Activation Steering: One Dial, Two Costs
Activation steering edits a model's internal activations to add or remove a concept at inference time. This explainer maps its one control knob, its two costs, and when it beats prompting.
Loading…
Topic
LLMs
648 posts about LLMs. Every summary links to the original.
- 648 posts
- 49 companies
- 5 explainers
See the idea
LLMs explainers
Antibodies as a Prediction Pipeline: Where AI Models Break Down
Antibody discovery is a three-stage prediction pipeline, and AI models that excel on familiar targets degrade sharply on unfamiliar ones. This explainer maps where that gap appears and what it costs.
Agent Memory: What Outlives the Run
Agent memory is the knowledge that survives a run, not the working state inside one. This explainer separates the two, then maps the write path, the read path, and the database limits that cap recall.
The Good Stuff
Paid picks we use ourselves. Links carry a referral code — a subscription through one earns us a commission at no extra cost to you.
Newest first
Posts
NVIDIA Vera CPU: Olympus Cores Built for Maximum Single-Thread Performance in Agentic AI
Agentic AI workloads shift critical execution paths onto CPUs through sandboxed environments that execute code, invoke tools, and manage runtime states. The NVIDIA Vera CPU addresses these compute requirements via Olympus cores built to maximize single-thread performance and instructions per cycle under full socket load. To handle irregular control flows and pointer-heavy structures, Olympus combines a ten-wide decode engine, neural branch prediction supporting two taken branches per cycle, dependency-breaking acceleration, and dedicated graph prefetchers. Across scale-up systems, dual-socket Vera configurations leverage second-generation NVLink-C2C interconnects to maintain a clean two-NUMA-node topology alongside PCIe 6.4, CXL 3.1, and Arm CCA confidential computing. These architectural choices provide predictable latency and memory bandwidth across highly concurrent reinforcement learning and agent loops.
Scaling document classification to 100k+ labels
Large-taxonomy classification must map freeform text to 35,000–100,000 labels while meeting quality, cost, and throughput requirements. The benchmark compares vector search, a two-step vector-search-plus-AI-Classify workflow, and direct frontier-model calls across transaction, company, and biomedical datasets, using hybrid semantic and BM25 retrieval with Reciprocal Rank Fusion for the search stages. The combined workflow performed best, reaching 0.81 average accuracy versus 0.76 for Gemini 3.5 Flash, while costing roughly one-hundredth as much per document; its strongest average setting shortlisted the top 20 labels. Vector search alone was nearly 100 times cheaper than the workflow but scored more than 20 points lower, while full-taxonomy frontier calls could exceed context limits. Taxonomy updates can be handled by embedding and ingesting changed labels rather than retraining or redeploying classifiers.
Online Learning for Cost-Efficient LLM Routing
Ramp describes an internal LLM gateway that processes trillions of tokens daily and uses dynamic failure-aware routing to reduce AI spend. It combines an EWMA of genuine provider-side failure rates with Thompson Sampling over a Normal-Inverse-Gamma posterior for the mean and variance of log-latency. For each request, the router estimates the probability of outright failure or missing the caller's deadline, incorporates model cost, and orders options while preserving caller-defined model preferences and fallbacks. In a reranker with a six-second deadline, the strategy reacted to Gemini 3.1 Flash Lite rate limiting by switching to GPT 5.4 Nano Flex, whose latency matched Standard at half the price. Larger experiments reported more than 25% cost savings without degradation, while a streaming-equivalent deployment for Ramp Inspect produced 30% savings.
NVIDIA NVLink: The Scale-Up Network for AI Factories
Scale-up networking has emerged as a critical architectural component for AI factories supporting trillion-parameter models, mixture-of-experts architectures, and disaggregated serving. NVIDIA presents its sixth-generation NVLink scale-up fabric and NVLink 6 Switch to address latency-sensitive all-to-all GPU communication and offload collective operations. In simulation benchmarks on models like DeepSeek-R1, Qwen 235B, and a simulated 2T parameter model, NVLink achieved up to 2.3X higher decode throughput than off-the-shelf Ethernet in a 72-accelerator domain. The architecture integrates NVLink-C2C to supply 1.8 TB/s of coherent CPU-GPU bandwidth in the Vera Rubin NVL72 platform, as well as NVLink Fusion for integrating semi-custom silicon XPUs. Through co-designed hardware, SHARP in-network compute, and NCCL libraries, the platform optimizes token throughput, factory uptime, and deployment economics.
How our universal content processing platform Riviera evolved for AI and beyond
Dropbox developed Riviera, a centralized content processing platform that evolved from an internal file preview generator into a shared infrastructure supporting over 300 file formats. By decomposing complex conversion tasks into modular, reusable transformations, the system now powers search indexing, media transcoding, and AI context extraction across Dropbox products.
Coding Agent Horror Stories: The Agent That Deleted Production
Recent industry updates focus on AI agent security, container maintenance, and developer workflows within Docker ecosystems. The OpenAI and Hugging Face incident involved 17,600 attacker actions, demonstrating that AI agent security requires automated system controls to constrain, observe, and govern agents rather than relying on human review. In software maintenance, Docker Extended Lifecycle Support (ELS) addresses escalating supply-chain risks and MinIO end of life by maintaining patches, building software from source, and enforcing local developer policies. For continuous integration, Docker Sandboxes enable isolated AI agents to run Testcontainers tests, implement code fixes, and submit draft pull requests within GitHub Actions. Additionally, Docker Verified Publisher applications have transitioned to a self-serve process accessible directly through Docker Hub.
In-House LLM Serving at Netflix
Netflix established an in-house serving platform to run large language model inference directly inside existing production environments alongside traditional machine learning models. Built upon NVIDIA Triton Inference Server and vLLM, the unified architecture handles member-scale routing, candidate generation, feature fetching, and model execution over gRPC and OpenAI-compatible HTTP endpoints. The team shifted their primary engine from TensorRT-LLM to vLLM to support custom architectures, simpler debugging, and non-trivial constraint logic. Addressing production obstacles required patching Triton's frontend for guided decoding, pinning dependent library versions, rewriting logits processors in C++ for vLLM V1, and handling state machine resets during engine preemptions. The resulting platform unifies deployment flows while preserving operational stability across evolving model schemas.
The cost of saying yes has changed
Generating initial code patches has become significantly cheaper with AI agents, shifting the primary expense of small feature requests from writing code to debating scope in meetings. Teams can use agent-generated patches as diagnostic probes rather than final deliverables, turning abstract scope debates into concrete artifacts that reveal true system touchpoints and risks. However, low generation costs do not translate to low ownership costs, as changes touching areas like authorization, compliance, or public contracts still demand significant human review. Constrained attempts allow engineers to price uncertainty quickly and shift scope discipline from pre-implementation speculation to evidence-based code review.
DE Dost: Teaching Machines to Be a Friend on the Last Mile
DE Dost is a GenAI onboarding companion embedded in Swiggy’s delivery-executive app, designed for new DEs who receive information but lack confidence-building, immediate guidance. It supports questions in nine languages across SOPs, earnings, payouts, app navigation, shifts, and support, adapting detail and tone to conversational context. Before the first LLM call, a graph state assembles user context, preprocessing resolves references, and embedding-based semantic routing separates general knowledge-base questions from personal diagnostics. Retrieval uses authored QnA pairs plus hybrid semantic and lexical scoring to handle paraphrases and terms such as “MG” and “COD,” while guardrails constrain follow-ups to available knowledge. Evaluation spans about 800 cases per language across fleet and payout combinations, with macro accuracy trending at 90%+ and a target P50 latency of 4.5 seconds after reducing an initial 15-second pilot.
The skills gap behind agentic AI — and how Databricks is closing it with a new context engineer certification and agent trainings
Databricks introduced the Databricks Certified Context Engineer Associate beta exam to validate skills for building reliable, production-grade AI agent systems as organizations scale agentic AI. Context engineering is presented as the practice of curating, maintaining, and filtering tokens, memory banks, and tool parameters so an LLM can solve a task; the certification is intended to benchmark that technical skill set. Databricks also added AI Agent Fundamentals, Building Retrieval Agents on Databricks, and Agent Evaluation, covering agent reasoning, retrieval-augmented architectures, and systematic performance testing and improvement. Its AI-first certification-prep guide is available on every certification page, works with free tiers of major LLMs, includes guardrails and disclaimers about hallucinations, and shows candidates how to use Free Edition for hands-on practice; registration for the Context Engineer Associate is open, with the first exam scheduled for July 29, 2026.
Integrating Context-Aware Video AI Agents Into Enterprise Workflows
Video analytics AI agents often struggle to integrate into enterprise workflows because video platforms, organizational knowledge bases, and operational systems are isolated. To resolve this, NVIDIA NemoClaw orchestrates the Metropolis Video Search and Summarization blueprint alongside the Retrieval-Augmented Generation blueprint into an automated pipeline. Upfront human-in-the-loop interactions collect analysis parameters, objects of interest, and relevant retrieval queries before video ingestion begins. The system combines hierarchical video summaries with contextual enterprise reference documents to generate structured reports with timestamps and citations. Finally, NemoClaw executes programmatic downstream tasks, such as creating prioritized Jira tickets or dispatching industrial maintenance work orders directly into management tools.
Scaling Agentic AI Factories Through Extreme Co-Design with NVIDIA BlueField
Agentic artificial intelligence workflows transform data center infrastructure by turning context management, security enforcement, and inter-service communication into active components of the inference data path. To prevent host central processing units and graphics processing units from stalling on these tasks, NVIDIA introduces BlueField-4 data processing units, Vera BlueField-4 STX storage processors, and the DOCA software platform. The BlueField-4 processor offloads Layer 3 routing, encryption, and telemetry using a sixty-four-core Grace processor with up to eight hundred gigabits per second connectivity. Simultaneously, the Vera BlueField-4 STX powers the NVIDIA CMX context memory platform alongside DOCA Memos to store, protect, and share key-value cache across flash storage tiers. This co-designed infrastructure isolates tenant traffic and accelerates context reuse without degrading host-side agent execution.
AI Agents Explained: How to Build with Them Safely
AI agents combine reasoning models, external tools, memory, orchestration, and execution environments to autonomously pursue goals across iterative perceive-reason-act loops. While simple chatbots only generate text responses to direct prompts, autonomous agents execute multi-step operations such as querying databases, modifying codebases, and executing shell commands without requiring continuous manual approvals. However, this unchecked operational autonomy significantly widens the blast radius when models encounter logic errors or vague instructions, risking destructive filesystem operations and credential leaks. Safe deployment requires shifting focus from model-level guardrails to infrastructure-level containment strategies, such as scoping tool access and isolating execution inside disposable microVM sandboxes. Controlling networking boundaries, file access, and credential exposure ensures teams achieve automated execution velocity without exposing developer host environments to uncontained systemic risk.
Your agent should understand what you see
Sentry's Seer Agent initially parsed user screens using DOM-scraped character-by-character ASCII grids, leading to severe token consumption, model context rot, and an inability to support interactive actions. To resolve these bottlenecks, Sentry replaced visual text scraping with a composable semantic context system where React UI components directly declare their state, metadata, and interactive affordances. A root React context provider maintains a registry of component nodes organized hierarchically through higher-order wrappers, serializing the resulting structure to JSON and backend-generated Markdown within the system prompt. Across twenty-four live pages, this approach reduced dashboard context tokens from roughly 5,500 to 1,300 while maintaining identical satisfaction rates and tool-call volumes across more than 14,000 evaluated conversations.
Security incident disclosure — July 2026
An intrusion into part of production infrastructure was driven end to end by an autonomous AI agent system, accessing a limited set of internal datasets and several credentials used by services. The attack began through two code-execution paths in dataset processing, then escalated from a worker to node-level access and moved laterally across internal clusters while running thousands of actions in short-lived sandboxes. AI-assisted anomaly detection surfaced the compromise, and LLM-driven analysis of more than 17,000 attacker events reconstructed the timeline, identified indicators, mapped credentials, and separated impact from decoys. Hosted frontier models initially blocked forensic requests containing attack artifacts, so the investigation used zai-org/GLM-5.2 on internal infrastructure. The initial vulnerabilities were closed, affected infrastructure rebuilt, credentials rotated, and cluster controls and alerting strengthened; assessment of possible partner or customer impact remains ongoing.
July 2026 newsletter
The July 2026 newsletter rounds up ClickHouse releases, community projects, benchmarks, engineering work, and upcoming events in real-time data warehousing. Version 26.6 adds hypothetical skip indexes, cascading refreshable materialized-view dependencies, experimental continuous queries, AI embedding functions, PNG output, CLI documentation, and geospatial improvements. A CostBench comparison running the full analytics path over 28 hours found better cost-performance for ClickHouse Cloud than Snowflake Interactive Tables, which required a continuously running refresh warehouse for pre-aggregated data. Other features cover conversational analytics with ClickHouse Cloud, LibreChat, and MCP; Postgres change-data capture through Debezium and Kafka; local agent queries with chDB; high-concurrency tuning; and Silk, whose fiber runtime targets lower I/O tail latency. Quick reads report ingestion, caching, security, and multi-tenant SQL developments.
Develop Lightweight USD Runtimes Faster with AI Agents
Building custom Universal Scene Description (USD) implementations traditionally required adapting large existing codebases to meet specific memory footprints, application binary interfaces (ABIs), or performance requirements. The nanousd-labs project introduces an alternative approach by using AI agents to generate lightweight USD runtimes directly from the Alliance for OpenUSD machine-readable USD Core Specification. Agents parse the specification section by section, write conforming code, and validate output against specification-derived tests under engineer guidance. Written in C++ with a public C API, nanousd functions as an independent data layer that handles parsing, composition, queries, and writing without rendering pixels. This methodology enables developers to regenerate runtimes for varying deployment constraints while maintaining standard compliance through reusable workflows and skill graphs.
Inkling model from Thinking Machines Lab now on Databricks
Databricks announces that Inkling, Thinking Machines Lab’s first open-weights model, is available to enterprise customers through the Unity AI Gateway. The model is positioned for coding and agentic reasoning workflows, supports multi-modal inputs, and can be applied to enterprise data, including proprietary codebases, internal documentation, and domain-specific data. Unity AI Gateway provides centralized security, permissions, audit logging, policy enforcement, cost controls, budgets, and observability, while data remains within the governed environment; Inkling is invoked through a REST API, with SQL query support planned. Teams can try it in AI Playground, deploy a governed endpoint, connect coding agents such as Cursor, OpenCode, or Pi, and build agents with Agent Bricks, with open weights enabling customization and inference-cost optimization without per-token API pricing.
Model Routing Is Simple. Until It Isn’t.
Model routing in agentic systems is presented as a systems optimization problem rather than a simple classification task. On 417 AppWorld Test Challenge tasks with the same CodeAct agent, Claude Sonnet 4.6 cost $79 total ($0.19 per task), versus $155 ($0.37 per task) for GPT-4.1, despite GPT-4.1 having lower listed token prices; cache-read pricing and repeated context changed the effective result. The discussion also shows why task difficulty and model speed are incomplete signals: hidden execution work, specialization, governance constraints, routing overhead, cache state, hardware, and endpoint load all affect decisions. The described router optimizes cost, quality, and latency together, producing configurable cost-accuracy operating points; one latency-oriented configuration reached 84% accuracy for $93 and 83 seconds, while adding about 6 ms and 2 kB per task.
Exploring Hierarchical Interest Representation For Meta Ads Deep Funnel Optimization
Hierarchical Interest Representation is an upstream representation layer for Meta Ads that addresses sparse deep-funnel feedback by connecting users’ inferred interests with advertisers, products, services, and other ads entities. It models a typed, heterogeneous engagement graph with a transformer architecture, sparse attention for long-range relationships, bias-aware attention, self-supervised cross-view distillation, and multimodal content features processed through vision or language models. The system projects the raw graph into learned interest primitives and cascading super-graphs, producing universal embeddings and Bag-of-Meaning tokens at multiple granularities. Temporal cutoffs and chronological splits prevent future-information leakage, while a pipelined online graph engine reportedly achieved a 30x wall-clock speedup over a synchronous baseline during testing. Trained end-to-end on billions of interactions, the representations are designed for personalization, retrieval, supervision, and ranking integrations including GEM, Andromeda, and the Adaptive Ranking Model.
Related topics
All topics →Weekly digest
The best engineering reads, once a week.