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
GGML and llama.cpp join HF to ensure the long-term progress of Local AI
Hugging Face announces that GGML, the creators of llama.cpp, are joining the company to support the ggml and llama.cpp community as local AI advances. Georgi Gerganov and his team will continue dedicating 100% of their time to maintaining llama.cpp, retaining autonomy and leadership over its technical direction and community, while the project remains fully open-source and community driven. The technical focus is tighter integration between llama.cpp’s local-inference stack and Transformers’ model definitions, aiming to make shipping new models from Transformers’ “source of truth” nearly single-click, alongside better packaging and user experience for ggml-based software. The longer-term goal is an efficient inference stack that runs on devices and helps make open-source superintelligence accessible, with llama.cpp becoming ubiquitous and easier for casual users to deploy and access.
Train AI models with Unsloth and Hugging Face Jobs for FREE
The post presents a workflow for fine-tuning the LiquidAI/LFM2.5-1.2B-Instruct small language model with Unsloth on Hugging Face Jobs, including optional use of Claude Code, Codex, or another coding agent. It explains that Unsloth is described as providing about 2x faster training and 60% lower VRAM use than standard methods, while the model runs under 1GB of memory and is optimized for on-device deployment. The tutorial covers claiming free Hugging Face Jobs credits through the Unsloth Jobs Explorers organization, installing the hf CLI or a model-training skill, and submitting a job with a dataset, GPU flavor, timeout, and output repository. An agent can generate a UV training script using 4-bit loading and LoRA, submit the managed cloud-GPU job, report Trackio monitoring details, and push the trained model to the Hugging Face Hub.
Scaling Localization with AI at Lyft
Lyft re-architected its batch localization pipeline to combine artificial intelligence with professional linguist oversight. The previous workflow relied exclusively on human translation, which resulted in multi-day turnarounds and linear cost scaling that hindered rapid market expansions. Under the new architecture, a dual-path system generates three candidate translations using a fast generation model, evaluates them against defined rubrics using a reasoning-focused model, and immediately releases approved drafts. Linguists review the drafts asynchronously in the Smartling translation management system to establish the final system of record. This pipeline reduced delivery latency from days to minutes while achieving a state where 95% of machine-generated translations require no significant changes during human review.
Minions: Stripe’s one-shot, end-to-end coding agents—Part 2
Stripe’s minions are an unattended agentic coding flow whose pull requests are human-reviewed but contain no human-written code; more than 1,300 merge weekly. The system runs agents in AWS EC2-based devboxes that are prewarmed to become usable within about 10 seconds, providing parallel, predictable, and isolated environments. Its custom harness uses code-defined blueprints that combine deterministic nodes, such as linting and pushing changes, with agent loops for tasks such as implementation and fixing CI failures. Minions receive automated feedback through local linting, one full CI iteration, and a possible second repair attempt, while devbox isolation and security controls limit destructive actions; Stripe presents the approach as an extension of long-running investments in human developer productivity.
IBM and UC Berkeley Diagnose Why Enterprise Agents Fail Using IT-Bench and MAST
IBM Research and UC Berkeley analyze how agentic LLM systems fail in enterprise IT automation, including incident triage, logs and metrics queries, and Kubernetes actions across long-horizon tool loops. To move beyond benchmark success rates, the analysis applies MAST (Multi-Agent System Failure Taxonomy) to execution traces, converting failures into structured signatures across system design, inter-agent misalignment, and task verification. It covers 310 ITBench SRE traces from Gemini-3-Flash, Kimi-K2, and GPT-OSS-120B, whose mean recalls were 75.5%, 28.6%, and 12.4%, respectively. Gemini averaged 2.6 failure modes per failed trace, compared with 4.7 for Kimi and 5.3 for GPT-OSS-120B, while FM-3.3, Incorrect Verification, was identified as the strongest predictor across models; the analysis recommends external verification, explicit termination controls, and clarification branches.
CX Score: How we built a metric support leaders can defend
CX Score was built to give support leaders a defensible view of customer experience across conversations, addressing whether an automated metric can be trusted. The team grounded its criteria in how experienced support professionals assess strong, mixed, and poor interactions, then created a dataset of thousands of conversations reviewed by experts, with two reviewers where possible and disagreement resolution. It operates as a production system using specialized prompt engineering, support-focused model configuration, neutral defaults for ambiguous evidence, and multi-stage evaluation across LLM models. Precision, recall, F1, regression, and bias checks set a required F1 above 0.8, while field testing with real customers examined varied tickets, languages, agent types, and conversation lengths. The stated conclusion is that CX Score is designed to be inspected, explained, and defended rather than treated as a black-box output.
OpenEnv in Practice: Evaluating Tool-Using Agents in Real-World Environments
OpenEnv is an open-source framework from Meta and Hugging Face for evaluating AI agents against real systems rather than simulations, with stateful environments that connect to tools and APIs. Turing's production-grade Calendar Gym applies this model to calendar operations, combining access control lists, partial visibility, temporal reasoning, and multi-step workflows in isolated sessions. Agents often handled individual actions but became less reliable on longer or ambiguous tasks: success was close to 90% with explicit calendar identifiers and roughly 40% when requests used natural-language descriptions. More than half of failures involved malformed arguments or incorrect action ordering despite selecting the correct tool, while structured validation, permission, and datetime errors provided concrete recovery signals.
Minions: Stripe’s one-shot, end-to-end coding agents
Stripe’s Minions are homegrown, fully unattended coding agents designed to complete tasks in a single run, beginning with an invocation such as a Slack thread and ending with a pull request ready for human review. More than 1,000 pull requests merged weekly at Stripe are minion-produced and contain no human-written code, although engineers review them and can provide follow-up instructions. Each run uses an isolated, pre-warmed devbox and a customized fork of Block’s coding agent goose, combining agent loops with deterministic operations for version control, linting, and testing. Minions use curated MCP tools, including Stripe’s Toolshed with more than 400 tools, to retrieve internal context before and during work. Local checks and selective CI provide feedback, with no more than two CI rounds, balancing correction against token, compute, and time costs.
Introducing SyGra Studio
SyGra 2.0.0 introduces Studio, an interactive environment for building, previewing, executing, and debugging synthetic data generation flows on a visual canvas. Users can configure connectors and models, preview source rows, reference automatically created state variables in prompts, define structured outputs, and inspect the generated YAML or JSON graph configuration. Studio streams node-level execution status, token usage, latency, cost, logs, and guardrail outcomes, while storing run history in .executions/runs/*.json and autosaving drafts. It also runs existing workflows, including a Glaive Code Assistant example with conditional revision loops, tunable dataset and batching inputs, and observable intermediate critiques; generated outputs can support training, evaluation, or annotation workflows.
The Future of the Global Open-Source AI Ecosystem: From DeepSeek to AI+
This third and final article in a series examines how Chinese AI organizations are shaping an open-source ecosystem after DeepSeek R1's January 2025 release. It argues that open sharing of models, papers, deployment infrastructure, and techniques has become the dominant near-term approach, with organizations ranging from Alibaba, Tencent, ByteDance, and Baidu to startups and application companies contributing at different layers. Qwen illustrates the ecosystem strategy: by mid-2025, more than 113,000 Hugging Face models used Qwen as a base, while organizations also connected models with chips, cloud platforms, toolchains, evaluation, and deployment systems. The article places this shift on infrastructure built since 2017, including the East Data, West Compute plan, and says R1 catalyzed wider use of existing compute, energy, and data capacity. It concludes that open-source models are becoming reusable, composable components in industrial workflows and localized AI systems as China moves from DeepSeek toward AI+.
We Got Claude to Build CUDA Kernels and teach open models!
Domain-specific coding tasks like writing optimized CUDA kernels often require capable, expensive frontier models, leading to high operational costs. The upskill tool addresses this challenge by capturing specialized workflows from teacher models such as Claude Opus 4.5 into portable skill files that can be consumed by smaller or locally hosted models. The framework automatically generates input-output test cases from execution traces to benchmark student model accuracy and token efficiency with and without the skill. In an evaluation targeting HuggingFace kernel-builder development, a local GLM-4.7 Flash model improved its pass rate from 40% to 85% when supplied with the generated skill. Measuring token usage across different models remains essential, as skills can lower token consumption for some architectures while introducing unnecessary overhead for others.
Architectural Choices in China's Open-Source AI Ecosystem: Building Beyond DeepSeek
Chinese open-source artificial intelligence development has shifted focus toward architectural choices, ecosystem toolchains, and domestic hardware integration following the January 2025 DeepSeek release. Leading groups have adopted Mixture-of-Experts architectures like Kimi K2 and Qwen3 to balance capability, dynamic compute allocation, and operating costs. Development expanded beyond text into multimodal domains including speech, video, and 3D systems while prioritizing smaller 0.5B to 30B distilled models under permissive Apache 2.0 licensing. Furthermore, model releases increasingly provide day-zero alignment with domestic inference and training infrastructure, including Huawei Ascend, Cambricon, and Baidu Kunlun chips. Production serving engines such as Moonshot AI's Mooncake and Baidu's FastDeploy 2.0 reflect a transition from standalone models to comprehensive system-level design.
Alyah ⭐️: Toward Robust Evaluation of Emirati Dialect Capabilities in Arabic LLMs
Standard Arabic language model evaluations predominantly assess Modern Standard Arabic while overlooking regional spoken dialects that feature distinct syntax, vocabulary, and cultural grounding. To address this evaluation gap, researchers created Alyah, a benchmark comprising 1,173 manually curated multiple-choice questions collected from native Emirati speakers across seven cultural and linguistic categories. Distractor answer choices were synthetically generated using language models and manually reviewed for semantic plausibility. Evaluating base and instruction-tuned Arabic-native, adapted, and multilingual models revealed that instruction-tuned systems consistently outperformed base variants, with falcon-h1-arabic-7b-instruct achieving the top instruction accuracy at 82.18%. However, questions involving implicit meanings, rare expressions, and specific dialect phenomena remained difficult across all tested model families and sizes.
Unlocking Agentic RL Training for GPT-OSS: A Practical Retrospective
Engineers evaluated the GPT-OSS model for agentic reinforcement learning across multi-step tasks such as ReTool and GSM8K using the verl framework. Initial training runs suffered from exploding gradient norms, increasing KL divergence, and stagnating rewards. Investigation revealed that non-deterministic expert routing in the Mixture of Experts architecture caused log-probability mismatches between forward passes, which falsely triggered Proximal Policy Optimization clipping. The team resolved this by substituting detached current log-probabilities for old log-probabilities during on-policy updates. To address further training instabilities and memory bottlenecks, the implementation integrated an attention-sink-aware backward pass into FlashAttention v3 and applied sequence parallelism using all-to-all communication around attention layers.
AI in Investment Management: 2026 Outlook (Part II)
Two Sigma leadership outlines the evolving role of artificial intelligence across quantitative investing workflows heading into 2026. The firm is embedding frontier large language models into internal systems, incident management, and feature generation pipelines to accelerate research tasks that previously took months into days. Technical focus across the broader field is shifting from raw parameter scaling toward efficiency optimizations, multimodal unified representations, and mechanistic interpretability circuits. In forecasting pipelines, rapid automated hypothesis generation introduces severe risks of overfitting and compromised backtesting, particularly when pre-trained models already contain historical regime knowledge prior to their cutoff dates. Consequently, engineering success requires strong institutional research discipline, production monitoring, and skepticism alongside the adoption of automated agentic tooling.
AssetOpsBench: Bridging the Gap Between AI Agent Benchmarks and Industrial Reality
AssetOpsBench is an evaluation framework designed to benchmark agentic AI systems within industrial Asset Lifecycle Management operations, such as chillers and air handling units. The benchmark incorporates 2.3 million sensor telemetry points, 4,200 work orders, 53 structured failure modes, and over 140 curated scenarios spanning four agent roles. Systems are assessed across six qualitative criteria, including task completion, retrieval accuracy, result verification, sequence correctness, clarity and justification, and hallucination rate. Evaluation traces are analyzed using a trajectory-level diagnostic pipeline called TrajFM, which extracts and clusters failure patterns. In community tests covering over 300 agents, no evaluated model family reached the 85-point readiness threshold.
One Year Since the “DeepSeek Moment”
The January 2025 release of DeepSeek's R1 model transformed China's artificial intelligence sector from a closed ecosystem into an organic open-source ecosystem. By releasing under the MIT license and sharing reasoning paths and post-training methods, DeepSeek turned complex reasoning into a reusable, downloadable engineering asset suitable for distillation in compute-constrained environments. Chinese technology organizations subsequently shifted from comparing standalone benchmark scores to developing ecosystem interfaces, application scenarios, and system-level infrastructure. Major players rapidly expanded their open footprints, driving downloads of newly created Chinese models on Hugging Face past those of any other nation. This momentum catalyzed global open-source adoption across international markets while prompting Western organizations to develop competing open-weight alternatives or build directly upon open Chinese foundations.
Differential Transformer V2
Differential Transformer V2 introduces an architectural refinement to differential attention that pairs and subtracts query heads within the same grouped-query attention group. The architecture doubles query heads while preserving key-value head counts, which avoids increasing key-value cache demands and removes the need for custom kernels. DIFF V2 replaces the per-head RMSNorm of its predecessor with a per-token, per-head projected lambda passed through a sigmoid function. This formulation expands the context root-mean-square range to between zero and root two, preventing massive gradient magnifications, mitigating attention sinks, and stabilizing large learning rate pretraining. Ablations confirm that sharing keys and values across subtracted heads and constraining lambda scaling are necessary to prevent training instability and elevated loss.
How Etsy Uses LLMs to Improve Search Relevance
Etsy’s Search Relevance team addresses a limitation of engagement-based search signals: popular listings can attract clicks even when they are not the best match for a query. It introduces semantic relevance, measured through a framework combining human-curated golden labels, an LLM annotator, cascaded models, and production integrations, with categories for relevant, partially relevant, and irrelevant query-listing pairs. The data pipeline uses sampled search-log pairs, dual human annotation, an evolving guideline, few-shot chain-of-thought prompting with o3 in LangGraph, and self-consistency sampling; a Qwen 3 VL 4B teacher then labels data at scale before distilling into a BERT-based two-tower student. The student supports real-time use with less than 10ms additional latency, while the framework evaluates and enhances search relevance alongside engagement signals.
MongoDB.local San Francisco 2026: Ship Production AI, Faster
MongoDB announced several features aimed at streamlining AI application development across its database and search ecosystem. The Voyage 4 embedding model family is now generally available, introducing cross-model compatibility within a shared embedding space alongside the video-capable voyage-multimodal-3.5 model. For in-database operations, MongoDB Community Edition introduced Automated Embedding in public preview to eliminate external synchronization pipelines for vector generation. Vector Search gained Lexical Prefilters, replacing the knnBeta operator with a dedicated vectorSearch operator and vector data type inside $search aggregations. Additionally, MongoDB made its intelligent assistant generally available in Compass and released the source code for its mongot search engine under the SSPL.
Related topics
All topics →Weekly digest
The best engineering reads, once a week.