News

Securing Quick access from corporate networks with AWS PrivateLink

AWS has published a walkthrough for connecting on-premises data sources to Amazon Quick through AWS PrivateLink. It describes how to keep analytics traffic off the public internet. The setup covers private user and API access alongside connections to internal data sources.

Private access follows different paths for sessions and data

The walkthrough assumes AWS Direct Connect or AWS Site-to-Site VPN already links the on-premises environment to a VPC. User sessions and API traffic use VPC endpoints, while data fetching uses VPC connections. Amazon Quick Sight queries PostgreSQL, MySQL, SQL Server and Oracle through network interfaces placed in the VPC. Quick Flows, Quick Automate and Quick Agents use VPC connections to invoke privately hosted MCP servers or internal APIs. Quick Index knowledge base VPC connections currently support only Amazon Redshift, excluding Confluence, SharePoint and Amazon S3. The walkthrough also covers DNS, identity federation and enforcing access only through VPC endpoints. Security groups and endpoint policies provide additional controls over those connections. (AWS Machine Learning)

thinkidiot take: Quick Index knowledge base VPC connections support only Amazon Redshift. I would check that source restriction before working through DNS and endpoint policies. A private connection does not make Confluence, SharePoint or Amazon S3 supported sources for this setup. For me, source compatibility is a deployment decision to address ahead of the network configuration.

How Specializing the Vocabulary of SLMs Can Cut Inference Cost by 77% in Real Applications

Authors writing on Hugging Face report a 77% reduction in measured inference cost with VocabX. They achieved it by specializing vocabulary without changing the base model. Their production Portuguese essay grader uses Qwen3-4B, a model with 4 billion parameters.

Every generated token starts with a large scoring task

The original vocabulary contains 151,669 entries. All of them are scored at every generated token. VocabX selects vocabulary around token cost rather than frequency. The approach retains the same data and training recipe. The grader receives an essay, a prompt and supporting texts. It returns a JSON assessment containing a grade and comment for each competency. The reported results come from a paper still under review and are subject to change before publication. (Hugging Face)

thinkidiot take: The reported 77% cost reduction came without changing the base model. I would test vocabulary specialization on a similarly constrained JSON task before replacing the model. Scoring 151,669 entries at every generated token gives that experiment a concrete target. I consider this a stronger reason to inspect vocabulary costs than to assume model size alone determines the inference bill.

Blackstone’s Jas Khaira says fears over AI-buildout debt are overblown

Blackstone’s Jas Khaira says fears over debt funding the AI buildout are overblown. The firm manages $1.3 trillion in assets and offers AI investments through debt and equity. In 2026, Blackstone and Coatue co-led a $10 billion credit facility for Australian GPU provider Firmus.

Customer contracts and pledged GPUs shape the lending

According to Khaira, CoreWeave approached Blackstone for a $2.3 billion loan in 2023. At the time, it had $30 million in revenue and minimal profitability. Blackstone’s neocloud financing structure considers committed customer contracts. It also considers GPUs pledged as collateral. The firm committed $5 billion to Crux AI, a neocloud launched with Google to commercialize its custom TPUs. Blackstone also launched its San Francisco AI hub, Blackstone N1, or BXN1, under Khaira in 2026. (TechCrunch)

thinkidiot take: CoreWeave sought a $2.3 billion loan with $30 million in revenue and minimal profitability, according to Khaira. I would start with the committed customer contracts and pledged GPUs when assessing that financing. Those are the concrete supports Blackstone identifies beyond the borrower’s revenue. I find them more useful for judging the deal than a broad dismissal of debt fears.

Data centers could erase decades of air quality gains

HARC modeled Texas air pollution under three projections for data center energy demand. If all planned projects connect to the anticipated grid mix, power-sector soot pollution would more than triple within five years. That scenario would put soot pollution at its highest level since 2006.

The electricity source changes the emissions outlook

In the same scenario, NOx emissions would rise 136%. Carbon dioxide emissions would rise 90% above 2025 levels. Combining grid power with renewable electricity generated onsite produces a different modeled outcome. Under that approach, NOx emissions would be up to 10 times lower. Texas has 26 gigawatts of existing and under-construction data center capacity. A March Wood Mackenzie report found that Texas had four times more proposed data center capacity using onsite power than any other state. (The Verge)

thinkidiot take: HARC’s grid-plus-onsite-renewables scenario puts NOx emissions up to 10 times lower. I would make the electricity source a deciding factor when evaluating data center capacity. The modeled difference gives me a concrete reason to ask how a facility gets its power. I would favor the lower-emissions supply arrangement over treating every gigawatt of capacity as equivalent.

Meta to create a new class of shares to fund AI push

Meta announced a new class of non-voting shares in its third-quarter earnings report. Trading is set to begin in 2027. The shares will fund AI initiatives, potential acquisitions and employee compensation. Meta also raised the lower end of its expected 2026 capital expenditures from $130 billion to $135 billion, with the upper end staying at $145 billion.

Muse adoption comes alongside a sharing rollback

Meta launched Muse in September. The AI agent can clean email inboxes, book restaurant reservations and plan trips. Its app reached 10 million US downloads during the week of the October 21 report. Meta released its Muse Wonder model during the same week. Researchers found Muse chats publicly available on the web. Meta then paused the agent’s conversation-sharing feature. (The Verge)

thinkidiot take: Meta paused Muse’s conversation-sharing feature after researchers found chats publicly available on the web. Before using it to clean an inbox or plan a trip, I would examine exactly how its sharing controls work. The app’s 10 million US downloads do not answer that practical question. For an agent handling those tasks, I put control over conversations ahead of adoption figures.

Trending AI Papers

Ranking source: Hugging Face Papers for 2026-10-22.

Agent World Model in the Wild: Computer-Use Agent Environment Synthesis

Teaching software agents on real computers creates practical hurdles and security risks. This paper builds artificial computer environments where agents can practise completing tasks. The aim is to make that training easier to expand while teaching abilities that carry over to unfamiliar tasks.

• Problem: Agents need experience using computers, but collecting that experience in real systems is hard to scale safely. Existing artificial environments usually support just one way of interacting, which limits practice that combines clicking, using tools and running commands.
• New idea: WildAWM is a system that uses open-source models to build working computer environments and tasks. Each environment supports graphical interfaces for visual interaction, model context protocol services for accessing tools, and command-line interfaces for entering commands. It checks tasks by having agents attempt them and comparing automated assessments of their results. It also tests whether those assessments hold up when relevant conditions change, aiming to keep tasks solvable, demanding and reliably checked.
• Simple example: Think of a computer training room built for practice. A learner can click through screens, call a tool or type commands to finish an assignment, while checks establish whether the assignment is possible and whether the result is correct.
• Evidence: The pipeline produced 341 environment-task pairs, with tasks that can take over 100 agent steps. GPT-6-Astra solved only 60% of them. Two Qwen models trained exclusively in these environments improved substantially across 8 benchmarks covering unfamiliar web, claw and computer-use tasks.
• Limitation: The abstract reports broad benchmark gains but does not give their sizes or establish how training affects security during use in real computer systems.
• Why it matters: Artificial practice environments could help agents learn computer tasks without making real systems their training ground.
• Paper: Agent World Model in the Wild: Computer-Use Agent

Cache Engineering for Agents

An agent working through a long task accumulates information from earlier steps. Keeping that information ready for use can become a major expense. This paper gives agents direct control over which stored calculations remain available, aiming to lower costs without repeatedly processing the same history.

• Problem: Long tasks make storing and moving cached calculations expensive, and their histories can outgrow the model's input limit. Cutting or summarizing earlier text invalidates later cached calculations, forcing the model to process retained text again. Keeping less text can reduce that work but hurt task accuracy.
• New idea: Cache engineering lets agents manage the key-value cache, which stores calculations the model uses to consult earlier text. The drop operation excludes selected stored calculations from future attention, the process that determines which earlier information the model uses. The repos operation adjusts position encodings, which represent where retained information sits in the sequence. The accompanying AKV serving engine, the software that runs model requests, tracks these changes so it can correctly reuse, remove and restore stored calculations.
• Simple example: Imagine working with a notebook and a desk covered in reference cards. Rewriting the notebook can force you to remake the cards; this approach instead lets you clear selected cards from the desk and adjust the remaining ones while leaving the notebook intact.
• Evidence: A policy that drops cached tool responses on a rolling basis raised average benchmark scores by 1.7 points and cut average cost by 27.7% compared with summarization triggered at a threshold. With AKV, it achieved up to 1.2x the output throughput and a 1.5x speedup in P95 request latency, the response-time measure at the 95th percentile. Repeated self-improvement of cache policies reduced cost by another 7.1% relative to expert design.
• Limitation: Correct reuse depends on the history of cache operations, so matching text alone is insufficient. The abstract does not identify the benchmarks or show how consistently the reported savings hold across different tasks.
• Why it matters: Better control of stored calculations could make long agent tasks cheaper while preserving useful information.
• Paper: Cache Engineering for Agents

Socratic Skill Learning: Interleaving Teaching and Practice for Transferable Agent Skills

An agent that succeeds once has not necessarily learned a method it can use again. This paper explores how teaching during practice can help agents write reusable instructions. The learner tries the advice itself and revises what it records. The aim is to produce instructions that remain useful when the task changes.

• Problem: A successful attempt may leave important questions unanswered about why a procedure works and when to use it. Turning that attempt into reusable instructions can preserve those gaps, leaving the agent poorly prepared for variations.
• New idea: Socratic Skill Learning alternates expert guidance with practice by a learner agent. An Educator, the teaching agent, directs investigation, while a Student, the learner agent, alone operates the task environment and writes the Skill, a reusable set of task instructions. The Student can question advice, test it and revise those instructions. An issue-guided version keeps track of unanswered questions and uses them to decide where further review is needed.
• Simple example: Think of learning a recipe with a teacher beside you. You do the cooking and write the instructions, while the teacher prompts you to test uncertain steps. The useful lesson is a recipe you have checked and revised for later use.
• Evidence: The researchers built SkillTransferBench with 35 task families and 408 executable cases. On the main cohort of 22 families and 155 cases, Fixed-Window SSL scored 0.801 utility, compared with 0.723 without a Skill, 0.761 for adapted Trace2Skill and 0.751 for adapted SKILL-KD. Guidance during practice beat expert-written instructions and teaching after practice across three model pairs on shared families. Issue-Guided SSL reached 0.805 utility while using 61.7% fewer Educator tokens and 16.9% fewer total learning tokens than Fixed-Window SSL; tokens are the text units processed by the models.
• Limitation: Reviews at fixed intervals become increasingly costly as the learner needs fewer new lessons. The main numerical comparison covers 22 task families and 155 cases, rather than the full benchmark.
• Why it matters: Teaching agents to test and revise their own instructions could make what they learn more useful on later tasks.
• Paper: Socratic Skill Learning: Interleaving Teaching and Practice

Trending AI Repositories

Ranking source: GitHub Trending.

vllm-project/vllm

vLLM is a Python project for serving language models. Its stated aim is to make that work easy, fast and cheap.

• What it is: It sits in the LLM serving layer, where models are made available for use.
• What it does: A high-throughput and memory-efficient inference and serving engine for LLMs
• Who it helps: It helps developers who need to serve language models. They can use vLLM to run inference.
• Limitation: The supplied excerpt gives no setup requirements or benchmarks to assess its speed and cost claims.
• Repository: vllm-project/vllm

supabase/supabase

Supabase is building Firebase features with open source tools. Its README lists a hosted database as an available feature, giving readers a concrete starting point.

• What it is: It is a TypeScript project in the application backend space. Hosted Postgres is part of its offering.
• What it does: The Postgres development platform. Supabase gives you a dedicated Postgres database to build your web, mobile, and AI applications.
• Who it helps: It helps developers building web, mobile and AI applications. They can use its hosted database for those projects.
• Limitation: The excerpt does not establish which Firebase features are available beyond the listed hosted database.
• Repository: supabase/supabase

twentyhq/twenty

Twenty is an open source CRM project. Its TypeScript codebase gives readers an open source option to examine in this software category.

• What it is: It belongs in the customer relationship management category. The README links to a website and documentation.
• What it does: The open alternative to Salesforce, designed for AI.
• Who it helps: It is relevant to teams looking for an open source CRM. They can follow the linked documentation to explore the project.
• Limitation: The supplied excerpt lists no CRM features or setup requirements.
• Repository: twentyhq/twenty