ReadingResearchRadarInvestment framework
Sign in / Sign up
AllTechnologyProductsInvestingPublic marketsOther

Reading

2026-11-2154 posts

Refresh

AWS launches managed cluster extensions for SageMaker HyperPod Slurm

Material
2026-11-21 07:07 GMT+8

Amazon Web Services (AWS) has launched managed cluster extensions for Amazon SageMaker HyperPod Slurm clusters. This new feature brings the same declarative, API- and console-driven approach used for infrastructure configuration to the software running on the cluster's nodes, specifically targeting observability and user management.

Previously, managing software on shared Slurm clusters relied heavily on lifecycle scripts that ran only during node provisioning. As researchers joined or left, or as teams requested specific metrics, these scripts required manual editing and node replacement, leading to configuration drift where nodes became inconsistent over time. The new extensions allow administrators to declare desired software states via an API or the AWS console. AWS validates these settings before application and ensures consistency across all nodes, including those added by scaling or replacement.

Two initial extensions are available: sagemaker-observability, which collects node, GPU, Elastic Fabric Adapter (EFA), and Slurm job metrics to send to Amazon CloudWatch or Amazon Managed Service for Prometheus; and sagemaker-user-management, which creates users and groups with consistent IDs across nodes, controls SSH/sudo access, registers users with Slurm accounting, and can integrate with Active Directory. Users retain full direct access to nodes for custom configurations, but the extensions provide a managed path for common operational tasks.

Sources:Amazon Web Services

Research

Ben Carlson: Why 'Buy the Dip' Survived the Dot-Com Crash but Not the Financial Crisis

Material
2026-11-21 05:30 GMT+8

Ben Carlson, a portfolio manager at Ritholtz Wealth Management, argues that the retail investor habit of buying the dip in tech stocks has a specific breaking point: not the depth of the drawdown, but the severity of the economic conditions that accompany it.

The observation starts with Nvidia. After Jensen Huang's scandal and resignation in October 2026, the stock fell more than 40% since early-October. On the day of the resignation itself, it dropped 9%. Retail investors responded by buying aggressively—Vanda Research data showed what Carlson calls "an insane amount" of net buying. This is not unusual behavior for this decade. Nvidia had already suffered three separate drawdowns exceeding 30% (the 2020 crash, the 2022 bear market, the Liberation Day sell-off) plus eight other double-digit corrections, and every time dip-buyers were rewarded.

Carlson's argument is that this conditioning creates a specific vulnerability. If the next downturn is accompanied by a shallow recession—like the 2001 contraction that lasted eight months with GDP falling just 0.3% and unemployment peaking at 6.3%—investors will likely hold or buy more, as they did during the dot-com bust. Household equity ownership remained stable from 2000 through 2002 and into the initial years after the crash, despite the S&P 500 falling 50% and the Nasdaq 100 dropping more than 80%. Maggie Mahar's documentation in Bull: A History of the Boom and Bust showed that net redemptions from equity funds in the twelve months ending March 2003 amounted to just 3.6% of the sector's assets.

But if the next downturn brings the kind of economic pain that accompanied the Great Financial Crisis—a year-and-a-half recession, GDP falling more than 5%, unemployment exceeding 10%, and a housing collapse with mass foreclosures—then dip-buyers will run out of cash and conviction. Household stock ownership fell from 63% in the mid-2000s to 52% by 2013, and it took that combination of job losses and housing destruction to break the habit.

The mechanism is straightforward: "It's much easier to step in and buy with the unemployment rate at 4%. Once you lose your job it's much harder to worry about the stock market." Risk tolerance is not just a psychological trait; it's constrained by income stability and balance-sheet health. A buy-the-dip mentality can survive a minor recession but faces a much harder test during a financial crisis.

Carlson does not predict which scenario will materialize. He explicitly says "I don't know" whether Nvidia will recover like Apple did after Steve Jobs, or whether this is a case where a huge company stumbles and never regains its perch. He cites JP Morgan's Michael Cembalest on the history of mega-cap companies knocked from their position who never return. The uncertainty is part of the argument, not a hedge against it.

Carlson closes by invoking the old saying that during bear markets, stocks return to their rightful owners—and then adds the qualification that matters: "Those rightful owners aren't always the people with the most discipline. Sometimes it's simply the people who can survive the downturns with their job and savings in tact." The relevant constraint on dip-buying resilience is not conviction or analytical skill but whether one's income and balance sheet can withstand the economic conditions that accompany the crash.

Sources:A Wealth of Common Sense

Research

Google Launches Semantic Governance Policies for AI Agents in Preview

Material
2026-11-20

On November 20, 2026, Google announced the preview release of Semantic governance policies within its Gemini platform. This feature addresses a critical gap in enterprise AI security: traditional permissions define what an agent can do, but not whether a specific action aligns with business rules or user intent. By introducing a policy enforcement point outside the agent's code, Google aims to prevent unauthorized actions such as data leaks or hijacked tasks.

The mechanism operates by intercepting every proposed tool call from an AI agent before it executes. An LLM judge evaluates the call against two criteria: first, against plain-language policies written by compliance teams (e.g., "Do not share quarterly results before earnings"); second, against the original user request to ensure the agent has not deviated from its intended task. If either check fails, the call is blocked, and a denial message is returned to the agent. This process runs through Agent Gateway, ensuring that governance logic remains separate from the agent's underlying code and model instructions.

This approach distinguishes itself from existing controls like IAM (Identity and Access Management) or Model Armor. While IAM restricts access to tools and Model Armor screens for prompt injections, Semantic governance policies add an inline, semantic layer of defense. For example, if a user asks an agent to send financial data to an external investor under the guise of an NDA, the system can deny the request based on a policy prohibiting pre-earnings disclosures, even if the agent technically has permission to send emails. The verdicts and rationales are logged in Cloud Logging for audit purposes.

Google notes that because the judgments are made by an LLM, they are probabilistic and may vary. Enterprises are advised to use dry-run mode to test policies before enforcing them. The feature supports various models, including those accessed via the Gemini API, Anthropic Messages API, and OpenAI-compatible endpoints, allowing organizations to apply consistent governance across different AI stacks.

Sources:developers.googleblog.com

Research

AWS Launches SageMaker HyperPod Global Inference Gateway for Cross-Region AI Routing

Material
2026-11-21 06:35 GMT+8

Amazon Web Services (AWS) has introduced the Amazon SageMaker HyperPod Global Inference Gateway, a new managed service designed to route artificial intelligence inference requests across multiple compute clusters and AWS Regions through a single endpoint. This addresses a critical infrastructure challenge where GPU capacity is fragmented across different locations, leading to idle resources in one region while requests queue in another.

The gateway operates as a second tier in the HyperPod architecture. While the existing Inference Gateway routes requests to specific pods within a single cluster, the Global Inference Gateway selects the best-suited cluster from all registered clusters. It uses four weighted signals to make routing decisions: KV cache headroom (available memory), queue depth (current load), region proximity (latency), and error rate (health). By default, the system prioritizes proximity but can be configured to favor resilience by increasing the weight of error rates.

For enterprises, this update eliminates the need to build custom cross-region networking, load balancers, and DNS management. Traffic is carried over the AWS backbone using AWS PrivateLink, ensuring privacy and reducing latency compared to public internet paths. Automatic failover moves traffic away from failed clusters without requiring client-side configuration changes. Authentication is handled via AWS Signature Version 4 (SigV4) and Identity and Access Management (IAM).

Jinhyung Park, VP of AI Platform at Bagelcode, noted that the beta version allowed their small platform team to achieve cross-region failover without maintaining their own routing layer. The feature is available now for users with multiple SageMaker HyperPod clusters.

Sources:Amazon Web Services

Research

South Korea deploys emergency disaster alerts to catch AI cheating in college entrance exam

Material
2026-11-20

The South Korean government deployed emergency disaster alerts during the College Scholastic Ability Test (CSAT), also known as Suneung, to detect prohibited electronic devices. This novel enforcement method caught 54 students attempting to use AI-assisted cheating tools, including micro-cameras and smartphones.

The crackdown was prompted by a recent confession from Han Won-seok, the top scorer of the Certified Tax Accountant exam, who admitted to using a micro-camera that fed test images to an AI on his smartphone. Following this incident, Prime Minister Lee Jae Myung stated that AI abuse in public examinations undermines societal fairness and called for a system overhaul.

During the CSAT, three alerts were triggered at specific times: 8:30 AM, 2:21 PM, and 2:57 PM. The first alert caught 31 students before the exam started, while the subsequent two caught 23 more. In total, 111 students were identified with unauthorized devices when combined with other detection methods such as metal detectors and WiFi signal monitoring.

Sources:Tom's Hardware

Research

Independent mathematicians verify AI proof of Riemann hypothesis, error chance near zero

Structural
2026-11-21 05:00 GMT+8

Mathematicians are now almost certain that the AI-created proof of the Riemann hypothesis is valid. Anthropic mathematician Levent Alpöge claimed to have used the company’s Claude AI model to solve the problem, which was first posed in 1859 and suggests a hidden pattern underlies the distribution of prime numbers.

Alpöge initially checked the work by writing out the proof using the programming language Lean; if a Lean-coded proof compiles error-free, this is usually taken as confirmation that the logic and math in each step is correct. Independent researchers, including Andrew Sutherland, a mathematician at M.I.T., have now separately compiled Alpöge’s result in Lean without issues.

“The Lean proof has been compiled and independently re-checked by the Lean kernel with only the three standard axioms,” Sutherland says. “Honestly the chance that anything is fishy here is approximately zero.”

However, some questions remain regarding whether the proof addresses the exact statement of the Riemann hypothesis or establishes a subtly different point. Researchers are still making sure that nowhere in the proof was there a hack, dodge or shortcut that changed the meaning of the initial hypothesis or any other mathematical rules or truths. This process, called a “comparator run,” typically takes a matter of hours.

Sources:Scientific American

Research

Why AI Companies Are Now Honey-Trap Targets, and Why Their Culture Has Not Caught Up

Material
2026-11-21 04:58 GMT+8

The article's central argument, attributed to former CIA officer Aaron Brown and cybercrimologist Perry Carpenter, is that the AI industry became a national-security-relevant sector faster than its organizational culture adapted to the resulting threat model. Where honey traps historically targeted politicians, the military, and national intelligence—the people behind clearances and a workforce trained to expect an approach from spies and other attackers—AI companies now hold the secrets that matter: chip roadmaps, export decisions, customer lists, and who gets the technology.

The mechanism Brown describes is straightforward. A spy wants what the CEO can touch. The roadmap, the export decisions, the customer list, who gets the technology. The CEO is the shortest path to all of it. But the attack does not require romance. Friendship is the oldest honest trap scheme, Brown says. Attackers gain access as investors, customers, recruiters, and consultancies. Their requests look legitimate on paper. They do not need targets to fall in love—they need targets to feel comfortable talking.

What makes this newly urgent is the cost curve. Planning a scheme used to take months or years: research, crafting a persona, knowing the target, executing a plan. AI now compresses that timeline dramatically. It can research a target, generate fake social profiles, fabricate backgrounds, even produce fake voice calls. Carpenter notes that AI can help sustain an entire fictional backstory, though the deception ultimately plays out through human relationships.

The practical detection advice is concrete but not reassuring. Carpenter suggests watching for coincidental encounters that happen a little too often, persistent curiosity about work and colleagues, requests for secrecy about meeting up, and pressing for introductions. Brown offers a simpler gut check: would this person still be around if you lost your job tomorrow?

The structural vulnerability is organizational. Carpenter suggests that within companies, it also helps to eliminate executive exceptionalism—the idea that rules do not apply to people in power. Brown adds that company culture determines whether employees feel comfortable raising concerns, challenging unusual requests, and escalating possible threats. Carpenter's observation that one of the most dangerous security cultures is one where nobody feels comfortable questioning the CEO is the article's sharpest point.

The strongest objection to this framing is that it describes a threat landscape without quantifying it. The article provides no evidence of how many AI executives have actually been compromised beyond the single contested Nvidia case. The detection heuristics are generic security awareness advice without empirical validation. And the claim that AI has made honey trapping much easier and cheaper is asserted by practitioners but not measured—no study compares pre-AI and post-AI attack success rates or costs. The argument works as a warning about a structural gap between industry importance and security culture, but it cannot yet establish how large or imminent the problem is.

Sources:TechCrunch

Research

Amazon Textract Insights GA: Verifiable generative AI for document extraction

Material
2026-11-21 04:22 GMT+8

Amazon Web Services announced the general availability of Amazon Textract Insights, a new generative AI feature within its document processing service. The tool allows users to define desired fields in natural language using JSON Schema "blueprints" and submit single or mixed multi-document packets (up to 3,000 pages). Unlike general-purpose large language models, Textract Insights returns structured results where each extracted value carries a confidence score and precise page-level grounding (bounding box), enabling auditable automation without model training.

The service addresses challenges in high-value workflows like mortgage underwriting and insurance claims, where documents arrive as interleaved scans of W-2s, payslips, and handwritten notes. Textract Insights automatically splits these mixed packets, matches sub-documents to appropriate blueprints, and extracts fields. It distinguishes between explicit fields read verbatim and inferred fields derived from content, providing traceability for reviewers. This capability aims to reduce reliance on custom OCR pipelines or ungrounded LLM outputs, which often lack calibrated confidence metrics necessary for automated decision-making.

Sources:Amazon Web Services

Research

Gemini Enterprise: What Authentication a Custom UI Inherits From the Platform

Quick take
2026-11-20

["When enterprises build a custom interface on top of Gemini Enterprise rather than using its built-in web app, they inherit three authentication flows the platform previously handled automatically. Neeraj Agrawal, a Forward Deployed Engineer at Google Cloud, walks through each layer in a November 2026 guide: authenticating the user to the StreamAssist API (via Google Identity, Workforce Identity Federation, or Workload Identity Federation), authorizing Federated Connectors through 3-Legged OAuth with tokens stored in Gemini Enterprise's managed Keystore, and handling per-tool OAuth for custom ADK agents deployed on Vertex AI Agent Engine.","The most transferable insight concerns the third layer. Agrawal explains that attaching static OAuth configurations at the root agent registration level blocks the entire agent upfront — a user cannot ask a general policy question without being forced through every connector login. The solution is wrapping protected sub-agent tools with GcpAuthProvider backed by GCP Auth Manager's Token Vault, deploying with Agent Identity enabled, and triggering 3LO popups only when a protected tool is actually called. After the first authorization, subsequent queries run in a single turn with no popups.","This is vendor-authored guidance for Google's own product. The specific components (GcpAuthProvider, AuthenticatedFunctionTool, SPIFFE identity, Keystore) are tightly coupled to Google Cloud infrastructure, and no independent validation of the pattern's effectiveness or adoption beyond this ecosystem is provided."]

Sources:developers.googleblog.com

Research

Private MCP Connectivity for Gemini Enterprise: Egress Gateway, PSC Network Attachment, and FQDN-Scoped DNS

Material
2026-11-20

When an enterprise moves a custom MCP server from prototype to production on Cloud Run, three constraints collide simultaneously. Security policy demands internal-only ingress, blocking all public traffic. The platform requires both service-level and user-level authentication on every tool call. And in large Shared VPCs hosting hundreds of Cloud Run services, a naive DNS override can hijack resolution for every other service on the network. Neeraj Agrawal, a Forward Deployed Engineer at Google Cloud, describes a three-part architecture that resolves all three at once.

The first component is an Egress Agent Gateway in Agent-to-Anywhere mode, connected via a PSC Network Attachment to a regional subnet. This places a managed network interface inside the enterprise VPC without public IPs or inbound firewall rules — the packet enters the private network at the attachment point rather than traversing the public internet. From inside the VPC, the request reaches Cloud Run's serverless frontend through either Private Google Access virtual IPs (199.36.153.8/30) or a global PSC endpoint (240.0.0.10), depending on whether the subnet retains a default internet route.

The second component addresses the DNS problem specifically. Without Private Cloud DNS, public resolution maps .run.app domains to public IPs that time out because the PSC Network Attachment does not route to the internet. The intuitive fix — a wildcard *.run.app private zone — breaks every other Cloud Run service in the Shared VPC. The solution Agrawal describes is to create a Private Cloud DNS zone scoped exclusively to the single MCP service's exact fully qualified domain name, then peer only that domain with the Agent Gateway. Because Cloud DNS uses longest-suffix matching, all other services continue resolving through their normal paths.

The third component handles authentication. Gemini Enterprise sends two distinct headers on every request: an X-Serverless-Authorization header containing a Google-signed OIDC ID token for the service agent, which Cloud Run IAM verifies at the serverless edge and strips before the request reaches the container; and an Authorization Bearer header containing the signed-in user's OAuth 2.0 access token, which passes untouched to the MCP application code for per-user permission enforcement. The provided Python middleware example shows how to validate the user token while allowing health checks to bypass authentication.

The architecture is documented in a companion GitHub repository with sequential setup scripts covering both the Private Google Access and global PSC endpoint options, along with Cloud Console instructions and Cloud Logging verification steps.

Sources:developers.googleblog.com

Research

How vLLM Serves a Looped Transformer: Per-Pass Caches as Regular Layers, Fused Message Reads, and a Variable-Pass Tradeoff Table

Material
2026-11-20

Inkling 2 Mini's architecture creates a concrete serving problem: its 24 physical layers split into a prelude (layers 0–5), a loop core (layers 6–17) and a coda (layers 18–23), and the core runs three times per token with shared weights. Each pass maintains its own KV cache and reads gated messages from the pass before it. The model therefore has 48 KV caches and 48 conv-state caches despite having only 24 layers.

vLLM's solution is to register each extra pass of a core layer as its own attention layer with its own KV and conv-state cache. The KV cache manager and prefix caching then treat these caches like any others, so serving the loop needed no new cache-management logic. This is the central engineering decision: the loop is absorbed into the existing layer abstraction rather than requiring a new execution path.

The cross-pass messages are the second challenge. From the second pass onward, each core sublayer input — the attention's query, key and value inputs and the MLP input — gets its own message, read from the previous pass's two blocks by an attention-residual read with a learned query per layer and input type. A rank-64 controller then merges the message into the input using a gated formula: the output is (1 − tanh(W_forget · h)) · input + tanh(W_gain · h) · message, where h is computed from RMS-normalized input and message via a down-projection. With 12 core layers and 4 inputs each, every later pass reads 48 messages per token.

A direct implementation would launch 4 message reads and 4 controllers, several kernels each, for every core layer in a later pass. vLLM instead reads all four messages of a layer in one attention-residual kernel and stacks the four controllers' weights, so their down and gain/forget projections run as batched GEMMs. One Triton kernel computes the controllers' RMS normalizations, and another fuses the gated merge with the RMSNorm of the sublayer the message feeds. The merge runs in FP32 without FMA contraction, matching the reference implementation's numerics.

For the layers themselves, vLLM reuses Inkling 2 Small's fused kernels — each sublayer's all-reduce, short convolution, residual update, attention-residual mix and RMSNorm fused into one kernel — generalized from hidden size 4,608 to 2,304, with a new mode that returns the FP32 mix before the RMSNorm so later passes can merge their messages into it.

Thinking Machines also released TML-d24-varloop-research, a checkpoint trained to run with a variable number of passes. vLLM exposes two options through --hf-overrides: loop_num_passes (at least 1, default 3) and loop_share_cache (all passes of a core layer share one KV cache, each overwriting the previous pass's keys and values; conv states stay per pass). On one GB200 at TP1 with 4K context and forward pass only, the measured decode-step times are 1.51 ms for 1 pass, 2.67 ms for 2 passes, and 3.85 ms for 3 passes at batch size 1; at batch 32, 3.81 ms, 5.92 ms and 8.11 ms respectively. KV cache capacity drops from 352K tokens at 1 pass to 175K at 3 passes. With 3 passes and a shared KV cache, decode time is essentially unchanged (3.86 ms at batch 1) but capacity recovers to 194K tokens — sharing frees less memory than dropping a pass because the per-pass conv-state caches remain.

Sources:vllm.ai

Research

Thinking Machines Lab Releases Inkling 2 Mini: A Looped MoE Model for Local AI

Material
2026-11-20 08:00 GMT+8

On November 20, 2026, Thinking Machines Lab released Inkling 2 Mini, a new multimodal Mixture-of-Experts (MoE) model designed specifically for local AI setups. The model features approximately 26 billion total parameters and 3 billion active parameters per token, supporting a context window of up to 1 million tokens and accepting both text and image inputs.

The key architectural innovation is a "looped" transformer design. While the model has 24 physical decoder layers, the middle 12 layers are visited three times in a loop during inference. This allows the model to execute 48 logical layer visits per token, effectively increasing its capacity without increasing the number of stored weights. Information moves between these passes through cross-pass messages, where each pass refines the computation of the previous one using learned gating controllers.

Inkling 2 Mini uses mixed-precision formats (NVFP4 and MXFP8 expert weights) and includes a DSpark drafter for speed-up. It supports day-zero integration with major frameworks including Transformers, Llama App (llama.cpp), MLX, and SGLang. A research variant with variable loop passes is also available, allowing users to trade off inference speed against generation quality by adjusting the number of loops.

According to vendor self-reported benchmarks, Inkling 2 Mini leads in 32 out of 48 tests compared to models like Gemma 4 31B, Qwen3.8-27B, and Nemotron 3.5 Super VL. It claims state-of-the-art performance in its class on the Artificial Analysis Intelligence Index v5. However, these results are based on internal testing and have not yet been independently reproduced.

Sources:huggingface.co

Research
Next →
Sign in / Sign up
ReadingResearchRadarInvestment framework

The accuracy claim is that vLLM keeps the reference implementation's numerics: same data types at every step, same NVFP4 and MXFP8 activation quantization, and FP32 message merges, with only reduction order differing. On a fixed prompt, hidden states at each of the 48 layer visits differ from the reference by about as much as the reference's own FlashAttention-4 and PyTorch attention paths differ from each other, and MoE outputs match bitwise given the same inputs. End-to-end accuracy was tested on needle-in-a-haystack up to 1M tokens, MMMU-Pro, HLE and GSM8K with DSpark enabled.

One boundary worth noting: the DSpark speculative decoder reads the target's hidden states at four fixed layer visits (one in the prelude, one each in the second and third passes, and the final output), so it requires the default 3-pass configuration and cannot be used with a variable pass count.