ReadingResearchRadarInvestment framework
Sign in / Sign up
AllTechnologyProductsInvestingPublic marketsOther

Reading

2026-10-29198 posts

Refresh

Pzena Podcast: Portfolio Manager John Flynn Discusses Index Concentration and Tracking Error Risks

Material
2026-10-28 22:07 GMT+8

US investment management firm Pzena Investment Management released its podcast episode "The Index Is Now a Single Theme" on October 28, 2026. The episode features Portfolio Manager John Flynn as the main speaker, moderated by Andy Williams, Director of Client and Portfolio Services at the firm.

According to the official show notes, the core topics discussed include the historic concentration of today's equity indices, what "effective N" reveals about real diversification, why low tracking error may now hide risk rather than reduce it, and possible investment opportunities in sectors that the market fears AI will disrupt.

For institutional investors and professionals concerned with index investing risks, this episode offers an opportunity to examine current market concentration from an active value investing perspective. As no full transcript is available, this recommendation is based solely on the publisher's listed topics, providing a specific listening lead for relevant readers.

Sources:pzenaperspectives.castos.com

Research

AWS Launches Durable Execution Telemetry Exporter and MCP Server to Solve Lambda Observability Gaps

Topic · AWS持久函数可观测性Material
2026-10-29 03:36 GMT+8

Amazon Web Services (AWS) introduced the Durable Execution Telemetry Exporter on October 28, 2026, an observability plugin designed for AWS Lambda durable functions. These functions can run for minutes, days, or even up to a year, with their execution often spread across many separate Lambda invocations interspersed with waits and retries. Traditional Lambda observability tools are built for single invocations, causing one logical execution to be fragmented across multiple log streams. This makes it difficult to answer critical questions such as "how far did my durable execution get," "which step failed," or "how many times did it retry."

The new Exporter plugin runs inside the Lambda function as part of the SDK, subscribing to the SDK's instrumentation lifecycle. It builds a cumulative snapshot of the entire execution on every invocation, rather than a delta view. This means developers can generate a single JSON record containing all steps, waits, callbacks, timing, retry attempts, status, and errors with just a one-line configuration change. This record can be sent to 13 different destinations, including S3, DynamoDB, and CloudWatch, turning cross-invocation workflow monitoring into standard database queries.

Simultaneously, AWS released a companion Durable Execution Telemetry MCP server. The Model Context Protocol (MCP) is a standard that allows AI models to connect to external tools. This MCP server enables AI coding agents to directly query the telemetry data described above. Developers can ask AI agents questions in natural language, such as "Which step got slower after last week’s deployment?" or "Which customers did last night’s payment failures affect?", and the agent will retrieve answers by querying the structured telemetry records.

Both tools are released as open source, supporting TypeScript, Python, Java, and Go. For enterprises building complex agent-based workflows, this update provides a concrete path from black-box log troubleshooting to structured, queryable telemetry analysis, lowering the technical barrier to debugging long-running AI applications that involve multi-step interactions.

Sources:Amazon Web Services

Research

Replit Releases ViBench 2.0: Top AI Models Pass Only 49-65% of Full Product Builds

Material
2026-10-29 02:50 GMT+8

Replit released ViBench 2.0 on October 28, 2026, a new benchmark designed to evaluate AI coding agents on "long-horizon software development." Unlike most coding benchmarks that provide well-defined tasks within an existing codebase, ViBench 2.0 requires agents to start with an empty project and build a whole product one feature at a time, modeled after applications like Uber, GitHub, or Google Docs.

The benchmark consists of 17 products, 241 feature requests, and 544 test plans. It tests complex scenarios such as data consistency, concurrency, and state management across interacting features. For instance, in a code-hosting simulation, the agent must ensure that an editor opened before someone else saved does not erase their newer changes.

According to results published by Replit and its partner Fireworks, the best-performing models—GPT-6.1 Sol, GPT-6 Astra, and Opus 5.5—pass all test plans in only 49% to 65% of their builds. This means that in more than a third of attempts, the resulting product contains at least one critical failure. GPT-6.1 Sol achieved a 65% pass rate at a cost of about $12 per build, while Opus 5.5 had a 49% pass rate but completed builds faster, averaging 70 minutes.

Failures were often due to subtle details rather than core logic errors, such as mishandled edge cases, database initialization issues, or conflicts at feature intersections. Nearly three-quarters of failed builds broke specifically where features interacted. These findings indicate that while AI has made significant strides in assisted coding, it still faces substantial reliability challenges in autonomously building and maintaining complex software products.

Sources:Replit Blog

Research

OpenAI Staff Essay: The More Capable AI Becomes, the Easier It Is to Adopt—Three Paths for Organizational Change

Topic · OpenAI组织变革框架Material
2026-10-28 18:00 GMT+8

Hemanth Asirvatham and Elliott Mokski are twenty-four years old and already manage their own teams at OpenAI. In the second essay of their Intelligence Age series, they propose an analytical framework: there are three structural positions for AI entering corporate organizations—beneath, beside, and between—and they argue a key asymmetry: AI with a broader scope of capability is actually easier to adopt.

The three scenarios are: 'Beneath,' where everyone keeps their original position and a new layer of AI labor grows underneath, turning everyone into managers; 'Beside,' where AI joins the team as a colleague, reporting to existing human managers and gradually taking on broader functions; and 'Between,' where humans are above and humans are below, with AI coordinating and monitoring work in between, executing executive vision while identifying unique contributions from grassroots employees. The authors believe these three scenarios will not play out exactly as described, but all three are already unfolding in subtle ways.

The core mechanism running through all three scenarios is the asymmetry of adoption. The authors' reasoning is that new technology traditionally makes a restricted set of tasks cheaper, then shifts the burden to the buyer—the enterprise must retrain, reorganize, and rebuild processes around the new tool. Spreadsheet software did not install itself into the accounting department; rather, the accounting department reshaped itself around the software. Labor is the opposite: an intern, an employee, or a good team figures out how to be useful to you. They absorb your context, notice your needs, and tell you things you didn't even realize you didn't know. The best labor onboards itself.

Therefore, the more AI resembles labor—with initiative and adaptability—the less the enterprise needs to pay organizational costs to adopt it. A tool that can only perform specific tasks requires broad buy-in and retraining across the entire workforce; whereas an AI team capable of completing an entire marketing function end-to-end can be used directly by executives, much like saying a word and having someone build the whole department for you. The authors call this asymmetry 'the core of the entire transition.'

This mechanism explains why AI might enter the organization simultaneously from the bottom and the top. At the bottom, the 'Beneath' scenario relies on frontline employees adopting it first—the people best at getting AI to do the work are precisely those who used to do the work themselves. At the top, the adoption threshold for the 'Between' scenario may be lower: an AI consultant does not require technical literacy from executives, just like traditional consulting; you simply talk to your consultant, and it handles the technical details. This wave of adoption can start from the corner office.

The structural consequence of the 'Between' scenario is centralization. When an AI consultant can scan hundreds of Slack channels, documents, and code files to proactively identify communication breaks between teams, information friction between executives and frontline workers is significantly compressed. In the past, information asymmetry protected the decision-making space of middle management; now, executives can understand their large enterprises as well as a founder understands a twenty-person startup. The authors argue this does not necessarily mean employees lose autonomy—an AI provides a monitor that understands the true richness of the work. If the monitor understands multiple different paths to good results, employees can work more in their own way.

Sources:OpenAI

Research

Apple Launches First Touchscreen MacBook Pro, Reversing Decade-Old Design Stance Under New CEO Ternus

Material
2026-10-29 01:08 GMT+8

Apple announced on October 28, 2026, its first laptop with a touchscreen, named the "MacBook Pro with ProLED XDR Display." This marks a formal reversal of Apple's years-long resistance to putting touchscreens on its laptops, executed under new CEO John Ternus. The new models go on sale November 6, starting at $2,899 for the 14-inch version and $3,399 for the 16-inch version, representing notable price increases over existing models.

The launch comes during one of the worst slumps in the global personal computer (PC) market. According to IDC data, total PC shipments dropped 20% year-over-year in the third quarter of 2026, with full-year shipments expected to fall 11%. Analysts attribute higher laptop prices and hurt sales to an AI-driven global shortage of memory. Despite generating less than 10% of Apple's total revenue ($10.35 billion in the most recent quarter), Ternus highlighted the growing strategic importance of the Mac, particularly for local or edge AI that runs on-device.

Apple is leveraging hardware-software integration to attract developers. Recently, Meta's AI model development team purchased hundreds of Mac Minis to test agents, and Apple teamed up with Meta to announce that large, nearly frontier models can run on clusters of Macs connected with high-speed cabling, allowing developers to run agentic code without paying Anthropic or OpenAI. The new MacBook Pro uses OLED technology for better brightness and color, features M5 Pro or Max chips, and is lighter than current pro laptops.

However, whether touchscreens will drive purchasing decisions remains questionable. Ryan Reith, IDC VP for worldwide device trackers, noted that while touchscreen laptops have been available on Windows machines for a long time, they are just another feature, not a selling point behind most purchasing decisions. Only 17% of notebook PCs shipped globally this year will have touchscreens. Reith stated, "This is really a user interface and software thing... It's one more thing that can draw the developers in."

Sources:CNBC

Research

South Korean Delivery Platform Baemin Discloses AI Search Results: Semantic Matching Cuts Null Results by 80%, Boosts Order Metrics by 50%

Topic · Baemin语义搜索Quick take
2026-10-28 17:00 GMT+8

Woowa Brothers, the operator of South Korea's largest food delivery platform Baemin, disclosed advancements in its search technology at the 'Woowa Tech Conference 2026' held in Seoul on October 28. The company stated that it has enhanced its system using AI-driven semantic search to overcome the limitations of traditional keyword matching.

According to server developer Lee Seung-hyo and machine learning engineer Yeom Ha-neul, who presented the case study, the legacy system compared input terms strictly against store and product names. This often led to irrelevant results for intent-based queries, such as showing shrimp chips when a user searched for 'diet snacks'. After adopting semantic search, which analyzes the meaning behind the query, the rate of searches returning no results decreased by approximately 80%.

In specific experiments where exact keyword matches failed, the introduction of semantic search resulted in an increase of over 50% in order-related metrics. Engineer Yeom noted that while this segment represents a small portion of total traffic, the improvement was significant enough to impact overall search indicators. The company also implemented personalization techniques that rerank results based on users' past order history and is expanding towards conversational AI search capabilities.

However, challenges remain. The embedding models currently struggle with complex conditions involving negations, such as failing to exclude peanut-containing menus for a request stating 'I have a peanut allergy'. The company indicated it is developing a more robust query understanding model to address these nuances.

Sources:zdnet.co.kr

Research

Oracle Sets FY30 Revenue Target Above $225 Billion, Caps Additional Capital Raises at $20 Billion

Topic · OracleStructural
2026-10-29 02:21 GMT+8

US cloud database and infrastructure company Oracle Corporation (NYSE: ORCL) hosted an Investor Day in Las Vegas, Nevada, on October 28, 2026, releasing a long-term financial outlook through fiscal year 2030. Co-CEOs Clay Magouryk and Mike Sicilia, along with CFO Hilary Maxson, presented the strategy, focusing on AI-driven revenue growth and a clear path to self-funding capital expenditures.

According to the new guidance, Oracle expects total revenue to grow at a compound annual growth rate (CAGR) of more than 35% from FY26 to FY30, reaching at least $225 billion in FY30. The company also projects non-GAAP operating margins exceeding 40% and non-GAAP earnings per share (EPS) above $21. This forecast significantly exceeds traditional software valuation models, reflecting management's confidence in the expansion of its cloud infrastructure and AI application businesses.

Addressing investor concerns about capital intensity, Oracle defined strict boundaries for future funding. The company committed to limiting additional capital raised in the markets during the FY27-FY30 investment cycle to $20 billion, excluding refinancing of existing debt. This includes approximately $10 billion in equity and $10 billion in gross debt. This cap is in addition to the $20 billion at-the-market equity raise completed in the first quarter of fiscal 2027.

Regarding cash flow, Oracle expects to achieve positive free cash flow in the second half of fiscal 2028 and targets at least $30 billion in annual free cash flow for fiscal 2029. The company stated that growth will be self-funded beginning in FY29, with net debt to adjusted EBITDA falling below 3.5 times. Vendor financing, customer-provided hardware, and customer prepayments are expected to reduce upfront cash requirements during the transition.

Oracle also reaffirmed its fiscal 2027 short-term guidance, projecting total revenue of at least $90 billion and non-GAAP EPS of $8.10. Capital allocation priorities remain maintaining investment-grade credit ratings, paying dividends, and funding growth opportunities. These disclosures provide a concrete benchmark for assessing the capital intensity of the AI infrastructure sector and its impact on public company balance sheets.

Sources:prnewswire.com

Research

AWS Ray Train DLC: Distributed PyTorch Training Usually Stalls at the Container, Not the Model Code

Material
2026-10-29 02:19 GMT+8

People doing distributed PyTorch training often focus their efforts on model code, but what actually prevents training from scaling from a single machine to a multi-node cluster is usually the container. Each container requires the same set of components: a matching CUDA runtime, a PyTorch build compiled against that CUDA version, NCCL (NVIDIA Collective Communications Library), Elastic Fabric Adapter (EFA) userspace libraries, the AWS OFI NCCL plugin that connects NCCL to EFA, and a Ray version approved by the cluster operator. Building this image and rebuilding it whenever any one of these components changes is an engineering project in itself, entirely unrelated to the model you intend to train.

AWS has packaged this stack into the Ray Train Deep Learning Container (DLC), publishing it on Amazon ECR Public and keeping it patched. The article runs an example using this image as-is: fine-tuning Qwen2.5-1.5B with PyTorch FSDP across eight NVIDIA L4 GPUs on two g6.12xlarge instances. The training code consists of a single function, the cluster is defined by a RayCluster manifest plus a variables file, and no image is built at any point.

The networking stack is the part most teams least want to maintain themselves. The key risk here is not an error message, but silent degradation: if the EFA userspace library and NCCL versions do not match, NCCL falls back to TCP sockets. The job completes successfully, but at a fraction of normal speed. The DLC builds and tests these libraries together; AWS gates each release with multi-node EFA and multi-node KubeRay training tests, which are exactly the paths used in this example.

All differences outside the image are resolved through configuration rather than rebuilding. The cluster layout is described by a KubeRay manifest: a CPU system node runs the KubeRay operator and the Ray head (requesting no GPU), while two g6.12xlarge nodes each run a worker pod requesting four GPUs and one EFA interface. Network transport is determined by four environment variables: FI_PROVIDER is set to efa so libfabric uses only EFA; NCCL_NET_PLUGIN is set to ofi to load the AWS OFI plugin; NCCL_SOCKET_IFNAME points to the pod's eth0; and NCCL_DEBUG is set to INFO to verify the actual transport chosen. These variables are written into the manifest, making behavior visible in a single file and independent of image defaults.

Inside the container, Ray Train handles the launcher work. Multi-node PyTorch typically means starting a process per GPU on every host and assigning each process a rank, world size, and rendezvous address. With Ray Train, you write the training loop as a function and pass it to TorchTrainer; Ray automatically starts one worker per GPU wherever the cluster has GPUs, forms torch.distributed process groups, and runs the function on each worker. The example requests an FSDP strategy, sharding model weights, gradients, and optimizer state across all eight GPUs. For Qwen2.5-1.5B in bfloat16, these four tensors occupy about 12 GB; sharded eight ways, each L4 GPU holds approximately 1.5 GB.

Sources:Amazon Web Services

Research

US Consumer Confidence Rises in October for First Time in Four Months, Still on Downward Trend

Material
2026-10-28 06:30 GMT+8

According to data released by The Conference Board, US consumer confidence rose in October, marking the first increase after four consecutive months of decline. This shift was partly helped by easing retail gas prices, although the report noted that despite the monthly gain, "confidence remained on the long-term downward trend."

Energy costs have been a significant driver of recent consumer sentiment. Since the start of the Iran war, US gas prices are 40% higher than one year ago, spiking shipping costs and the price of many household items. This inflationary pressure affects not only household budgets but also the political landscape. A former economic adviser to US President Donald Trump told The Associated Press, "People hate paying more money for gas," suggesting Republicans may pay a high price for this in the upcoming midterm elections.

Meanwhile, markets expect the Federal Reserve to hit pause on interest rates this week, though investors predict this pause will not last long. Despite the short-term improvement in data, the gloomy economic outlook continues to cloud Republicans' midterm prospects.

Sources:semafor.com

Research

a16z Survey of 100 CISOs: AI Security Priorities Completely Rewritten, but Production Deployment Far Lags Procurement Intent

Topic · AI智能体安全治理Material
2026-10-29 00:00 GMT+8

The security priority list has been completely rewritten in one year. A survey by a16z of 100 CISOs at companies with more than 1,000 employees shows that when asked to recall their 2026 priorities, exposure and vulnerability management and SIEM ranked at the top; whereas all seven highest-ranked initiatives for 2027 involve AI. Internal AI agent security, SOC automation, and offensive security saw the largest gains, followed closely by code security, remediation, and AI governance.

Two factors are driving this shift: enterprises need to protect the AI that employees and customers are already using, and they also want to apply AI internally within security teams—to help investigate attacks, test systems, and fix discovered issues.

Budget directions align with outside expectations. 87% of surveyed CISOs expect an increase in cybersecurity spending in 2027, primarily driven by AI-enabled threats and the expanded attack surface resulting from their own AI deployments. Most additional spending flows into tools, but not at the expense of headcount: nearly half expect security team sizes to increase, another half expect them to remain flat, and only 4% expect cuts. The survey interprets this choice as a coverage requirement—teams want to investigate alerts that would otherwise backlog, test more applications, and close vulnerabilities earlier.

However, there is a significant gap between intent and production. Only 18% of enterprises report that AI SOC tools have entered production environments, and 13% report that AI offensive security or AI exposure remediation has gone live. Most enterprises are still piloting or expect to reach production within the next year. Procurement plans are more urgent: 45% plan new purchases or upgrades for internal AI agent security within twelve months, 42% for offensive security, and 38% for AI usage and governance.

Offensive security is shifting from periodic assessment to a continuous capability. a16z notes in its article that Bain & Company estimates AI can compress a typical attack sequence from about four weeks to about eighteen hours, significantly shortening the time defenders have to find and fix exposures. 83% of respondents consider more frequent security posture testing important, with 54% considering it very important. Currently, 18% of enterprises say most of their offensive testing runs continuously against production assets, and 52% expect to reach this state within a year; five years out, 44% expect testing to be embedded in CI/CD processes, triggered by code or infrastructure changes.

Discovering vulnerabilities is the fast part; fixing them is the bottleneck. A fintech CISO stated, 'These AI tools allow you to find vulnerabilities faster, but not patch them faster.' 70% of respondents want to retest and verify that vulnerabilities have been fixed, 67% want to prioritize remediation using validated exploitability, and 57% want automated generation of remediation tickets. An executive at a large cybersecurity platform described the opportunity as 'find it, fix it'—tracking from discovery to deployable changes, then rerunning attacks to confirm the exposure is closed.

Sources:a16z.news

Research

South Korean Delivery Platform Baemin Discloses Robot Tech Upgrades: AI Simulation Cuts Outdoor Testing Time by 40%

Topic · DeliQuick take
2026-10-28 17:40 GMT+8

Woowa Brothers, the operator of South Korean food delivery platform Baemin, disclosed technical progress for its delivery robot 'Deli' at the Woowa Tech Conference 2026 held in Seoul on October 28. The company stated that outdoor testing time was reduced by 40% through the use of artificial intelligence for training autonomous driving models and verifying performance in virtual environments.

According to Woowa Brothers, traditional modular autonomous driving systems required separate modifications to perception, planning, and control modules for new scenarios, with each improvement and validation taking 3 to 4 weeks. To address this, the company is transitioning to an 'End-to-End' learning approach, training the entire model directly on driving data to allow performance to improve continuously as data accumulates.

In terms of data collection, the company optimized its strategy to prioritize data from alleys and unexpected situations rather than relying solely on main roads. As a result, the proportion of roads with collected driving data rose from 13.7% to 28.2%, nearly doubling. Woowa Brothers emphasized that the key to improving model performance lies in the novelty and diversity of data, not just the volume.

Additionally, the company built an AI-based simulation verification system. Previously, constructing a 0.085 square kilometer 3D virtual road took 40 hours, whereas the automated system can now generate a 4.4 square kilometer model in 34 minutes. After software updates, the system automatically runs simulation tests in the early morning, with AI initially identifying potential error causes before human judgment. This process reduced the outdoor testing cycle from 5 days to 3 days.

Sources:zdnet.co.kr

Research

Anthropic Launches Three Claude Enterprise Features to Help Track AI ROI

Topic · Anthropic上市Material
2026-10-29 02:00 GMT+8

US-based AI company Anthropic announced that three new features for its Claude Enterprise plan are now Generally Available, designed to help business clients better track their AI return on investment (ROI). The features—smart reports, pooled budgets, and in-app surveys—were previously released in beta in September.

Smart reports, which are off by default, track what tasks employees are using Claude for, how often they run them, and the associated costs. They also document where sessions encounter friction, helping admins identify routine workflows that can be automated. Anthropic explicitly stated in its documentation that the tool “isn’t designed and should not be used for evaluating individual performance or making employment decisions,” a relevant disclaimer given AI's impact on HR.

Pooled budgets allow organizations to set monthly spend caps for specific groups while maintaining individual seat limits. Admins can optimize budgets based on historical spending data, ensuring highly active users aren't hamstrung by fixed limits while others leave spend unused. In-app surveys enable admins to ask employees about their Claude usage directly within the Chat interface and view aggregate responses in the admin console.

This update comes as Anthropic prepares for its Initial Public Offering (IPO). The company has filed its prospectus and faces market scrutiny over the sustainability of its enterprise AI growth. With competitors like OpenAI catching up in the enterprise market, providing clear ROI tracking tools helps Anthropic retain large customers and justify its premium enterprise subscription fees ($20 per seat per month plus API token rates).

Sources:ZDNET

Research
Next →
Sign in / Sign up
ReadingResearchRadarInvestment framework

But the authors also acknowledge this is a double-edged sword. The same AI consultant can increase trust through intelligent monitoring, but it can also more vigorously enforce the CEO's preferred way of doing things. Enterprises value consistency—not just good decisions, but decisions compatible with the company's culture, brand, and economies of scale. The ultimate direction depends on the choices of executives, not the technology itself.

It is important to clarify that this essay is a personal opinion piece by two OpenAI staff members. The authors note that it reflects their views, not those of OpenAI or their colleagues. They admit they do not know the actual capabilities of models in terms of management scale—existing evaluations mostly target execution-focused cognitive labor, rarely capturing large-scale orchestration at the manager and executive levels. These three scenarios are conceptual frameworks, not verified organizational trends.

The cost is communication: FSDP gathers layer weights before use and reduce-scatters gradients afterward, meaning the entire model crosses between ranks on every step. This inter-instance traffic is carried by the EFA stack included in the DLC. The article concludes with an attribution result: after enabling the EFA path via four environment variables, throughput doubled compared to TCP with unchanged code. This is an AWS demonstration conclusion; the example uses random token batches, and the evaluation criteria are whether FSDP collectives actually crossed both nodes and whether the task completed, not training convergence quality.

The value of this entire solution lies in removing container building from the training workflow. The image is public, allowing nodes to pull it without credentials; the entrypoint script checks the host NVIDIA driver at startup and automatically enables bundled CUDA forward-compatibility libraries if the host driver is older than the image's CUDA. The same tag runs both head and worker pods with no role distinction. When additional Python packages are needed, they are installed via Ray's runtime_env at submission time, leaving the image unchanged. To change models, adjust steps, modify sequence length or batch size, only the parameters in the variables file need changing. Switching GPU instance families requires only confirming that the new family supports EFA and is on the device plugin allowlist; neither code nor image needs modification.

The core constraint of SOC automation is analyst time. Processing an alert still requires pulling evidence from multiple systems, reconstructing the event, and determining if there is a real threat; by the time the full picture is assembled, hours have passed and the response has not even started. AI SOC tools can handle evidence collection and routine checks, allowing analysts to focus on complex cases requiring judgment. However, the degree of authorization is an independent decision: today, 36% of respondents accept agents acting autonomously on low-risk actions, and this proportion is expected to rise to 60% in 12 months, with only 7% expecting to allow autonomous action in most cases.

The proliferation of internal enterprise AI is equally astonishing. 92% report at least limited production use of employee generative AI tools, 83% report coding assistants or agents are in production, and 55% have MCP servers or other agent-to-application integrations in production. These systems can access source code, retrieve company documents, and execute operations across applications. 47% cite employees leaking sensitive data to AI tools as one of their biggest concerns, 46% mention agents exceeding permissions or being hijacked, and 29% worry about prompt injection or data leakage through AI applications. 63% of responses describe security slowing down at least some AI adoption, including review, approval requirements, and restrictions on specific uses.

A structural dilemma lies in the fact that agents run on employee devices using the user's own permissions; therefore, for endpoint security tools built over the past decade, an agent reading a file or calling an API is indistinguishable from the employee doing it. EDR, DLP, and Zero Trust fail at this level.

In procurement decisions, existing platforms still get the first call. 44% of respondents said they first look for AI capabilities within existing security platforms, compared to only 21% who start with AI-native vendors. But the first call does not equal the last: 47% said they would consider choosing a startup if it could solve capability gaps that existing vendors cannot cover. One respondent bluntly stated, 'I don't want to buy another point product every time a new problem arises.'

Market noise is another obstacle. a16z cites Jake Black Hat 2026 vendor landscape, pointing out that some categories have up to 43 vendors. Buyers need to understand how much work each product completes, how much follow-up the team still handles, which operations can be safely executed, and what the running costs are—and these answers require time spent by the people the product is supposed to help. A CISO at a global apparel company described narrowing down seven or eight vendors to three for hands-on testing: 'Doing proof of concepts for all seven or eight is humanly impossible.'

a16z holds portfolio interests in several categories covered by the survey, including offensive security company Armadin, endpoint AI security company Neo, SOC automation tool Cotool, code security company Socket, social engineering simulation company Adaptive Security, and brand impersonation detection company Doppel. Armadin's Series B was co-led by a16z and Kevin Mandia; Neo's founding team previously participated in building SentinelOne. The survey conclusions have interest associations with these investment directions, and readers should evaluate their analytical stance accordingly.