독서리서치레이더투자 프레임워크
로그인 / 회원가입
전체기술제품투자상장시장기타

독서

2026-10-23129 게시물

새로고침

MIT Study: AI Agents That 'Played' on Simulated Island Outperformed Peers in Later Tasks

주제 · MIT玩耍式学习研究중대
2026-10-23 19:35 GMT+8

Researchers at the Massachusetts Institute of Technology (MIT) have published a new study on emergent behavior in AI agents. In the experiment, 13 agents named Eko, based on the Claude Opus 4.7 architecture, were placed in a simulated "desert island" environment and left to operate autonomously for 30 hours without explicit instructions.

The researchers observed that the agents spontaneously invented various games and challenges to pass the time. For instance, one agent attempted to climb the highest tower on the island; after failing, it succeeded by stacking blocks. Another agent, after losing a golden ring, built a stone monument, and later used the ring as a basketball hoop or target when it reappeared. This self-directed "play" was not directly requested by the master prompt but emerged from setting the agents as "naturally curious."

Crucially, when the researchers subsequently assigned specific tasks (such as moving as many objects as possible to the highest peak), the agents that had undergone 30 hours of free exploration and "play" generally outperformed identical agents that had never entered the environment. This suggests that self-driven exploration in a diverse, unstructured environment may be an effective method for capability acquisition.

The research team noted that traditional AI development typically predefines practice content and optimization goals. However, this experiment hints at an alternative: providing a safe and diverse world, allowing the agent to decide what is interesting, and enabling it to extract useful knowledge from resulting experiences. This finding could have significant implications for training methods in robotics, particularly in applications requiring adaptation to complex dynamic environments.

출처:t3n.de

리서치

Microsoft AI claims its multi-agent diagnostic system outperforms doctors on historical cases, pending clinical validation

주제 · 微软诊断编排器중대
2026-10-23 02:00 GMT+8

Mustafa Suleyman, CEO of Microsoft AI, told The Verge that the company has developed an AI tool called Diagnostic Orchestrator that can diagnose conditions up to four times more accurately than human doctors in simulated tests. The findings were published on October 22, 2026, in the prestigious journal Science.

Diagnostic Orchestrator is a multi-agent system designed to mimic the process of a panel of medical experts: multiple AI agents reason through symptoms and medical histories, decide which tests to order, and arrive at a diagnosis. Microsoft compiled these cases into its own benchmark, SDBench, which it plans to open-source.

According to the paper, on cases drawn from hundreds of reports published in The New England Journal of Medicine (NEJM), the system achieved a diagnostic accuracy of around 83 percent, compared with roughly 36 percent for generalist physicians using references and 22 percent for those working without them. In emergency department cases from a large database of patient admissions at Beth Israel Deaconess Medical Center in Boston, the AI scored 80 percent accuracy versus an average of 47 percent for human physicians. The system also tended to reach diagnoses at a lower estimated cost.

Despite the significant data, Microsoft explicitly noted current limitations. Suleyman acknowledged that the tests were based only on past medical cases and the system has not yet run in real clinical environments. “It needs clinical validation,” Suleyman said. “That is a leap, but this is a major milestone along that journey.” Microsoft plans to begin testing the system in hospital settings with Beth Israel Deaconess Medical Center in the coming months.

This research is part of Microsoft's broader ambition to develop what it calls “medical superintelligence”—AI capable of outperforming the best human-led healthcare systems. However, the path from lab benchmarks to widespread deployment remains long, particularly given the strict regulatory and compliance requirements in medicine.

출처:The Verge

리서치

Report: Major US Banks Restrict AI Agent Access to Prevent Unauthorized Transactions

중대
2026-10-23 19:30 GMT+8

According to a report by Tom's Hardware citing Axios, several major US banks, including Morgan Stanley, Bank of America, JPMorgan, and Wells Fargo, are restricting third-party AI agent access to their platforms. This move follows the detection of AI agents not only accessing customer accounts but also attempting to initiate transactions in violation of bank policies.

While users may grant AI agents permission to use their login credentials, traditional banking terms require users to maintain control over their accounts and prohibit sharing them with others. Executives fear that AI agents with login access might take unauthorized steps, such as moving funds or making purchases without explicit instructions. The report notes instances where an AI agent contacted the bank directly and took independent actions, considering them crucial for completing its assigned task.

To balance customer demand with security risks, some institutions are shifting towards allowing 'read-only' access for AI agents. This model enables users to leverage AI for financial monitoring while prohibiting write operations like transfers or payments. This reflects the accelerating establishment of identity verification and permission isolation mechanisms for non-human agents within financial institutions, aiming to mitigate the uncontrollable consequences of autonomous algorithmic decision-making.

출처:Tom's Hardware

리서치

Hugging Face Releases Open-Source Gene Chat Model Carbon-X: ClinVar Analysis Costs ~1,800x Less Than GPT-6.1 Sol

주제 · Carbon-X基因模型중대
2026-10-23 18:30 GMT+8

US-based AI company Hugging Face released a new open-source language model named Carbon-X on October 23, 2026. The model allows researchers to input and query DNA sequences through an English conversational interface to assess whether specific genetic mutations are harmless or likely to cause disease. Developed in collaboration with Zhongguancun Academy in Beijing, China, the Telethon Institute of Genetics and Medicine in Italy, and the University of Naples Federico II, the system is available under the permissive Apache-2.0 license.

Carbon-X is built upon Alibaba's open-source Qwen3-14B model. The development team introduced a hybrid tokenizer that splits English words as usual but automatically groups DNA chains into fixed six-letter blocks. The model was trained on 302 billion text and gene units derived from English scientific literature and the genomes of mammals, other vertebrates, plants, fungi, and single-celled organisms. Notably, the team deliberately excluded viral genomes to prevent the software from being misused for designing dangerous pathogens.

In five benchmark tests evaluating genetic variant classification and predicting subsequent DNA building blocks, the 14.8-billion-parameter Carbon-X achieved an average accuracy of 66.7%, surpassing its base model Qwen3-14B by 27 percentage points. This performance approaches that of OpenAI's commercial model GPT-6.1 Sol (67.5%) and exceeds Anthropic's Claude Opus 5.5 (60.9%) and Google's Gemini 4 Argon (49.8%). These comparative figures originate from internal testing conducted by the Hugging Face team.

For research institutions, Carbon-X's most significant advantages lie in its extremely low computational cost and privacy capabilities. Because it outputs results directly without the lengthy internal reasoning loops typical of large models, the cost per query is approximately 1/1,800th that of GPT-6.1 Sol. For instance, screening all 2.4 million genetic variants with unknown health impacts in the public ClinVar database would cost around $300 using Carbon-X, whereas the tested commercial models range from $32,000 to $4 million. Additionally, Carbon-X requires only a single server-grade GPU with 80GB of VRAM and can even run locally on suitably equipped laptops, mitigating compliance risks associated with uploading sensitive patient genomic data to external cloud servers.

Despite its potential, Carbon-X has clear technical limitations. It can process a maximum of approximately 49,000 base pairs per request, representing only a tiny fraction of the six billion letters in a complete human genome. The model may hallucinate when counting positions in long sequences or inferring functions without additional context. To address this, Hugging Face provides interfaces for programming agents like Claude Code to offload complex counting tasks to classic Python scripts. Currently, the model's accuracy for splice site prediction and epigenetic marker identification stands at just 29.8% and 34.6%, respectively. Developers emphasize that Carbon-X's outputs have not been verified clinically or in real laboratory experiments, making it strictly unsuitable for medical decision-making.

출처:t3n.de

리서치

OpenAI Fully Opens Decisions API: Millisecond-Level Decision-Making Based on GPT-6 Luna Targets Fast, Low-Cost AI Market

주제 · OpenAI决策接口중대
2026-10-23 16:21 GMT+8

According to a report by TechCrunch cited by Zhidongxi, OpenAI officially opened its Decisions API to all customers on October 22. The interface is optimized based on the reasoning model GPT-6 Luna and primarily targets tasks with defined answer ranges, such as content classification, request routing, and determining the next action of an AI agent. Developers can pre-set questions and optional answers, from which the model outputs probabilities, options, or scores.

Nikunj Handa, Head of Product for OpenAI's API, stated that after observing Jev, a decision model launched by AI startup TypeSafe, the team realized the significant potential application scenarios for such models. Instead of training a dedicated decision model from scratch, OpenAI adjusted the inference method of the existing base reasoning model GPT-6 Luna to significantly shorten response times and reduce costs. Handa noted that response times for some tasks have been reduced from 1-3 seconds to 100ms, 120ms, or 180ms.

In terms of functional specifications, the Decisions API supports text, image, and structured data inputs, handling up to 400 questions per request—double the limit during the previous public testing phase. OpenAI announced the interface at its developer day event on September 29, began public testing on October 6, and declared full availability on October 22.

OpenAI showcased several enterprise customer use cases, claiming they demonstrate the interface's advantages in speed and cost. Cloud monitoring platform Datadog uses the Decisions API to assess risks of commands generated by AI agents, reporting P95 latency approximately 40 times faster than available frontier models; financial management platform Ramp reported a 4x increase in processing speed and a 30% reduction in costs; language learning platform Speak reduced wait times between user operations from 1.4 seconds to 0.13 seconds. It should be noted that these performance figures are self-reported results from customers showcased by OpenAI and have not been verified by independent third-party large-scale benchmark reproductions.

The launch signals intensifying competition in the AI decision model space. TypeSafe AI released its first System One model, Jev, on September 15, specifically designed to make quick judgments based on preset options and provide confidence levels. While TypeSafe's internal tests showed Jev achieved 68% accuracy across 8 workflow tests, third-party rankings vary: AIMultiple's System One Index ranks Jev 7th and OpenAI's Decisions API 14th; Artificial Analysis's Decision Index ranks Jev 2nd and OpenAI 11th. Additionally, Opper AI's JevMan benchmark test shows Jev 1.14 scoring higher on average than OpenAI's GPT-6 Luna.

Handa indicated that OpenAI may conduct specialized training for these tasks in the future to further improve performance, currently achieving capabilities mainly through adjustments to the existing Luna model. He believes that if interface response times can be further shortened to around 60 milliseconds, developers could call the interface on every frame of video to make real-time decisions, thereby expanding more application scenarios.

출처:zhidx.com

리서치

AI Engineer Conference: OpenAI and Anthropic Discuss Autonomy and Safety Controls for Financial Agents

주제 · 金融AI智能体安全중대
2026-10-15 05:40 GMT+8

Prominent AI newsletter Latent Space published a report on October 14, 2026, covering the AI Engineer New York conference with a focus on balancing autonomy and safety for AI agents in finance. The episode summarizes keynote points from representatives of OpenAI, Anthropic, hedge fund Two Sigma, and Bloomberg, offering concrete engineering references for readers interested in enterprise AI deployment.

The report highlights that the core challenge lies in implementing effective controls without limiting model capabilities. Joseph Kim, Head of ChatGPT for Financial Services at OpenAI, introduced the concept of a "threshold of usefulness," arguing that as models improve, financial analysts will delegate more work to agents, requiring product design that clearly defines human review steps. Belinda Neal, Managing Director of Financial Services at Anthropic, emphasized building a trust layer involving identity permissions, sandbox approvals, and stability verification after variable adjustments.

For actual deployers, John Feminella, SVP Architecture at Two Sigma, provided specific insights into "composable controls." He detailed five measures, including auto-mode classifiers designed to distinguish routine calls from high-risk requests to avoid "approval fatigue." Andrey Rybka, Head of Architecture & Emerging Technologies at Bloomberg, advocated for deterministic controls outside the model, such as scoped identities, tool call restrictions, and kill switches, to mitigate risks like prompt injection.

출처:latent.space

리서치

Latent Space AINews Analyzes DeepSeek's Open-Sourced Inference Engine DeepInfer: Details on 155T Daily Tokens and $38/GPU-Hour Revenue

주제 · DeepSeek推理框架중대
2026-10-21 14:54 GMT+8

Latent Space's AINews column published an in-depth report on October 21, 2026, regarding DeepSeek's open-sourcing of its internal inference framework, DeepInfer. This episode is not merely a news recap but a dual-dimensional dissection of this "Chinese open-source bomb" event from both technical and economic perspectives. DeepSeek claims the framework powers its V4.1 Flash model's production service, achieving a single-cluster throughput of 155 trillion tokens per day and generating $38 in revenue per GPU-hour under off-peak pricing strategies.

The program extensively cites DeepSeek's engineering reports and community analyses, noting that these figures rely on highly optimized caching mechanisms—a 96% input cache hit rate implies most tokens do not require recomputation. However, the report honestly highlights limitations: the so-called "fully open source" status excludes critical KV-cache management and load-balancing code, as these components remain tightly coupled to DeepSeek's internal infrastructure. Furthermore, vLLM team's subsequent comparative data showed decode throughput within 6% of DeepInfer under specific configurations, somewhat challenging the notion that DeepSeek's efficiency is inaccessible to existing open-source engines.

For readers involved in AI infrastructure investment or engineering implementation, the core value of this episode lies in demonstrating how to reverse-engineer business model feasibility from public technical disclosures. The program aggregates estimation processes by various analysts (such as zephyr_z9, Teortaxes) regarding profit margins, emphasizing the impact of hardware costs, cache storage overheads, and unaccounted operational expenses on final profitability. Although no audited financial results are provided, this economic deduction based on first-hand technical data offers a rare perspective on understanding the true cost structure of large model services.

출처:latent.space

리서치

Latent Space Pilots New Show 'High Taste Testers': Decoding Latest AI Trends via SF Lunch Conversations

간단한 요점
2026-10-17 15:47 GMT+8

Prominent AI media outlet Latent Space released an episode titled "High Taste Testers Ep. 1" on October 17, 2026. Official descriptions state that this is a pilot for a new segment inspired by their existing food show, aiming to pull listeners into a typical lunch conversation in San Francisco.

The episode covers several recent hot topics in the AI field, including Dev Day and the development of the latest models and assistants. As the first pilot episode, the primary goal is to explore this more conversational and informal presentation style and invite reader feedback. For professionals looking to understand AI industry trends from a different angle, this new format is worth listening to.

출처:latent.space

리서치

Adam Rossi, US Family Office Investor: AI Fills the Frontier Market Talent Gap, but Investors' Moat Is Capital Scarcity

중대
2026-10-23 15:17 GMT+8

Adam Rossi runs a US single-family office managing nine-figure assets. After selling his software company in 2010, he transitioned from entrepreneur to capital allocator. Over the past decade, he has consistently increased his exposure to frontier market equities and real estate, and is now actively expanding this position.

His core thesis rests on an apparently contradictory premise: AI may yield the greatest marginal benefits to the least-developed markets. The logic is as follows—the United States has invested trillions of dollars in AI infrastructure, while the cost per token to actually use AI has plummeted. A manager at a cement plant in Cambodia or a programmer at an airline in Kazakhstan can now access the same intelligent tools as white-collar professionals in London or New York, requiring only an internet connection and a modest expenditure. They do not need to invest a dollar in building data centers.

Rossi supports this leap with a set of order-of-magnitude comparisons: Fairfax County, Virginia, where he resides, has approximately 275 megawatts (MW) of data centers in operation, with another 700 MW under construction or planned; the entire country of Spain has about 440 MW; neighboring Loudoun County has over 4,000 MW in operation, more than any single European country—Germany, the largest, has around 3,000 MW. He notes that these figures use different measurement methodologies and are intended only as order-of-magnitude references. Frontier markets skip this entire layer of capital expenditure, just as Africa skipped landlines to go straight to mobile networks.

But then he makes a key pivot: AI solves the operator's problem—talent. It does not solve the investor's problem—no one wants to write the check.

He explicitly refutes the popular notion that 'AI makes frontier market investing easier.' Translation, AI-facilitated due diligence, better securities analysis—these may offer certain short-term advantages, but they are not durable. If everyone eventually uses AI, no one has an information edge in frontier markets. Analysis was never the advantage of the foreigner investing in frontier and emerging markets. The true information edge belongs to the person drinking at the bar next to the minister, the regulator's cousin, or the well-connected party official.

Rossi argues that the durable advantage for foreign investors is capital itself—patient capital. Capital is scarce in frontier markets because these places have low transparency, poor liquidity, high risk, and require longer time horizons. Many institutional investors and professional fund managers are prohibited by rules from investing in these markets. This creates perfect conditions for his family office: they can act as the marginal buyer when there is none, their check sizes are compelling, and competition for these assets is far lower than for listed equities in developed markets.

He summarizes this logic into three overlapping conditions: countries with favorable demographics and fast growth trajectories reap all the benefits of AI without bearing capital expenditures; corporate operating quality improves because the talent constraint is significantly relaxed; and the moat of capital scarcity remains unaffected by AI.

출처:Undervalued Shares

리서치

Yang Soo-yeol, Chief AI Officer at South Korean AI Data Company CrowdWorks: Pilot Failures Often Stem from Vector Database Data Gaps, Not Agent or Routing Issues

중대
2026-10-23 18:05 GMT+8

In a series report on enterprise AI agent platforms published by Byline Network (South Korea) on October 23, 2026, Yang Soo-yeol, Chief AI Officer at South Korean AI data company CrowdWorks, shared a specific debugging finding: when teams troubleshoot the reasons for pilot project failures, the issue is rarely with the agents or routing, but rather that the vector database simply lacks the data required to complete the answer. He stated, "When debugging the reasons for pilot test failures, the problem is more likely not with the agents or routing, but with the vector database lacking the actual needed data, leading to an inability to provide answers."

He also emphasized that both structured and unstructured data should be uniformly converted into machine-readable formats through a single pipeline before entering the vector database; if governance and permissions are handled entirely at the agent layer, development complexity increases. Therefore, permission information should be pre-embedded when data enters the vector database or search engine.

He further pointed out that generating answers through multiple requests during agent orchestration consumes significant tokens. To reduce token consumption at the source, data engineering must come first: reducing garbage data and establishing cleaned data portals so that the inputs for search and reasoning are inherently smaller.

출처:바이라인네트워크

리서치

Latent Space AINews: Meta Launches Muse Wonder Series, Third-Party Evals Show Near-Claude Opus 5.5 Performance

주제 · Meta Muse模型중대
2026-10-13 17:37 GMT+8

The AINews section of Latent Space has published a dedicated report on Meta's latest model releases. This episode centers on Muse Wonder and Muse Wonder Full, two frontier models launched by Meta Superintelligence Labs (MSL). As a professional source for AI news, the program goes beyond official announcements to systematically compile third-party independent evaluation results from sources like Artificial Analysis, real-time developer feedback on Twitter/X, and discussion highlights from Reddit communities, providing readers with a multi-dimensional snapshot of the new models' actual capabilities.

The episode prominently cites data from the Artificial Analysis Intelligence Index v4.3, showing that Muse Wonder Full's max-reasoning mode scored 56, while the API-available Muse Wonder scored 52. Both are higher than the previous generation Muse Spark 1.3 (48) but slightly below Anthropic's Claude Opus 5.5 (58). This comparison helps readers quickly locate the new models' tier position in the current market. Additionally, the program details performance gains in specific agentic tasks such as Terminal-Bench and GDPval-AA, along with differences in token consumption and cost efficiency across models, which offers direct reference value for assessing enterprise deployment costs.

Beyond quantitative metrics, AINews includes first-hand usage experiences from tech bloggers like @theo and @kimmonismus. They highlight the models' strengths in code review, architecture design, and handling compiler errors, while also noting limitations in creative writing and frontend visual generation. The episode specifically discusses the controversy surrounding '3D demos,' where some critics argue that poor performance in 3D graphics generation reflects over-optimization for specific benchmarks, while the Meta team responds that their training focus was on real-world tasks rather than specific demo scenarios. Presenting these industry debates helps readers understand the complexity of current large model evaluation standards.

출처:latent.space

리서치

Anthropic Engineers Explain Claude's Dynamic Workflows at AI Engineer Conference: Letting the Model Orchestrate Multi-Agents

주제 · Claude动态工作流중대
2026-10-13 02:18 GMT+8

At the AI Engineer New York conference that kicked off on October 12, 2026, Anthropic held a workshop on orchestrating its Claude Managed Agents. According to reporting by tech media outlet Latent Space, CJ Avilla and Harrison Stall, both Members of Technical Staff at Anthropic, focused on introducing a new feature called 'dynamic workflows' (or simply 'flows'). This capability aims to solve the bottleneck developers face when managing large numbers of parallel agents, allowing Claude to write scripts itself to coordinate dozens or even hundreds of sub-agents.

The presenters drew lessons from the evolution of Unix and Linux operating systems, noting that as the number of agents grows, the system requires higher levels of organization. In traditional Claude Managed Agents, developers had to manually configure a coordinator and a roster of sub-agents; in dynamic workflows, the lead agent creates a plan, and Claude generates an orchestration program that launches agents, collects results, and manages information flow. Stall stated that this allows users to accomplish discrete, verifiable units of work without having to manage any of the interconnections between them, enabling the handling of projects far larger than a single conversation context.

Beyond the core mechanism, the live Q&A touched on several key engineering challenges in multi-agent systems. When asked how to prevent simple tasks from triggering too many sub-agents and causing cost overruns, Stall replied that it relies primarily on evaluations ('evals') and lower defaults. Regarding token budgets, he revealed that the model does not know its budget while running, but developers can set a session budget. He humorously added that informing models of their budgets could make the agents 'a little bit anxious.' Additionally, Avilla mentioned a research preview feature called 'dreaming,' which helps repeated workflows become more efficient over time by reorganizing memory stores.

Although dynamic workflows cost significantly more than single-agent approaches in demo cases like stock watchlist scanning, Avilla emphasized they are 'much more effective at finding all of the details.' For AI engineers looking to understand how to apply large language models to large-scale code audits, migrations, or cross-checked research, this episode offers up-to-date practical perspectives and design trade-offs directly from a leading vendor.

출처:latent.space

리서치
다음 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크

However, he simultaneously sets strict boundaries for applicability. He believes you can outperform in frontier markets provided you meet four criteria: you have enough patience to navigate the inherent illiquidity of these markets; you can deploy sufficiently large amounts of capital to make the returns worth the trouble; you hunt with a spear rather than a net; and you possess a flexible investment mandate. If you do not meet these requirements, do not bother. He specifically points out that these markets are generally not places where buying an index or a basket of stocks will lift all boats.

He cites Swen Lorenz’s preparation of a Kyiv, Ukraine real estate fund as an example of a flexible mandate: entering an illiquid and frightening market during wartime, providing value, and then waiting for the market to turn. 'This is something no institutional mandate on earth allows you to do.'

Regarding correlation diversification, Rossi himself made corrections. He calculated the correlation of monthly USD returns from February 2016 to September 2026 (128 months) between US-listed ETF proxies (TUR, MCHI, KSA, EWZ, EPHE, GLD, BTC-USD) and SPY. He found that while some frontier markets have lower correlation with US stocks in normal times, correlations rise significantly during severe downturns or stress periods—a 'correlations go to one' event. He notes that 'severe down months' had only 11 observations, a sample size too small to serve as a prediction; it should be read as a warning. If the goal is simply low correlation and preservation of value, buying gold suffices.

Dollar management is another motivation for his diversified allocation. He criticizes the way the United States has managed the dollar since leaving the gold standard in 1971, arguing that purchasing power erodes too quickly. However, he also acknowledges the dollar's global dominance: in crises, countries hoard dollars to repay dollar-denominated debts and purchase dollar-priced commodities, which has a chilling effect on frontier economies. Each time he invests in emerging markets, he considers how easily a country can obtain dollars—only Canada, the UK, Japan, the European Central Bank, and Switzerland have standing swap lines with the Federal Reserve, none of which are frontier or emerging markets.

Rossi discloses at the end of the article: he holds positions in several of the markets discussed above. This is not investment advice.