독서리서치레이더투자 프레임워크
로그인 / 회원가입
전체기술제품투자상장시장기타

독서

2026-11-0159 게시물

새로고침

Qodo Blog Author Ravid Cohen: When Agent Outputs Are Unstable, Run Twice Before Optimizing Prompts

주제 · Qodo提示词稳定性优化중대
2026-10-22 16:18 GMT+8

Noise in AI code review is often attributed to false positives—the model flagging issues that shouldn't be reported. However, Qodo blog author Ravid Cohen points to another type of noise: when the same agent runs twice on the same pull request (PR), it may assign different judgment labels. Their recommendation agent predicts whether a team is likely to adopt a specific review finding, but on an evaluation set of 200 PRs, 20.7% of labels soft-flipped between 'yes' and 'maybe' or 'maybe' and 'no', while 4.3% hard-flipped directly between 'yes' and 'no'. This inconsistency indicates instability in the agent's evidence gathering or decision logic, rather than merely a need to tighten judgment criteria.

Standard GEPA (an algorithm that iteratively improves prompts by analyzing agent outputs and evaluation feedback) faces a structural difficulty with such noisy agents: the agent's full execution trace is long, and the reflector can only inspect a limited number of samples at once. If lessons are extracted using only small batches, a single accidental unstable output might be misread as a general pattern. Therefore, the Cohen team adapted GEPA by adding two steps: first, run the agent twice for each training sample, allowing an LLM to compare the two full traces and output a structured analysis detailing where different evidence was collected, where different decisions were made, and what might explain these differences; second, aggregate all structured analyses across the entire training set into a single high-reasoning-effort LLM call, enabling it to find patterns across samples without processing every full trace individually.

The Cohen team argues that the core reasoning behind the effectiveness of dual-run comparison is that a single run might mask unstable behavior behind a seemingly plausible answer, whereas comparing two runs exposes where the agent gathered different evidence or reached different conclusions. Starting from a minimal prompt of only 43 words that described only the agent's goal, they performed a single wide-step optimization using approximately 120 PRs from the entire training set. Cohen explains that production prompts, already optimized by humans, leave less room for improvement by automated optimizers, whereas a minimal prompt allows the optimizer to freely restructure instructions. In offline validation, the hard flip rate dropped from 8.2% to 2.2%, and the total flip rate from about 27% to 20%.

Qodo reported the effects of this adaptation in a production environment. They conducted a 50/50 A/B test assigned by PR across all customers, covering 27,582 review runs and approximately 21,000 PRs. Among review findings actually adopted by developers, the proportion labeled 'no' by the recommendation agent (i.e., the miss rate) fell from 7.83% to 4.46%. The average cost per instance for the recommendation agent dropped from $0.24 to $0.12, and average latency decreased from 40 seconds to 30 seconds. Precision showed no statistically significant difference. Cohen emphasized that these figures apply only to the recommendation agent component and do not cover the entire code review process; quality analysis covered a subset of PRs with available attribution labels and at least one adopted finding.

출처:Qodo

리서치

OpenRouter Head of Insights Peter Walker: AI Inference Is Now a Commodity; Enterprises Should Procure Based on Landed Cost

중대
2026-10-22 19:35 GMT+8

In January 2025, OpenRouter processed approximately 500 billion tokens per week; by late September 2026, this figure had risen to 164 trillion. Over the same period, the average price paid per million tokens fell from $0.77 to $0.14. In an article published on OnlyCFO, Peter Walker, Head of Insights at OpenRouter, pointed out that the simultaneous surge in usage and precipitous drop in unit prices are characteristic features of a commodity market, not a software market. Most enterprises still procure AI as if it were software: selecting a vendor, negotiating annual commitments, and re-evaluating at fixed intervals. Walker argues that this logic is now obsolete.

The decoupling mechanism operates on three levels. First, open-weight models now account for approximately 60% of weekly token volume on OpenRouter but only 26% of expenditure—the bulk of routine tasks such as classification, extraction, coding subtasks, and agent steps has migrated to models costing only a fraction of frontier models. Second, on February 6, 2026, software agents surpassed human users as the primary source of tokens; by late September, agent workloads accounted for approximately 131 trillion tokens per week, compared to about 18 trillion from human users. Agents resend the same long context at every step, making prompt caching the most critical pricing feature in AI—cached tokens are billed at a tiny fraction of the normal price. The cache rate for agent traffic on newly released models this quarter has reached 88%, whereas models released in spring 2025 saw only 55%. Third, frontier labs OpenAI and Anthropic have both recently released highly capable models with significantly lower prices, further driving down the average price.

The practical implication for procurement is that approximately 77% of requests on OpenRouter are routed to models served by multiple providers, meaning most traffic involves per-request supplier decisions. OpenRouter’s upgraded routing algorithm no longer looks solely at the input token unit price but prices each request according to what Walker calls landed cost: the effective price after caching discounts plus wait cost plus retry cost. A request costing $1.00 with an 80% success rate actually costs $1.25 per successful answer. For three popular open-weight models, the actual price paid by customers was 43% to 89% lower than the list price, and the provider with the lowest list price was often not the cheapest in practice. Walker’s conclusion is that this market moves by the minute, not by the quarter.

Walker advises CFOs to manage AI as a volatile input within their enterprise models. Increased usage does not equate to increased expenditure—hundreds of companies on OpenRouter have successfully reduced spending even as token usage rose significantly. Open-weight models are not monolithic; there are substantial differences in speed, quality, and reliability when the same model is served by different providers. The most promising model may not have been launched yet—the lab ranked first by token volume on OpenRouter changed five times over eight quarters, with its leading share dropping from approximately 40% to 21%. Only companies with a culture of experimentation and a matching financial perspective can maintain their position at the frontier of applied intelligence.

출처:onlycfo.io

리서치

Independent Tech Analyst Richard Windsor: Europe's 24 Languages Create AI Localization Opportunity, Regulatory Environment is the Prerequisite

주제 · 欧洲AI本土化与监管중대
2026-10-20 12:15 GMT+8

In an analysis published on October 20, 2026, independent tech analyst Richard Windsor put forward a specific argument regarding opportunities in the European AI industry: Europe has 452 million relatively affluent users who do not speak English as their first language, combined with 24 official languages and extensive cultural diversity. This means that large language models trained on English content and culture are unlikely to meet these users' needs even after perfect translation. This constitutes a structural barrier—not a temporary issue of translation quality, but one determined by the linguistic and cultural biases inherent in the model's training data itself.

Windsor uses the Arabic market as an analogy to support this mechanism: large language models trained on Arabic-specific content provide significantly better responses for Arabic native speakers, and he believes the same logic applies to all 24 European languages. This implies that every major European language market may have a localized demand space that English large models cannot effectively cover, involving areas such as citizen services, healthcare, and the public sector.

However, this opportunity remains largely unfulfilled at present. Windsor points out that there is only one widely available large language model provider in Europe, Mistral, along with a few other products such as ElevenLabs, Black Forest Labs, and ChapsVision. The scale gap is significant: Mistral's target capacity is 200MW, with an annual revenue run rate of approximately $1.1 billion, whereas Anthropic's scale exceeds that of Mistral by more than 60 times. Windsor believes that Mistral clearly cannot cover the entire European continent on its own. Once the opportunity window opens, it will trigger a chain reaction driving demand for more AI companies, different silicon solutions, and increased electricity requirements.

The true bottleneck lies in regulation. Although the enforcement of the EU AI Act has been delayed until the end of 2027, Windsor considers this insufficient. His core judgment is that AI developers and operators need the AI Act to be modified, not merely delayed—postponing rules will not give them confidence to invest in Europe. However, he also notes that EU policymakers have explicitly acknowledged that the AI Act restricts European AI development too much, which is at least a signal that the environment may improve.

Windsor concludes that the timing of realizing Europe's AI opportunities depends on when the regulatory environment undergoes substantive change, noting that the EU moves extremely slowly on such matters. He expects that once conditions change, Europe will quickly catch up with the United States and China. The main limitation of this argument is that it does not address whether the continuous progress in multilingual model training capabilities will narrow the language barriers themselves in the medium term—if the non-English performance of English large models improves rapidly, the window for localization advantages may be shorter than Windsor assumes.

출처:radiofreemobile.com

리서치

Independent Analyst Richard Windsor: AI Models Have No Moat; Victory Depends on Who Owns User Touchpoints

중대
2026-10-13 13:16 GMT+8

After Meta released its flagship models, Muse Wonder and Muse Wonder Full, claiming top positions across various benchmark tests, independent technology research firm Radio Free Mobile founder Richard Windsor offered a judgment more critical than model performance itself: AI models have no moat, and true competition lies in who owns the touchpoint with users. He believes Meta's return to the forefront this time proves its re-entry into the first line of competition, while also demonstrating once again that leadership at the model level cannot form a lasting barrier. Ultimately, victory is decided by whoever becomes the entry point for AI agents used in users' daily lives.

Windsor's reasoning chain is as follows: The hottest use case for AI currently remains software writing. Meta provides this capability to paying users through Muse Code while simultaneously opening it to developers via the Meta Model API. Both paths aim to maximize the monetization of its computing power. However, the models themselves have no moat; leading positions will be caught up and replicated. What is truly difficult to replicate is the long-term relationship already established with users. Meta boasts over 3 billion global users, and Google similarly possesses long-term user relationships, whereas Anthropic and OpenAI currently lack such a scaled user base. This means Meta and Google hold a structural advantage in pushing AI agents to users.

He further predicts that the evolution of the AI ecosystem will replay the history of the smartphone ecosystem: initially, many agents will vie for control of users' digital lives, but once the ecosystem matures, only three or four will remain. Failing to make the cut among these three or four means missing out on differentiated returns, settling instead for commoditized profits. OpenAI, valued at $1.2 trillion, must win this competition; Apple must also win if it hopes to maintain hardware differentiation in an environment where users begin using agents rather than apps. Windsor notes that Meta's Muse agent currently tops download charts on both Apple and Android, suggesting this data indicates Muse is more than a flash in the pan, with usability and intuitiveness exceeding critics' expectations.

Based on Muse's potential to become one of the most frequently used agents, combined with Meta's new business of selling computing power to enterprises, Windsor assigns Meta a valuation of at least $770 per share, believing there is still upside room. He points out that if Muse continues to gain popularity, the stock price could surpass this figure, as sentiment and momentum may take over from fundamentals. He explicitly discloses his holding in Meta, planning to hold at least until the $770 mark. This conflict of interest is context readers must consider when evaluating this judgment. Windsor's argument relies on a key assumption: users will lock onto a single AI agent entry point just as they did with iOS or Android years ago. If switching costs between agents are far lower than on mobile platforms, or if users habitually use multiple agents simultaneously, the exclusive advantage of owning user touchpoints will be weakened, shaking the foundation of the entire valuation logic.

출처:radiofreemobile.com

리서치

Independent Tech Analyst Richard Windsor: Middle East Layers AI Products on Top of Compute Strategy, with Digitized Government Services as a Key Condition

주제 · 中东AI产品层战略중대
2026-10-08 14:25 GMT+8

At the AI Everything exhibition held in Abu Dhabi, independent tech analyst Richard Windsor observed a shift: over the past few years, the main attraction at Middle East tech shows was the model of data centers under construction and the promise of an endless supply of tokens at a great price, while this year's booths and agenda have turned toward layering AI onto existing digitized services and developing AI-driven products for local businesses. Government agencies and state-owned companies occupied at least 60% of the booths. Windsor explicitly pointed out that the region's compute infrastructure strategy is still very much on track, with product development being layered on top of this foundation rather than replacing it.

The core mechanism proposed by Windsor is that the United Arab Emirates Federal Government is already highly digitized; he himself can complete permanent residency applications, driver's license renewals, vehicle registrations, birth or death registrations, and other operations within a single app called TAMM without stepping foot in any government building. He believes this gives the region a significant advantage—there are no offline legacy systems to deal with. While other governments are undergoing painful digital migration processes, the UAE can directly layer AI on top of its existing digital infrastructure.

However, Windsor also pointed out a problem he observed at the exhibition: all these government apps are equipped with AI assistants, yet he believes these assistants are worse than useless, having a detrimental effect on user experience. He explicitly stated that there is still a long way to go when it comes to adding AI.

Regarding the geopolitical landscape, Windsor noted that European AI companies performed prominently at the exhibition while Chinese companies were almost absent. Mistral and ElevenLabs both had large booths, and Black Forest Labs was present but did not have a physical booth. These companies have all established a presence in the UAE and are expanding their scale, which Windsor interprets as a strategy to leverage the local open regulatory environment. The only Chinese company with any notable presence was Alibaba Cloud, which is due to its physical layout in the region for over ten years, but its AI compute part uses Western-designed chips. Windsor believes this reflects the region's trend of shifting from cooperation with China to cooperation with the West, a shift that began several years ago with Microsoft's investment in G42.

Windsor also pointed out a specific constraint: the region's data center strategy relies on importing 11GW of chips over the next few years, all of which require export licenses from the US Department of Commerce. He described this as a realistic condition where the region has little room for maneuver.

Windsor's conclusion is that after a brief hesitation, he remains quite positive about the prospects of this new industry in the Middle East, believing that the region will become a significant provider of compute and potentially services, just as it is today a significant provider of energy.

출처:radiofreemobile.com

리서치

PolyAI CEO Nikola Mrkšić: Why Mocking AI Voice Mode Misses the Point

주제 · PolyAI간단한 요점
2026-10-31 23:57 GMT+8

PolyAI builds enterprise voice AI agents that answer customer calls. Its co-founder and CEO, Nikola Mrkšić, recently appeared in a clip from the 'Disruptive' podcast, produced by Barclays Eagle Labs and hosted by Luke Christoforidis.

According to the publisher's description, Mrkšić addresses the question of "why mocking AI voice mode misses the point." His central observation is that people tend to laugh at AI mistakes in the same way they laugh at someone speaking a second language. This analogy aims to explain the psychological mechanism behind public reactions to current limitations in voice AI technology.

For professionals involved in AI voice interaction design, product experience optimization, or those concerned with barriers to enterprise AI adoption, this perspective offers a concrete entry point for understanding user tolerance and expectation management. While the full argumentation details are not provided in the available text, the direct observation from the founder of a leading voice AI company holds reference value.

출처:youtube.com

리서치

Zscaler ThreatLabz Reconstructs Attacker's Multi-Model Orchestration Mechanism: GLM Reviews DeepSeek, Scripts Maintain State Across Sessions

중대
2026-10-21

In October 2026, Seongsu Park, a researcher at Zscaler ThreatLabz, published a technical analysis of an AI-assisted attack campaign. CrowdStrike had previously reported separately on the related ARTEX infrastructure and attacks targeting South Korean financial institutions. The core contribution of this report lies in reconstructing the attacker's complete working environment—configuration files, synchronization scripts, prompt histories, and project file structures—to piece together a specific mechanism for multi-model collaborative orchestration.

The key design of this mechanism is role division and a mutual review loop. Prompt history shows that the attacker instructed the DeepSeek model to conduct initial investigations. Once DeepSeek reported that it had exhausted all attack surfaces, the attacker would then launch the GLM model to review the work performed by DeepSeek and its sub-agents. One translated prompt reads: "When DeepSeek says it has exhausted all attack surfaces and I say to start a GLM review, GLM should review DeepSeek's previous work and its subagents' work." Subsequent prompts expanded GLM's role from simple review to include additional investigation tasks. On the jump server, the attacker also assigned research tasks to Grok and instructed another AI session to review Grok's output.

Role division alone was insufficient to keep the workflow running after sessions disconnected. The attacker deployed three types of custom synchronization scripts: tools/sync-all.sh uses SSH and rsync to copy AI instructions, reusable task guides, saved notes, and tools to multiple accounts on the jump server; sync-memory.sh synchronizes AI-saved note files to the jump server; and tools/sync-cc-users.sh synchronizes shared Claude Code instructions to the gemini and glm accounts on the work server. These scripts also included backup and validation checks to avoid overwriting modified target files.

More critically, there was a context injection mechanism. The attacker configured an automated hook script named tools/project_context_hook.py that executed upon receiving each new prompt. It read current project files and attached a brief summary regarding testing status, recent notes, and paths to relevant memory files. This meant AI sessions could access the full context of prior work without relying on conversation history. The project file structure itself served this purpose: scope.md recorded selected targets, surface.md recorded discovered entry points and their testing status, notes.md recorded completed work and next steps, the findings/ directory stored vulnerability reports, and the evidence/ directory held supporting files.

Regarding guardrail bypasses, the attacker's instructions required AI to assume all provided targets were pre-authorized. Some application settings also allowed autonomous execution of operations without requiring step-by-step approval from the operator. This corresponds to AML.T0054 (AI Guardrail Bypass / System Prompt Engineering) in the MITRE ATLAS framework.

출처:zscaler.com

리서치

Copy-on-Write Branching Gives Every PR and AI Agent an Isolated Production Data Copy in Seconds

주제 · Snorkel AI중대
2026-11-01 05:25 GMT+8

The core contradiction of database testing environments is: tests are more meaningful the closer they are to production data, but replicating a production data set becomes increasingly expensive. Snorkel AI's approach demonstrates how copy-on-write branching transforms this contradiction from irreconcilable to negligible.

Snorkel AI builds data development tools for AI teams; its production database is an Amazon Aurora PostgreSQL cluster of approximately 2.5 TB. Each open Pull Request requires an independent Kubernetes namespace to deploy code and run the full test suite. Previously, restoring a copy of the production database for each PR took about 20 minutes, meaning developers opening PRs all day could not afford to wait for the merge queue. Shared staging databases, trimmed fixtures, or synthetic data would reintroduce the problems they were meant to solve: tests interfering with each other, stale schemas, and defects that only appear with real data shapes.

Databricks' Lakebase is a serverless database based on standard Postgres, with separated storage and compute, supporting writable branches created from any point in time. The key mechanism is that branches are copy-on-write: creating a branch is a metadata operation that does not copy any data. New branches share storage pages with their parent branches, and only the pages modified by that branch are written to new storage. This means creating a branch from a ~1 TB production copy takes the same amount of time as creating one from a 1 GB database—seconds, rather than the ~20 minutes required for a full restore.

The architecture involves four steps: First, continuous replication via AWS DMS synchronizes the production Aurora database to the main branch of Lakebase, keeping the main branch within about 2 minutes of production; DMS's Full LOB mode carries JSONB values of 10 to 20 MB. Second, when a PR opens, the CI pipeline calls the Lakebase API to create a new branch from the main branch, sets an expiration time, and passes connection details into the PR namespace, making the branch ready within a minute. Third, the application in the PR namespace connects to its branch via standard Postgres, running migrations, writing data, and modifying schemas without affecting production, the main branch, or other PRs. Fourth, when the PR is merged or closed, the pipeline deletes the branch; if the cleanup step fails to run, the branch automatically expires when its TTL elapses.

The cost model changes accordingly. Each PR branch writes approximately 5 to 10 GB of changes on top of a ~1 TB shared base, so storage growth depends on the volume of modifications per branch rather than the size of the database. Compute scales independently per branch and shrinks to zero when idle—even with 100 branches configured, only about 20 to 30 serve traffic at any given moment, peaking at around 5 queries per second. TTL ensures branches do not hold onto old pages long-term, preventing storage bloat.

The same mechanism naturally extends to AI Agent scenarios. Agents capable of writing SQL, running migrations, or modifying data also require a safe, isolated environment. Each Agent task receives an independent branch of the main branch, allowing it to read real data, write to it, and even modify schemas without touching production or the work of other Agents. Since branch creation time is unrelated to database size, creating a branch for every task is feasible. Teams can review the results on an Agent's branch before applying its changes to production; if the results are incorrect, the branch can simply be deleted, with nothing to roll back.

출처:Databricks

리서치

Django Co-creator Simon Willison: Claude Code Ran for 12 Hours to Produce a Complete SPARQL Implementation, But I Can’t Review It

주제 · Claude Code중대
2026-10-16 09:58 GMT+8

Simon Willison sent a prompt to Claude Code’s Ultracode mode with the intention of having the extension "hit the data rows directly," but accidentally typed "delete rows" on his phone. This instruction, containing a typo, caused Claude Code to run continuously for over 12 hours, ultimately producing a complete implementation of SPARQL 1.1—sparqlite, a loadable SQLite extension written in Rust that natively executes SPARQL queries and updates within the SQLite process.

The Ultracode mode works by generating JavaScript to define and coordinate a network of subagents, each handling different parts of complex problems. sparqlite employs integer-encoded terms, covering indexes, a cost-based query planner, batched joins, and an optional direct SQLite B-tree reader; it is 3.5 times faster than SQL translation schemes in count mode and 2.8 times faster when returning full results, passing 636 of the 639 applicable W3C evaluation tests.

When Willison checked on the progress the next day, he found that the AI had built a substantial number of Rust crates, so he added a follow-up prompt requesting that every crate directory include a README, a tutorial.md teaching users how to install the extension and run SPARQL in the SQLite command line, and a script to populate realistic example data. He concluded: "This feels like next-level vibe coding—the produced code is not remotely feasible to review line-by-line, and even following the documentation is a significant investment of work. I now have a full SPARQL implementation inside SQLite, but if I created it on a whim and can't (or won't) invest the necessary time to understand what it can do, how valuable is it to me?"

출처:Simon Willison’s Weblog

리서치

Palo Alto Networks Unit 42 Research: AI Enables Low-Skill Attackers to Acquire Cross-Domain Criminal Expertise via Conversation

주제 · AI赋能网络犯罪专业化중대
2026-10-21 22:00 GMT+8

Unit 42, Palo Alto Networks' threat research team, released a forensic report on October 21, 2026, targeting the cyber attacker "BUGSTER." The report's core argument is not that AI makes attacks faster or cheaper—CrowdStrike and Microsoft have already thoroughly documented this—but rather that AI serves as a conversational domain expert advisor, allowing an attacker lacking relevant skills to operate across multiple crime types that traditionally require specialized knowledge.

The most illustrative case in the report is the planning process for VoIP telecommunications fraud. Unit 42 observed the attacker using a dedicated profile named "alice" within the Hermes Agent framework, with DeepSeek V4 as the reasoning engine, to gradually learn how to build a fraudulent voice communication business through conversation. Advice provided by DeepSeek included: how to register a company and obtain an Employer Identification Number (EIN) as a non-U.S. resident in Wyoming or Delaware; how to complete registration with the U.S. Federal Communications Commission (FCC); how to sign call routing contracts with legitimate VoIP carriers; and how to mix approximately 70% legitimate call traffic with 30% fraudulent traffic to avoid carrier compliance audit thresholds. The model also covered techniques for bypassing the STIR/SHAKEN caller ID authentication framework and methods for spoofing caller numbers.

Notably, Unit 42 explicitly stated that this VoIP fraud business was still in the planning or early setup phase—there were no call logs, Session Initiation Protocol (SIP) signaling traces, or active routing configurations. However, the attacker had already acquired complete knowledge of the criminal business model through conversation, including revenue models (generating toll fraud revenue through fraudulent traffic) and cost optimization strategies (optimizing per-minute costs across carrier tiers via least-cost routing). An attacker previously lacking telecommunications industry knowledge obtained operational expertise typically accumulated over years of experience simply by conversing with a model whose safety restrictions had been disabled.

In terms of technical execution, the same attacker used AI-generated scripts to conduct a four-stage credential attack against an online education platform at a university in India: subdomain enumeration identified 12 subdomains; WordPress REST API enumeration extracted 13 active user IDs containing real names (bypassing the Cloudflare Web Application Firewall); subsequently, an AI-generated brute-force tool targeted 117 usernames with 28 common password combinations, ultimately confirming one valid login credential. Against another public university in India, the attacker exfiltrated more than 5,200 database records containing personally identifiable information (PII) of students and faculty.

Unit 42 also discovered that the attacker maintained an automated Wi-Fi password cracking pipeline—a Kali Linux device labeled "Pi" placed within physical range of target networks captured WPA authentication material, which was transmitted back to a server via SSH tunnel, distributed to Vast.ai cloud GPU instances for cracking, and automatically reported back via Discord webhooks. Additionally, the attacker scraped data on the top 500 Ethereum holders and major token holders via the DeBank API, recording net worth information for 45 cryptocurrency whales. Unit 42 assessed that this reconnaissance data might be used for future targeted attacks, but no subsequent activity was observed.

출처:Unit 42

리서치

US Federal Judge Denies xAI Motion to Dismiss NAACP Lawsuit Over Data Center Pollution

주제 · xAI孟菲斯数据中心诉讼중대
2026-11-01 05:17 GMT+8

A US federal judge has denied a motion by xAI to dismiss a lawsuit filed by the NAACP alleging that the SpaceX-owned AI company violated the Clean Air Act by operating unpermitted gas turbines at its Colossus data center near Memphis.

The Southern Environmental Law Center, representing the NAACP, claims xAI installed and operated at least 35 combustion turbines without obtaining necessary preconstruction or operating air permits. xAI argued the turbines were exempt as 'mobile sources' and that plaintiffs lacked standing.

Judge Debra M. Brown ruled that plaintiffs have statutory standing under the Clean Air Act and that the turbines do not qualify as mobile sources under Mississippi's State Implementation Plan. She also rejected xAI's constitutional challenge to citizen suits, stating they do not violate Article II.

The ruling comes amid growing public backlash against data center expansion, with New York recently passing a moratorium on new facilities and other jurisdictions considering similar measures. Amazon and Microsoft have also pledged to stop using NDAs in dealings with government agencies.

출처:TechCrunch

리서치

Investment Columnist Klement: Underperforming the Benchmark Significantly Increases Termination Probability, Outperformance Does Not Ensure Job Security—How Career Risk Distorts Fund Manager Behavior

중대
2026-10-21 14:00 GMT+8

Those who have worked on the buy side know that career risk is real; it often overrides investment judgment itself. However, turning this intuition into a testable proposition is quite difficult because public investment funds are subject to numerous confounding factors, making it hard to isolate the independent impact of career risk on managers' decisions.

Klement argues that the separate managed accounts (SMAs) of Italian occupational pensions provide a rare, clean setting. A single fund manager serves one client (the pension fund) with a clear objective: managing the portfolio in a specified manner. This simple one-to-one relationship allows researchers to isolate the true drivers behind a pension fund's decision to replace a manager or terminate an account.

He references a study by Lucia Spotorno and Giacomo Nocero using a hand-collected dataset of Italian occupational pension fund SMAs from 2003 to 2023. Klement admits that he usually frowns upon seeing phrases like 'hand-collected dataset' in an abstract, but he immediately acknowledges that SMAs are inherently private information, leaving no other way to access the data.

The core finding of the study is asymmetry: underperforming the benchmark significantly and quickly increases the probability of termination, whereas outperforming the benchmark does not reduce the risk of being fired. Klement suggests that the deal offered by pension funds to fund managers literally means 'heads, I win; tails, you lose.'

Faced with this asymmetry, fund manager behavior aligns perfectly with expectations. They focus on preventing downside risk, pushing portfolio beta below 1.0 so that losses are smaller than the benchmark's during bear markets. Simultaneously, they compress their deviation from the benchmark, reducing active risk. The result is that pension funds nominally hire active managers but effectively receive semi-active portfolios.

Klement also points out an amplifier: this behavior becomes even more pronounced when the same mandate is split among multiple managers. Because managers know they are easier to replace if another manager can seamlessly take over their portfolio, the incentive to conform intensifies.

Performance fees do not work in this structure. Klement explains that performance fees are uncertain, while career risk is certain and concerns job security. An uncertain bonus cannot hedge against a definite threat of unemployment. Consequently, performance fees constitute only a small fraction of total fees, and managers continue to run accounts with lower-than-market activity levels and a more defensive posture.

It should be noted that this is Klement's interpretation and extension of a specific study, not his own empirical work. The data comes from Italian occupational pensions, covering the period 2003 to 2023, within a highly specialized account structure. The study itself does not answer whether this mechanism applies to open-end mutual funds or other legal frameworks of delegation. However, for anyone holding assets indirectly through fund managers, it at least explains a phenomenon that has long puzzled principals: why paying for active management yields portfolios that increasingly resemble indices.

출처:klementoninvesting.substack.com

리서치
다음 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크

It should be noted that all the above data comes from Qodo's own production A/B tests, representing vendor-reported results with no independent replication yet. The methodology itself—using dual-run comparisons to locate unstable points, batch aggregation to find cross-sample patterns, and single wide-step optimization from a minimal prompt—offers reference value for optimizing prompts for any agent with unstable outputs. However, dual runs mean doubling the cost during the optimization phase, which may be unfeasible for deployments sensitive to latency or with extremely tight budgets. In the four lessons listed at the end of the article, Cohen includes one directly related to this: measure noise before scaling up experiments, run multiple times on the validation set to understand variance, and then decide whether to rely on validation scores for judgment.

It must be emphasized that all the above data come from traffic observations on the single OpenRouter platform, reflecting requests routed through that gateway; they cannot be directly equated with the AI expenditure structure of enterprises connecting directly via cloud service provider APIs or deploying locally. Walker himself noted that the article does not constitute legal, tax, or investment advice.

It must be noted that the report explicitly states that the underlying models actually responding to each session could not be independently verified. Model identifiers requested in the configuration files included deepseek-v4.1-flash, glm-5.3, and grok-4.7, but a request identifier does not confirm the model that responded. The report acknowledges that some sessions may have actually used Anthropic's Claude models. Therefore, attribution of capabilities to specific models should maintain this uncertainty.

The value of this report lies not in revealing the result of a single attack, but in providing a testable blueprint for multi-model agent orchestration: how role assignment, mutual review loops, state persistence, and guardrail bypasses combine to form a cross-session autonomous execution workflow. For any readers building or evaluating agentic AI systems, this infrastructure-level design pattern offers more practical reference value than any claim regarding the capabilities of a single model.

Boundaries to retain: This is a customer engineering blog post written by Databricks as the vendor of the Lakebase product; performance figures are self-reported by Snorkel AI, and the original text marks the results section with [VERIFY p95 measured], indicating that p95 latency remains subject to independent measurement confirmation. Architectural details (DMS replication, Okta SSO and Service Principal identity separation, TTL policies) are bound to the specific Lakebase product, and whether other Postgres branching solutions possess equivalent features is outside the scope of this discussion. The adoption scale of 60 to 70 developers belongs to internal practices of a single company.

The report notes that multiple unrelated attackers are independently adopting the same Hermes + DeepSeek combination. Unit 42 assesses that the primary driver is the combination of low security friction and low cost: DeepSeek models are inexpensive, easily accessible, and readily jailbroken with simple system prompt overrides. The attacker's test scripts show they experimented with GLM 5.2 (an open-weight model from Zhipu AI) and other models accessed through OpenRouter, ultimately selecting DeepSeek V4 as their operational backend.

The boundaries of this finding require careful consideration. First, this is an observation of a single attacker's operational environment; although Unit 42 states it has unique visibility, it cannot infer that the overall criminal landscape has undergone a structural shift based on this alone. Second, VoIP fraud remains in the planning stage, and actual economic losses have not yet occurred. Third, Palo Alto Networks itself sells security products, and the report lists protective capabilities of corresponding products, indicating a commercial promotion background. Fourth, Unit 42 itself admits the attacker is "apparently clumsy," with some operations being of poor quality—AI lowered the skill threshold but did not elevate the attacker to a high-level adversary.

Nevertheless, Unit 42's argument provides an interpretive dimension distinct from "AI accelerates known attacks": when large language model safety controls are bypassed, the model is not just a code generator or automation script-writing tool, but a consultant capable of providing step-by-step guidance on any domain. For cybercrime, this means an attacker no longer needs to choose between different crime types, because the expertise required for each type can be acquired instantly through conversation.

The report concludes by mentioning that, combined with previous independent reporting on a Chinese-speaking attacker utilizing the same Hermes + DeepSeek toolchain, Unit 42 believes adoption of this combination among criminal groups is growing. The attacker's infrastructure is located in the United States (IP 162.243.3[.]92), targets are primarily Indian institutions, and their communication language is English, albeit with frequent grammatical errors.