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
673 posts about LLMs. Every summary links to the original.
- 673 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
17,600 Actions: Agent Security Is a Systems Problem
Security incidents involving AI agents, such as the OpenAI and Hugging Face event where attackers executed 17,600 actions, demonstrate that system security cannot rely entirely on manual human review. Managing agent velocity requires dedicated system-level controls to properly constrain, observe, and govern their behaviors during execution. To secure autonomous development workflows, Docker Sandboxes allow AI agents to operate within GitHub Actions, run Testcontainers tests, adjust code, and automatically generate draft pull requests inside isolated environments. In addition, escalating software supply-chain threats and the increased volume of AI-generated code drive the adoption of Docker Hardened Images and Docker Extended Lifecycle Support. These updates incorporate software built from source into container images, extend security patch coverage past end-of-life milestones, and move organizational policy enforcement directly onto developer machines.
When it comes to Governance, Retailers need a control plane for context
Retailers are moving from isolated generative-AI pilots to enterprise-wide adoption, where each store, headquarters, and digital workflow needs different data, permissions, models, and cost controls. The post argues that useful AI depends on governed business context, while fragmented tools can produce separate provider contracts, logs, permission models, cost centers, and answers to the same question. It proposes an open control plane that centralizes governance and model access without forcing employees into one application or vendor. The described Databricks design combines Unity Catalog for data and business context, Unity AI Gateway for routing and oversight, Foundation Model APIs for commercial, open-source, and custom models, and Genie for natural-language queries over governed data; the intended result is faster experimentation and deployment with visibility into access, spending, lineage, and auditability.
Evaluating AI Agents Live at the Grounded Reasoning Cup
The Grounded Reasoning Cup evaluated whether AI agents could generalize grounded reasoning improvements from OfficeQA to a newly released, enterprise-style document benchmark. Eleven academic teams had about two months to build agents with a partner lab’s model family, then applied them live to OfficeQA Pro V2, based on U.S. Treasury Accounts of Receipts and Expenditures, after a 36-hour release window. Stanford won with 63.3% accuracy, while the top systems combined document preprocessing, targeted or hybrid retrieval, specialized tools, parallelism, and answer verification; 18.8% of questions remained unsolved by every team. The results show that benchmark gains did not reliably transfer, and that parsing, retrieval, tool use, verification, and operational infrastructure can matter as much as model choice in end-to-end enterprise reasoning.
Coding Agent Horror Stories: The Command You Already Approved
The OpenAI and Hugging Face security incident involved 17,600 attacker actions, demonstrating that AI agent security cannot rely exclusively on manual human review. Modern development workflows require systems controls to constrain, observe, and govern autonomous agents operating at high speeds. To address related isolation and automation requirements, developers can run AI agents inside GitHub Actions using Docker Sandboxes to execute Testcontainers tests, fix code, and generate draft pull requests. In addition, Docker Extended Lifecycle Support provides continued security coverage past software end of life while moving policy enforcement directly to developer workstations. Finally, Docker Hub has transitioned Docker Verified Publisher applications to a self-serve model, enabling publishers to distribute verified content directly to developers.
Consistency is the new latency: AI at the data layer
Autonomous AI agents depend on the accuracy of the underlying data layer because retrieved data populates the LLM context window for multi-step reasoning. In distributed architectures, standard asynchronous replication lag causes agents to ingest stale data, leading to flawed execution and self-reinforcing hallucination debt when errors are saved back to memory. To resolve this, architectures must align replication strategies with specific task truth requirements across three primary patterns. High-stakes workloads utilize Amazon Aurora with global write forwarding or native Aurora DSQL for multi-region strong consistency, while high-scale scenarios leverage DynamoDB Global Tables with conditional writes to prevent lost updates. Finally, high-velocity ingestion pipelines enforce LOCAL_QUORUM reads and writes in Amazon Keyspaces to capture critical data spikes reliably.
Agent Memory Inside the Harness
The post distinguishes agent state, the run-bound working data of a single execution, from agent memory, knowledge that outlives the run and is read across sessions. Context compaction, which decides what to keep when the window fills, is lossy by construction and sets both token cost and reliability; Chroma’s study found models become unreliable well before their advertised 200k-token window, and teams compact at 95% or proactively at 60%. System prompts and skills must be versioned, governed infrastructure, not text files, with prompt registries and the SKILL.md standard for procedural memory, and skills are executable so they require sandboxing. Enterprise data enters memory through pipelines like change data capture, not migrations, with chunking, embedding, and provenance stamps. The database choice caps filtered recall under selective scoping, the consistency window, and temporal modeling, setting the ceiling for later decisions.
DeepSeek V4 Pro 0813 vs GPT-5.6 Sol on DeepSWE: Cost, Coding, and Routing
DeepSeek V4 Pro 0813 costs $0.24 per rollout and achieves 62.8% pass@1 on DeepSWE, while GPT-5.6 Sol reaches 72.7% pass@1 at $8.37 per rollout—a 35x price gap. At four attempts, DeepSeek V4 Pro 0813 surpasses Sol, with 88.5% pass@4 versus 85.8%. Sol is faster and more reliable, completing a rollout in 17 minutes and 53 steps, and breaking tests in only 11% of its failures. A cascade strategy—running DeepSeek V4 Pro 0813 first and escalating to Sol on test failure—solves 83.0% of tasks at $3.35 each, outperforming either model alone. Per $100, DeepSeek V4 Pro 0813 solves 260 tasks versus Sol’s 9, making it the value leader for budget-constrained workloads.
How to Evaluate Production AI Agents: Measure System Outcomes, Not Conversations
Evaluating AI agents solely on conversational transcripts fails when agents must invoke tools to modify external production systems. Language models can convincingly narrate that an action occurred while completely omitting the underlying tool call, leaving database or billing records unchanged. Outcome-based evaluation addresses this divergence by treating external system mutations rather than generated dialogue as the actual source of truth. Frameworks such as CRMAgentBench implement persistent, stateful environments with strict all-or-nothing validation, verifying tool arguments, execution sequence, final record modifications, and the total absence of unauthorized actions. Furthermore, measuring repeated success across independent executions using pass^k metrics exposes operational variance and inconsistency that standard single-shot scores and pass@k metrics conceal.
The prototyping tax is killing your AI roadmap
The post argues that the “prototyping tax”—delay between an AI-powered idea and a working MVP—causes initiatives to lose momentum, often because organizational context is fragmented across teams, tools, and codebases. Traditional encapsulation and siloed domain knowledge can leave agents with syntax and contracts but not business intent, making production codebases harder to navigate than personal projects. The proposed response is agentic development grounded in business semantics and governed data, with alignment occurring through working builds while CI/CD, code review, and production rigor remain unchanged. The post recommends tracking time-to-prototype, first-pass acceptance rate, and PoC-to-production rate to distinguish faster demos from durable delivery. It cites a benchmark of 401 data tasks and Abacus Insights results: 77% accuracy at roughly half the cost for a platform-native agent, plus about 40% less manual effort and onboarding to first value in roughly half the time.
Developing Nemotron 3.5 Lightning NVFP4 with QAD Using NVIDIA Model Optimizer
NVIDIA’s post presents a workflow for developing an NVFP4 checkpoint of the Nemotron 3.5 Lightning 30B-A3B model, targeting lower memory use and higher inference throughput without sacrificing accuracy. The method starts with post-training quantization (PTQ), producing a W4A16 student from the BF16 teacher, then applies quantization-aware distillation (QAD), where simulated quantization and KL-divergence loss align student logits with the frozen teacher. For the selected four_over_six recipe, 32K calibration performed best, and the recipe using W4A16 Mamba linears offered the best reported trade-off between accuracy degradation and inference performance; the checkpoint is 22 GB versus 66 GB and delivers up to 4x faster throughput. The workflow includes Model Optimizer examples and launchers for Megatron-LM and Megatron-Bridge, covering calibration, distillation, evaluation, and export.
A/B test models in production
A/B testing model changes in production by sending fixed shares of live endpoint traffic to one control and up to 20 variants, allowing teams to compare user-facing outcomes rather than relying only on shadow-traffic checks. The endpoint router keeps variants out of the base traffic split, resamples requests that reach the control, and treats experiment percentages as absolute shares independent of replica counts or autoscaling. A 95/5 experiment can be updated to 80/20 and then 50/50 by resending the complete member list; an etag rejects stale concurrent updates. Measurement joins per-deployment platform metrics with product signals using the deployment ID in response metadata. To finish, a winning variant is promoted through a blue-green rollout before deleting the experiment, while deletion alone returns traffic to the control and can unwind a losing test.
Using AI_Functions in Your Data Warehouse: Top Use Cases
Databricks AI Functions bring model inference into SQL queries, addressing the need to combine warehouse-resident structured data with unstructured reviews, tickets, PDFs, and transcripts without exporting rows to a separate service. The post presents six Lakehouse use cases: document parsing and extraction, sentiment classification, translation, ticket routing, sales-call extraction, and generative drafting with ai_query. Task-specific functions such as ai_parse_document, ai_extract, ai_classify, and ai_translate handle specialized work, while ai_query sends prompts to an accessible Databricks-hosted Foundation Model serving endpoint. Unity Catalog permissions, existing pipelines, and system.billing.usage keep governance and billing within Databricks, while SQL execution manages planning, parallelization, and retries. Production guidance includes tagging jobs, sampling at least 10,000 rows, choosing models deliberately, requesting structured output, and versioning prompts.
No Dumb Questions: What is AI context architecture? Why not just build your own?
AI context architecture defines the boundaries, data access, memory, and fallback behavior that constrain AI agents and make their outputs more predictable. The discussion distinguishes architecture as system design, infrastructure as the storage and delivery of context, and engineering as implementation; MCP is treated as an architectural protocol, while RAG spans all three layers. Examples include restricting retrieval to car or sports-car tires, preserving agent history, and routing incomplete or low-trust knowledge to subject-matter experts through a human-in-the-loop flow. The buy-versus-build discussion notes that coding a custom architecture is possible, but buying can provide accumulated experience with data treatment, edge cases, and recurring operational problems, while good architecture also reduces token costs and aims for consistent, predictable answers.
How Cloudflare detects MCP traffic and helps secure it
Traditional security boundaries assume human speed and judgment, but AI agents execute nondeterministic actions rapidly and can invoke unapproved Model Context Protocol servers through standard HTTPS traffic. To mitigate this shadow traffic, Cloudflare One introduces detection and inspection capabilities across clients, network boundaries, and MCP servers. Cloudflare Gateway leverages protocol-level headers and JSON-RPC inspection to identify direct MCP connections, verify managed paths, and enforce access through approved MCP Server Portals. Furthermore, Cloudflare Agents SDK v0.20.0 implements the stateless 2026-07-28 MCP specification while providing fallback handshakes for legacy servers. Organizations can combine Gateway traffic source rules with server middleware like WriteGuard to block unauthorized tool invocations before execution.
State of Open Models: Summer 2026 Observations
The report analyzes Hugging Face Hub activity from January through August 2026 to assess how open-model releases, usage, licensing, and agent traffic are changing. Public repositories expanded from 2.43 to 2.96 million models, while 85.6% of models had fewer than 200 lifetime downloads and 1.5% of repositories captured 99.2% of downloads. Chinese labs reached monthly ceilings of 754B to 2.78T parameters, while broader portfolios such as Qwen’s generated far more downloads than frontier-only strategies; likes and downloads measured attention and infrastructure dependence differently. Large Chinese releases were often Apache 2.0 or MIT licensed, though newer trillion-parameter models introduced restrictions, while agents became the Hub’s leading user and autonomous-agent activity exposed the practical value of running quantized open models locally.
Reading the signals: Helping your AI-built payment integration understand business intent
AI agents can quickly produce Stripe payment flows that create a PaymentIntent, render the Payment Element, and pass an end-to-end demo, yet a correct implementation can still violate the business intent behind the money. The post argues that payment expertise and business decisions are often missing from integrations built from API references, causing silent failures such as terminal declines being retried, intents stuck in requires_action, uncaptured authorizations, restricted connected accounts, and asynchronous payments being refunded. It recommends defining business intent before configuring the PaymentIntent, then writing those rules into code comments, metadata, documented requirements, or custom objects. A verifier agent can check rules on every build, either through an agent skill, automated tests, or CI, while shared ownership remains with the human who decides what correct means.
What are AI Hallucinations?
AI hallucinations are coherent, confident outputs that are factually wrong, fabricated, or unsupported, appearing in chatbots, image generators, and multimodal systems. Generative models predict likely continuations rather than retrieve guaranteed facts, while training incentives, limited or flawed data, generation settings, and error compounding in reasoning models can increase the risk. Examples include Google Bard’s incorrect JWST claim, Air Canada’s nonexistent bereavement policy, Microsoft Sydney’s erratic responses, and fabricated legal citations that led to sanctions. Hallucinations create enterprise risks involving safety, liability, compliance, security, reputation, and trust; recommended controls include guardrails, source verification, adversarial testing, monitoring, human review, retrieval-augmented generation, governed data, and evaluation throughout development and deployment.
When Can LLMs Replace Humans in A/B Tests?
Replacing human participants in A/B testing with large language model predictions eliminates causal identification guarantees by design, forcing reliance on surrogate endpoint assumptions. An empirical evaluation using the Upworthy dataset revealed that raw gpt-4o-mini click-through rate predictions exhibited systematic attenuation bias, capturing only 39 percent of observed human treatment effects. To address this discrepancy, researchers formalized surrogacy and comparability conditions, finding that nonlinear machine learning calibrators like random forests and gradient-boosted trees successfully recovered treatment effects where linear ordinary least squares failed. In addition, averaging multiple sample draws per experimental unit mitigated variance and zero-bias caused by model generation temperature. Ultimately, because comparability cannot be verified for novel interventions that depart significantly from historical experiments, calibrated models cannot entirely replace human testing for genuine product innovations.
Smart Routing in Unity AI Gateway: Match frontier quality with 30%+ lower cost per task
Unity AI Gateway’s Smart Routing, now in Beta, addresses the cost and choice overload of assigning coding tasks to increasingly diverse models and harnesses. It uses task-aware routing: a small, low-latency model classifies the initial task description and metadata, then a policy selects a cheaper, medium, or frontier model; Omnigent can also choose the coding harness and route sub-agents. The design keeps consecutive turns on one model to preserve cache-hit rates, while escalating complex work and delegating simpler tasks to lower-cost options. On internal coding workloads, Smart Routing achieved 35% savings and outperformed any single model at 65% of Opus 5’s cost; public benchmarks reported matched Opus 5 performance with 56% savings. The team identifies mid-session reassessment, real-session evaluation, and affordable switching as continuing challenges.
Track generative AI costs with Amazon Bedrock inference profiles
When multiple departments share a single foundation model in Amazon Bedrock under one IAM role, usage aggregates into one billing line item, complicating department-level chargebacks and budget tracking. This issue can be resolved by deploying Amazon Bedrock application inference profiles wrapped around the shared foundation model and tagging each profile for cost allocation. The application layer determines the user department during authentication and passes the corresponding inference profile ARN as the model identifier. Once the designated cost allocation tag is activated in the AWS Billing and Cost Management console, usage separates into per-department line items in AWS Cost Explorer. This method incurs standard per-token model rates without extra charges for cost attribution.
Related topics
All topics →Weekly digest
The best engineering reads, once a week.