself. Mistral CEO Arthur Mensch had previously labeled such a route as capitulation.

On the commercial front, Mistral raised three billion euros in its latest funding round, valuing the company at over 21 billion euros, and has opened its platform for third-party models (such as China's GLM-5.3). Stock emphasizes that the company remains a frontier lab, with model development as the core; the platform strategy merely expands coverage of token demand to feed back into data center investments. For enterprise customers, the value of open weights lies in the fact that no one can cut off your access—Mistral deploys models on customers' own GPUs, after which it neither sees the model nor the data. Mozilla has already made Mistral Small 4 available as one of the optional models in Firefox's Smart Window (currently in beta).

출처:heise online

리서치

German Study Warns: AI Data Centers Could Double Water Use to 27 Billion Liters by 2030

주제 · 德国数据中心耗水争议중대
2026-11-06 16:13 GMT+8

A new study published by the Freiburg-based Öko-Institut, commissioned by Greenpeace, warns that the expansion of data centers driven by the AI boom threatens to severely strain regional water supplies in Germany. The research indicates that German server farms consumed approximately 13 billion liters of water for cooling in 2025. By 2030, this demand could more than double to up to 27 billion liters per year due to new mega-data centers, a volume equivalent to the annual drinking water consumption of a city with around 580,000 inhabitants, such as Essen.

These calculations stand in stark contrast to the position taken by the digital industry association Bitkom, which recently classified water usage as largely unproblematic. Bitkom stated that nearly two-thirds of local facilities consume a maximum of 0.1 liters of water per kilowatt-hour of connected power, arguing that relevant consumption is limited mainly to evaporative cooling. However, after correcting erroneous datasets from the official government data center register, the Öko-Institut arrived at a weighted average of 0.67 liters of water per kilowatt-hour of IT electricity consumption. The current annual usage of 13 billion liters already corresponds roughly to the total water content of Berlin's Wannsee lake.

The study highlights particular risks posed by planned large-scale projects. Currently, nearly 30 data centers with connection capacities ranging from 100 megawatts to over 1 gigawatt are in planning in Germany. Experts estimate that a single 100-megawatt data center requires between 200 million and 930 million liters of water annually, depending on the cooling technology used—comparable to the drinking water needs of a small town of up to 20,000 people. In already dry regions, these projects could cause acute groundwater stress, distribution conflicts affecting local communities, and damage to local ecosystems.

Furthermore, the investigation refutes a central argument by operators that low energy consumption must technically be paid for with higher water usage. The study found no statistical correlation between power and water efficiency in practice; high energy efficiency can be achieved entirely without water cooling. By employing modern dry or compression cooling systems, direct water consumption in server farms can be reduced to zero. Greenpeace also criticized significant transparency issues, noting that less than half of the approximately 1,200 obligated operators submitted data to authorities, and called for mandatory water assessments before permits are granted and bans on water-intensive facilities in dry areas.

출처:heise online

리서치

From HRT Internship to Co-Investment by Founders Fund: The 'New Mafia' Network in the AI Industry Takes Shape

주제 · HRT AI人才网络중대
2026-11-06 18:30 GMT+8

In the winter of 2016, ten young people from Harvard University and the Massachusetts Institute of Technology (MIT) met during a three-week internship program at Hudson River Trading (HRT), a Wall Street quantitative trading firm. Ten years later, at least four of them are leading or managing AI and crypto-finance companies with a combined valuation exceeding $50 billion: Alexandr Wang, Chief AI Officer at Meta; Scott Wu, co-founder of AI coding company Cognition (valuation $48 billion); Jeff Yan, founder of on-chain derivatives platform Hyperliquid; and Jesse Zhang, co-founder of AI customer service company Decagon (valuation $4.5 billion).

Looking at these numbers alone, it would merely be a coincidence of several smart individuals succeeding independently. What gives the label 'HRT Gang' analytical value is the emergence of specific business connections. The first clue lies in talent mobility: Steven Hao, co-founder of Cognition, appeared on the engineering team list of Scale AI in 2018, introduced in an article written by Alexandr Wang himself; several years later, he co-founded Cognition alongside Scott Wu and Walden Yan. This indicates a traceable channel of personnel between Scale AI and Cognition, rather than just parallel entrepreneurship by the same group of people.

The second clue is at the capital level. Founders Fund, a venture capital firm co-founded by PayPal co-founder Peter Thiel, appears simultaneously in the existing investor lists for Scale AI's $1 billion financing round in May 2024 and Cognition's Series E financing round in September 2026. This is not easily explained away as coincidence: a fund founded by core figures of the previous generation of Silicon Valley power networks happens to bet on two of the most important companies in the 'HRT Gang' simultaneously.

Putting these two clues together, the structure indeed resembles the early stages of the PayPal Mafia. After eBay acquired PayPal for $1.5 billion in 2002, former colleagues maintained business ties in the tech industry: Peter Thiel participated in founding Founders Fund and became Facebook's first external investor; Reid Hoffman founded LinkedIn; David Sacks founded Yammer. The key is not what they each founded, but how they formed a network that continuously generates business opportunities through joint investments, talent referrals, and board seats.

However, there is a structural gap between the current HRT Gang and the PayPal Mafia. Members of the PayPal Mafia not only started their own companies but also continuously participated in the financing of next-generation companies through institutions like Founders Fund, forming a triangular closed loop of entrepreneurship-investment-management. The HRT Gang currently only has two lines: talent mobility and common investors. There is no evidence yet of HRT alumni founding their own investment funds or systematically injecting capital into each other's companies. Alexandr Wang jokingly suggested on a podcast that HRT founders directly invest in the interns from that year, and the official HRT account subsequently publicly responded welcoming alumni to contact their investment team—this suggests the closed loop may be forming, but is not yet complete.

출처:zhidx.com

리서치

World Models: The Core of Physical AI Lies in Predicting the Next Frame, Not the Next Word

주제 · 世界模型与物理AI중대
2026-11-05 11:12 GMT+8

In a demonstration released by Google DeepMind on October 27, 2026, for its Genie 4 system, a video was generated via text prompt: a crowned crane walks across a snowy landscape at dusk toward a white piano, pressing keys with its beak, accompanied by both piano notes and the sound of snow crunching. The technology behind this video is called a World Model. Its quality determines whether the scene is believable: if performance is insufficient, the sound may play before the beak touches the key, or the beak might pass directly through it; if performance is sufficient, the movement appears natural and physically consistent, resembling real footage.

A World Model can be defined as an AI that learns how the real world moves from data such as videos, thereby predicting what will happen next given a current state and an action. Byline Network reporter Oh Min-seon provides a clear analogy in her article: if language models like ChatGPT predict the next word, a World Model predicts the next frame. This distinction is critical for physical AI: robots need to understand not language, but the consequences of actions.

Physical AI—which enables entities like robots to act autonomously in the real world—relies on World Models because robots must understand basic physical laws such as gravity, velocity, and force. Oh Min-seon points out that the robotics industry currently applies World Models in two directions. First is generating training data: robots can learn not only from real-world collected data but also from synthetic data generated in virtual environments. Previously, these virtual environments were primarily built by developers manually inputting physical rules; the difference with World Models is that they learn these patterns directly from data rather than relying on human-defined rules, thus generating more realistic synthetic data. However, because they render scenes based on learned patterns rather than calculating physics, the results may look plausible but lack physical consistency. Second is planning actions during actual execution: Vision-Language-Action (VLA) models allow robots to decide what to do based on camera feeds and human instructions. By integrating a World Model, robots can virtually test the consequences of multiple actions before acting in reality, selecting the one most likely to succeed.

Google DeepMind demonstrated a combination experiment in a Korean-language blog post released on October 28: pairing Genie 4 with Gemini 4 Argon to simulate Apptronik’s Apollo humanoid robot (from the United States) playing checkers. After a user issues a task command, Gemini 4 Argon breaks it down into multi-step actions, and Genie 4 generates corresponding virtual world visuals for each step. However, the article explicitly notes this remains a demonstration in the early research stage, with no results released regarding actually training a robot using this video. NVIDIA’s Cosmos Policy, made public in January this year, showcases the second direction: the World Model not only predicts visual changes after an action but also evaluates the likelihood of goal achievement, enabling robots to compare expected outcomes of multiple candidate actions.

출처:바이라인네트워크

리서치

Germany's NRW State Deploys First Dedicated AI System 'Forensight' to Cut Financial Crime Evidence Analysis from Years to Weeks

주제 · Forensight중대
검증됨 2026-11-06 17:34 GMT+8

The financial administration of North Rhine-Westphalia (NRW), a state in Germany, has announced the first use of a dedicated artificial intelligence system named Forensight to evaluate digital evidence in financial crime investigations. The program was jointly developed by the NRW Office for Combating Financial Crime and the Fraunhofer Institute for Intelligent Analysis and Information Systems (IAIS). According to a spokesperson for the NRW Ministry of Finance, Forensight translates documents, summarizes content, transcribes audio and video files, identifies persons, companies, and locations, and maps their connections as networks.

The initiative addresses the challenge of data overload faced by investigators. The spokesperson noted that searches and seizures in financial crime cases regularly generate millions of files, including documents, chats, images, and recordings, yet often less than ten percent are relevant to the investigation. Manual evaluation of such volumes ties up investigators for months or even years.

Data provided by the NRW Ministry of Finance indicates that in one pilot project, investigators used Forensight to identify 1,000 relevant pages out of a corpus of 20 million within just a few weeks—a task that would have taken years manually. In further test cases, the system helped investigators find approximately 60 percent more relevant documents than previous methods.

NRW Finance Minister Marcus Optendrenk stated that effectively combating financial crime requires professionalism in data analysis comparable to the concealment tactics used by perpetrators. He emphasized that while Forensight provides a tool to filter decisive traces from millions of datasets, the final assessment and decision-making remain with human investigators, who must review, evaluate, and decide.

출처:n-tv.de

리서치

OpenAI Releases National Community Compact and Policy Blueprint, Reaffirms Full Data Center Power Cost Payment and Calls for Legislation

주제 · OpenAI社区协议중대
2026-11-06 13:01 GMT+8

U.S. artificial intelligence company OpenAI released the "National Community Compact" and the "Compute Campus Policy Blueprint" on November 5 Eastern Time. These documents aim to address voter opposition to data center construction observed during the recent U.S. midterm elections. OpenAI reaffirmed its commitment to fully pay for the electricity costs of its owned and operated data centers (termed "compute campuses") and pledged not to require government officials to sign non-disclosure agreements (NDAs) during negotiations.

According to the compact, OpenAI commits to covering the incremental costs of infrastructure required to serve the campus and actual electricity expenses, providing financial guarantees before major expenditures begin to prevent other users from bearing costs if projects are delayed, scaled down, or cancelled. This reiterates the "self-funded energy costs" promise made in January's Stargate community initiative and aligns with Anthropic's earlier pledge to pay for grid upgrades. However, the agreement's scope is explicitly limited to OpenAI's "owned and operated" future campuses. For facilities leased or co-developed with partners like SB Energy, responsibilities remain governed by specific project terms. For instance, in the Ohio PORTS-Pike project, SB Energy holds and operates the facility while OpenAI acts as a customer, meaning its power cost obligations revert to that specific contract.

Beyond voluntary commitments, OpenAI actively calls for legislation to codify these measures in the blueprint. The document suggests that state utility commissions establish dedicated tariffs for large-load customers, requiring them to pay for the incremental costs they cause, and mandates the inclusion of the "Ratepayer Protection Pledge" in law. Additionally, OpenAI requests that Congress grant the Federal Energy Regulatory Commission (FERC) clearer authority to expedite interconnection approvals for large projects and expand federal backup approval rights for key transmission plans. Chris Lehane, OpenAI's Chief Global Affairs Officer, stated that these proposals were developed after meetings with dozens of members of Congress, aiming to build momentum before the new session begins in late January.

While OpenAI describes the compact as an accountability framework for communities to measure progress—including annual public reports on water usage, electricity consumption, and noise—the document itself lists no penalty clauses or third-party audit mechanisms. In contrast, its Project Camellia in Georgia is already bound by state Public Service Commission rules prohibiting cost pass-through to existing users. This dual strategy of "voluntary pledges plus legislative advocacy" reflects AI infrastructure companies' attempt to reduce regulatory uncertainty through institutionalization amidst political resistance, while also highlighting complex industry debates over cost allocation and liability definitions.

출처:INSIDE

리서치

South Korean Tech Firm Kakao Releases Open-Source Decision Model Kanana-2-v-decide: Outputs Probabilities Instead of Text for AI Agents

중대
2026-11-06 09:16 GMT+8

South Korean technology company Kakao has released Kanana-2-v-decide, an open-source multimodal decision model designed specifically for AI agents. Unlike standard Large Language Models (LLMs) that generate text responses requiring further parsing, Kanana-2-v-decide outputs structured data such as probabilities, choices, or scores directly. This allows AI systems to bypass the step of extracting conclusions from natural language, enabling faster execution of subsequent actions.

The model is built upon Kakao's multimodal language model Kanana-2-v-3b and utilizes an embedding-based architecture. It converts questions and options into vectors to determine similarity, which reduces computational load and response time for repeated requests. Kakao states that the model responds within an average of 0.04 seconds on a single H200 GPU, approximately 43 times faster than OpenAI's GPT-6 Luna.

In terms of accuracy, Kakao evaluated the model using its proprietary 'Kanana Decision Bench' (15,909 Korean items) and the public 'Decision Index'. The company reports that Kanana-2-v-decide achieved the highest scores among compared models, including Microsoft's Decision-1, TypeSafe AI's Jev, Perplexity's pplx-decider-v1.1-27B, and Upstage's Solar Decide. These performance metrics are self-reported by Kakao and have not yet been independently verified.

The model is available under the 'Kakao Open License', permitting both research and commercial use. Kakao plans to integrate it into internal projects such as web search and product category classification before broader deployment. The release addresses a growing need in the AI industry for specialized 'decision-only' models that can handle the rapid, structured judgments required by autonomous agents, distinct from the generative capabilities of general-purpose LLMs.

출처:스타트업 생태계의 모든 것 \'플래텀(Platum)\'

리서치

OpenSSH Releases Emergency Update 10.6p2 to Fix 'SSHole' Authentication Bypass Vulnerability Affecting musl-Based Linux Systems

주제 · OpenSSH중대
2026-11-06 11:43 GMT+8

The OpenSSH development team released an emergency update, version 10.6p2, on November 5, 2026, to fix a high-severity authentication bypass vulnerability named "SSHole" (CVE-2026-123158). The vulnerability has a CVSS severity score of 8.1 and allows attackers to potentially log into accounts that have no password set by entering any string between 257 and 1,024 characters in the password field during an SSH connection.

The flaw traces back to OpenSSH 7.3p1, released in 2016, meaning versions from the last decade may be affected. However, SSHole does not impact all users. It specifically affects Linux operating systems that meet three conditions: they use the musl C library, OpenSSH does not have PAM (Pluggable Authentication Modules) enabled, and SSH password authentication remains active. Major affected distributions include Alpine Linux, OpenWrt, and postmarketOS, as well as certain Docker images based on these systems.

Security research firm Wakefield identified that the vulnerability resulted from an unintended combination of security adjustments across three independent projects: musl restricted password length in 2012 to prevent denial-of-service attacks; OpenSSH modified password verification for some accounts in 2016 to avoid timing side-channel leaks; and Alpine adjusted the root account's password field settings in 2022 to facilitate SSH key login. These changes, each solving a different problem, ultimately created exploitable authentication bypass conditions. Wakefield discovered SSHole with the help of AI tools on November 3 and reported it to the relevant teams, leading to rapid patch releases from both OpenSSH and Alpine Linux.

For administrators unable to immediately update OpenSSH, Wakefield provided temporary mitigation measures: disable the PasswordAuthentication option in the SSH server configuration and restart the service. Additionally, administrators should check SSH login logs; if the password field shows "*" and an "Accepted password" record appears, it should be treated as a potential sign of SSHole exploitation, requiring further investigation into system compromise.

출처:iThome

리서치

AI Company Anthropic Warns in IPO Filings That R&D May Be About to Accelerate Dramatically; Frontier Model Astra Demonstrates Zero-Shot Direct Editing of Other Models' Weights

주제 · Anthropic구조적
2026-11-06 02:37 GMT+8

Anthropic's amended S-1 filing now states that the company may be imminently meeting the definition of 'dramatically accelerated progress attributable to AI' found in its Responsible Scaling Policy.

This warning is not a generic risk disclosure. Anthropic's Responsible Scaling Policy defines this threshold as: models being able to fully replace all research scientists and engineers at competitive costs, or a dramatic acceleration in the pace of AI progress attributable to the automation of R&D.

Specific technical support for this assessment comes from recently published experimental results. Researchers Rikiya Takehi et al. reported that GPT-6 Astra can directly read and edit the raw weight parameters of other neural networks, requiring no gradients, no training, and no tool calls. They termed this 'Weight-Space Programming.'

These results require clear attribution: they are benchmarks run by researchers themselves and have not yet been independently replicated. However, Zvi Mowshowitz noted in his analysis that the significance of these experiments lies not in a specific numerical improvement, but in the demonstration of a mechanism—frontier models can treat other models' weights as readable and writable code to diagnose and patch them. If this capability continues to strengthen, a self-accelerating cycle in AI R&D may already have begun.

The strongest counterargument is the scale issue. Astra edited a small model with 0.8 billion parameters, primarily fixing a specific defect in reasoning loops. For models that are already functioning well, Astra could only provide marginal improvements. There is a massive unverified gap between generalizing this technique from patching known issues in small models to autonomously designing next-generation frontier models. Anthropic itself acknowledges in the S-1 that 'we do not anticipate being able to safely use Claude to fully autonomously design and build successive generations of frontier models.'

But the core of Zvi's argument is that Anthropic, as the lab most likely to trigger this threshold first, proactively wrote this sentence in a legal document filed with investors, indicating that the company has internally observed signs of acceleration. Previously, Anthropic stated that they often only identified such accelerations after they had occurred. This time, they are attempting to warn before it happens.

출처:thezvi.substack.com

리서치

Brazilian Court Bans Meta from Using Media Content for AI Training, Imposes Daily Fine

주제 · Meta巴西媒体版权禁令중대
2026-11-06 06:22 GMT+8

A court in São Paulo, Brazil, has issued an injunction prohibiting US tech giant Meta from using content belonging to member companies of three Brazilian press associations to train its artificial intelligence (AI) models. The ruling directly targets Meta's previous attempt to force media organizations to license their content for free in exchange for continued platform access through modified terms of service.

According to heise online, Judge Guilherme Madeira Dezem of the 44th Civil Court in São Paulo granted the interim relief requested by the National Newspaper Association (ANJ), the Brazilian Association of Radio and Television Broadcasters (Abert), and the National Association of Magazine Publishers (Aner). These associations alleged that Meta's new terms of service for business users made the "free and unrestricted provision of copyright-protected material" a condition for continuing to use its platforms. The judge stated in the ruling that this requirement constituted an obvious abuse of rights.

In addition to suspending the execution of the relevant clauses, the court also banned Meta from sanctioning or penalizing media companies that refuse to accept these new terms. If Meta violates the injunction, it will face a fine of 1 million reais per day (approximately 162,000 euros). This decision is seen as an initial victory for the Brazilian media industry in protecting intellectual property and maintaining the sustainability of professional journalism.

출처:heise online

리서치

How Amazon Marketplace Evaluates Content Quality with Frozen Vision-Language Models and OCR Signals

주제 · 亚马逊A+内容质量评估중대
2026-11-06 02:06 GMT+8

A+ Content is the rich media widget on Amazon product detail pages, where brands use combinations of images, text, and video to tell their product stories. The challenge is that checking each asset in isolation—such as verifying image clarity or grammar errors—misses cross-module quality issues: for example, a feature list might advertise a 64-ounce capacity while a lifestyle photo displays a single-serving model. To address this, the Amazon team renders the entire page as a single snapshot, allowing the vision-language model to evaluate the content holistically, much like a customer would.

A core architectural decision is to not fine-tune the vision-language model. The team reasoned that fine-tuning a base model with only a few thousand labeled samples requires GPU training infrastructure and carries a risk of overfitting. As an alternative, they keep the VLM and its scoring rubric fixed. The model outputs scores from 0 to 10, confidence values, and natural language rationales across four dimensions: Visual Storytelling, Feature Presentation, Layout, and Conversion Optimization. These structured outputs are then used as features to train a CatBoost gradient-boosted decision tree classifier to predict expert-labeled high-quality or low-quality tags. Because only the classifier is retrained, updates required as annotations accumulate or criteria change can be performed as a short CPU-only task.

While vision-language models excel at capturing overall semantics and visual composition, they are less reliable for fine-grained typographic readability; they often overlook layout signals such as text density, line spacing, and character crowding. To capture these details, the team runs an OCR model in parallel to extract text blocks and their positions from the snapshot, calculating four readability metrics: Text Density, Character Density, Spacing, and Overall Readability. The fusion model inputs eight features—the four VLM scores and the four OCR metrics—into the same CatBoost classifier. On a 20% holdout set, the classifier using only VLM features achieved 89% accuracy and an F1 score of 0.90, while the one using only OCR features reached 82% and 0.83. The fused model achieved 94% accuracy and an F1 score of 0.95, representing a 5 percentage point improvement over the VLM-only approach.

Production validation of this method included several key steps: using SHAP to explain the contribution of each feature to decisions, feeding the most negative features along with product category and best practices into a language model to generate prioritized improvement suggestions, and running global A/B experiments that adjusted the visibility of low-quality content based on model judgments, which resulted in a statistically significant increase in Ordered Product Sales. Sellers who revised their content using these suggestions saw an average quality score increase of 1.4 points (on a scale of 10). It is important to note that all figures above are self-reported by Amazon, based on internal annotations and holdout sets, without independent replication. The annotations covered 10 product categories, and whether this method can be generalized to other multi-module content evaluation scenarios remains unverified.

출처:Amazon Web Services

리서치

Databricks Application Scaling Guide: Separating Soft and Hard State to Decide Between Adding Machines or Instances

중대
2026-11-06 02:00 GMT+8

Bruno Tafur and Tushar Madan of Databricks published an extension guide for application builders in November 2026, discussing how to maintain fast response times and high availability as a Databricks App grows from a few users to hundreds or thousands. The core judgment of the article is that the intuitive answer 'add more compute resources' is only effective once you know where requests spend their time and whether the application itself is built to scale.

The authors divide the scaling process into four stages: design, build, test deployment, and operations. In the architecture design stage, the most critical decision is dividing application state into two categories. Soft state is temporary working data, such as in-memory context, caches, and ongoing work, which can be rebuilt from hard state if lost; hard state consists of business records, checkpoints, and user preferences, which must be persisted in Lakebase or Unity Catalog tables so that any instance can recover the session. This classification determines whether horizontal scaling is truly feasible—if session state exists only in memory, users will lose their work when instances restart or migrate.

The article provides a specific retailer scenario to demonstrate load estimation: during the busiest hour of Black Friday, 200 out of 2,000 store managers are using an inventory query app simultaneously, with each performing two queries per minute, each executing three SQL statements. The result is 1,200 statements per minute, averaging 20 per second. The authors emphasize that this estimates query rate rather than concurrency; testing needs to validate burst traffic and growth, and database connections should be budgeted separately.

Regarding connection budgeting, the article provides an actionable formula: when each worker maintains a connection pool, the upper limit of connections is the number of instances multiplied by the number of workers per instance multiplied by the maximum pool size. For example, with five instances, eight workers each, and a pool limit of ten, up to 400 connections can be opened. After adding other clients and connections opened on behalf of users, this total must be compared against database documentation limits, leaving headroom for deployments. For scenarios requiring queries to Lakebase on behalf of users, each user requires an independent Postgres role, and Lakebase limits branches to 500 roles; adding instances does not change this upper limit.

The core recommendation for the build phase is to use asynchronous code. Most applications spend time waiting for external services—SQL queries, model calls, and external APIs. Blocking calls occupy an execution slot for the duration of the wait, meaning a few slow calls can exhaust capacity. When handlers await model and database calls, one worker can continue serving other requests while waiting. The article also recommends using asyncio.gather or Promise.all to launch independent queries in parallel, making request latency close to that of the slowest query rather than the sum of all queries, while limiting parallelism to avoid overwhelming the backend.

출처:Databricks

리서치
다음 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크

Stock also acknowledges limitations regarding safety. In benchmarks, ML4 refuses malicious operations more frequently than Chinese competitor models, but he states this is not magic; protection mechanisms can be bypassed, and the cost of bypassing should be comparable to the cost of training a model from scratch. He advises against over-relying on benchmarks—every benchmark can be gamed—and suggests letting models of equal capability monitor each other.

It should be noted that all the above performance data, power scales, and efficiency improvements are self-reported by Mistral and have not been independently reproduced. The claim that a trillion-parameter model was trained using only ten megawatts of power would be highly significant if true, but currently lacks third-party verification. The sovereignty argument directly links base model capabilities to geopolitical influence, an equation itself worth scrutinizing—whether application-layer advantages might offer greater practical economic leverage than model-layer autonomy is a path not fully discussed by Stock in the podcast.

For readers concerned with the landscape of the AI industry, the practical significance of this analysis is: when any of Cognition, Scale AI, Hyperliquid, or Decagon releases a major announcement, it is worth simultaneously checking whether the other three show corresponding talent changes, capital moves, or product synergies. If this frequency of cross-interaction increases significantly over the next two to three years, the 'HRT Gang' will cease to be just a media label and become a trackable power structure in the AI industry. If it remains at the level of independent operations, then the shared internship experience is merely narrative decoration.

The boundaries to retain are: all facts in this article come from a synthesis of secondary reports, including character profiles from the Financial Times, interviews with Jeff Yan and Scott Wu by Colossus, disclosures from Fortune magazine regarding Zuckerberg's recruitment process, and income estimates for Hyperliquid by Grayscale. The fact that Founders Fund invested in both companies is information from publicly disclosed investor lists, but their specific shareholding percentages, entry times, and additional terms have not been disclosed. Inferring 'network formation' from individual talent mobility and common investors is the author's analytical judgment, not an established organizational fact.

The fundamental limitation of World Models lies in the Sim-to-Real Gap—the disparity between simulation and reality. Even if a robot stably picks up a glass cup in a virtual environment, differences in the cup's weight, surface friction, and finger pressure in the real world could cause it to drop. Virtual training does not equal real-world deployment. Furthermore, error accumulation increases the longer the prediction horizon; Google DeepMind itself has noted that Genie 4 may experience object appearance and position drift under long sequences or complex instructions. Precise simulation of complex physical phenomena remains an unsolved challenge.

For readers, the practical contribution of this article is providing a judgment framework: World Models are a key enabling technology for physical AI, shifting robots from passively executing preset rules to actively predicting and planning the consequences of their actions; however, they are currently in a transitional phase from concept validation to actual deployment. The Sim-to-Real Gap and error accumulation are the core variables determining whether this technology can truly change the pace of the robotics industry.

Regarding graceful shutdown, Databricks sends SIGTERM upon redeployment, followed by SIGKILL if the application has not exited within 15 seconds. This window is intended for saving final checkpoints, but the authors emphasize that important work should already be saved during normal operation. For FastAPI applications, it is recommended to set Uvicorn's --timeout-graceful-shutdown to below 15 seconds (e.g., 10 seconds) to prevent long-running requests from consuming checkpoint time.

In the test deployment phase, the authors recommend using tools like Locust to perform load tests with real user journeys at expected peak levels, gradually increasing concurrent users until throughput no longer increases or latency and error rates exceed targets. The article lists a decision table to help select instance specifications: if requests are slow but application CPU usage is low, it indicates waiting on the backend, so queries or connection pools should be fixed before adding compute; if CPU is near the limit and the server runs multiple workers, upgrading to Large is appropriate; for single-process applications (such as Streamlit or default-configured AppKit), when CPU is near the limit, adding instances rather than upgrading specifications is better, because larger instances are only useful if the application can utilize the additional cores.

The operations phase suggests keeping telemetry enabled, wrapping SQL, model, and API calls with custom spans to locate time consumption, and tracking individual user requests across instances via traces and request IDs. When slow requests occur, the article provides a symptom-to-cause comparison table: slow first requests after idle periods point to cold starts; CPU or memory near the limit points to the application itself being the bottleneck; database connection timeouts point to pool contention or long transactions.

The key premise of the entire method is that application compute is responsible only for serving the UI and coordinating calls, while heavy lifting is delegated to specialized services. Long-running work should be placed in background tasks, allowing the UI to poll for status or receive streaming updates, rather than keeping a request hanging indefinitely. If the application needs to run code written by users or generated by agents, it should be placed in the Databricks Sandbox isolated environment, ensuring its CPU and memory consumption do not affect the instances serving users.