Loading…
Topic
619 posts about Machine Learning. Every summary links to the original.
See the idea
A walkthrough of how a 3D CT model turns a voxel volume into a structured diagnostic report: isotropic resampling, 3D patch tokens, spatial encoding, and language-model reasoning.
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
Anthropic and the Australian government signed a Memorandum of Understanding to cooperate on AI safety research and support Australia’s National AI Plan. The agreement centers on collaboration with Australia’s AI Safety Institute, including sharing findings about emerging model capabilities and risks, participating in joint safety and security evaluations, and conducting research with Australian academic institutions. Anthropic will also share Anthropic Economic Index data to track AI adoption, economic effects, and workforce implications, initially covering natural resources, agriculture, healthcare, and financial services. Separately, AUD$3 million in Claude API credits will support four Australian institutions working on clinical genomics, precision medicine, pediatric research, and computing education. The company is also exploring data-center and energy investments, launching a startup credit program, and preparing to open a Sydney office.
Orbax and MaxText introduce continuous checkpointing to balance training reliability and performance when hardware failures or preemption events interrupt long-running jobs. Fixed checkpoint intervals create a trade-off: infrequent saves risk substantial lost work, while frequent saves can block or bottleneck training, particularly over unstable networks. MaxText enables asynchronous checkpointing that attempts a new save only after the previous background request is fulfilled; its continuous mode overrides checkpoint_period, and max_num_checkpoints_to_keep can limit storage use. A llama-3.1-70B benchmark on two v5p-128 slices found smaller P50 checkpoint intervals, but average training step time increased because of more frequent device-to-host transfers; the analysis estimated resource conservation against mean-time-between-failure assumptions. Orbax also supports minimum intervals and customizable preservation policies, while co-locating storage with training is recommended because cross-metro bandwidth can reduce checkpointing speed and reliability.
TRL v1.0 marks a transition from research codebase to a library intended for production systems, with clearer stability expectations. It implements more than 75 post-training methods, but its design prioritizes trying, comparing, and using them over coverage alone. The library responds to a field in which PPO, preference optimization methods such as DPO, ORPO, and KTO, and RLVR-style approaches such as GRPO repeatedly changed which components of the training stack appear essential. TRL addresses that volatility with a stable core under semantic versioning, an experimental surface for faster-moving methods, and deliberately limited abstractions that favor explicit implementations over generic hierarchies. v1.0 also follows distributed breaking changes across the 0.x releases, so migration from the last 0.x version is described as minimal, while future work targets asynchronous GRPO, scaling, and actionable training diagnostics.
Lyft requires real-time predictions of whether a rider will request a ride after viewing price and ETA details to balance supply and demand and personalize user experiences. Traditional machine learning models overfit on high-cardinality, sparse context intersections, whereas deep neural networks introduce unacceptable inference latency during live user sessions. To address this, engineers created a hierarchical Bayesian tree framework that organizes session data into increasingly granular context partitions. Each node houses a parametric model trained top-down, applying Gaussian Bayesian priors and L2 penalties relative to parent parameters to smooth predictions in data-rare child segments. Using simple parametric models at each node also permits monotonic constraints, guaranteeing logically consistent predictions aligned with domain expectations.
Composer 2 is a coding model for agentic software engineering, and its technical report describes training from continued pretraining of the open Kimi K2.5 base model through large-scale reinforcement learning. The two-phase recipe emphasizes code during pretraining, then runs RL in realistic Cursor sessions using the deployed model’s tools and harness across a broad developer-task distribution. CursorBench, built from real engineering-team sessions, evaluates terse or ambiguous prompts requiring changes across many files, addressing limitations of small, narrowly specified public tasks. Composer 2 scores 61.3 on CursorBench, 37% above Composer 1.5, plus 73.7 on SWE-bench Multilingual and 61.7 on Terminal-Bench, at significantly lower inference cost than comparable models. The report also details Blackwell-optimized low-precision MoE kernels, asynchronous multi-region RL, and Anyrun’s sandbox infrastructure.
Fin announces Apex 1.0, a custom model trained by its AI Group for customer-service conversations, replacing frontier-lab models as the core answering model for nearly all English-language chat and email traffic. The post says Apex improves resolution, speed, hallucination rates, and cost, citing one gaming customer’s resolution rate rising from 68% to 75% overnight and describing a 22% reduction in unresolved conversations. Its development relies on proprietary evaluations and billions of human and agent interaction data points generated by Fin’s resolution engine, creating a training flywheel for further domain-specific improvement. The launch is presented as evidence that customer-service applications may need to become full-stack AI companies rather than relying on shared general-purpose models. It also argues that frontier labs face pressure to offer cheaper specialized models as post-training and task-specific data become more important.
EVA is an end-to-end framework for evaluating conversational voice agents across task accuracy and spoken conversational experience. It simulates multi-turn live-audio conversations with a user simulator, the evaluated voice agent, deterministic tool execution, validation, and metrics derived from recordings, transcripts, and tool-call logs. The framework reports EVA-A for Accuracy and EVA-X for Experience, combines code-based and LLM- or LALM-as-Judge metrics, and uses pass@3 and pass^3 across three trials to capture peak performance and consistency. Its initial release contains 50 reproducible English-language airline scenarios and 15 tools covering rebooking, itinerary changes, cancellations, standby, and vouchers. Results for 20 cascade and audio-native systems show a consistent tradeoff: agents stronger at task completion tend to provide worse experiences, while a substantial gap between pass@3 and pass^3 indicates inconsistent success.
The tutorial presents a pipeline for adapting a general-purpose embedding model when it misses fine-grained distinctions in domain documents and labeled query–document data is unavailable. It generates synthetic question–answer pairs with NeMo Data Designer, filters them for quality, mines hard negatives with a 95% positive-score margin ceiling, and unrolls multi-hop queries for contrastive training. The recipe fine-tunes Llama-Nemotron-Embed-1B-v2, evaluates base and adapted models with BEIR metrics, and exports the result to ONNX or TensorRT. On the reported datasets, it improved Recall@10 and NDCG@10 by more than 10%; Atlassian increased Recall@60 from 0.751 to 0.951 on JIRA data. The complete workflow can run in roughly 2–3 hours for about 500 documents, using a single 80GB Ampere-or-newer GPU.
Etsy’s Ads Search ranking model needed a stronger purchase-intent signal because clicks are noisy and purchases are sparse later in the buyer journey. It retained its shared feature-representation and interaction layers, then replaced the shared feed-forward network with Multigate Mixture of Experts (MMoE), using task-specific gates and towers for click-through rate (CTR) and post-click conversion rate (PCCVR). Add-to-cart was added as an auxiliary task, while favorites were rejected after producing negative transfer. Offline, the resulting model increased Purchase and Click PR AUCs by average 3.5% and 1%, respectively, and Purchase ROC AUC by 0.5%. Online, it improved purchases and purchase calibration, supplied better inputs to auto-bidding, kept model size flat through pruning, and reduced inference cost.
Dropbox Dash optimized its LLM-as-a-judge relevance scoring system using DSPy to migrate from expensive proprietary models to cheaper open-weight alternatives. By establishing automated feedback loops based on human agreement and strict JSON format validation, the team systematically generated robust prompts for new models. This reduced human-score disagreement by 45% and slashed model adaptation time from weeks to days while enabling 10x to 100x more data labeling.
The post examines how open source AI on Hugging Face changed over the past year across competition, geography, technical trends, and emerging communities. It combines Hugging Face and community-contributed metrics to assess users, models, datasets, downloads, repositories, and specialized sub-ecosystems. In 2025, the platform reached 13 million users, more than 2 million public models, and over 500,000 public datasets, while the top 200 models represented 49.6% of downloads and China surpassed the U.S. in monthly and overall downloads. The analysis also describes rising participation by independent developers, increased Chinese releases, and active communities in robotics and science, framing open source AI as overlapping ecosystems rather than a uniform market.
Holotron-12B is a multimodal computer-use model released by H Company and available on Hugging Face. Post-trained from NVIDIA’s Nemotron-Nano-12B-v2-VL-BF16 on proprietary data, it targets agents that must perceive, decide, and act in interactive environments while handling long contexts with multiple images. Its hybrid State-Space Model and attention architecture stores a constant state per layer per generated sequence instead of attention’s per-token key and value activations, reducing memory demands for long-context inference. In a WebVoyager evaluation on one H100 with vLLM, it reached 8.9k tokens/s at concurrency 100 versus 5.1k for Holo2-8B, and its benchmark score rose from 35.1% to 80.5% after training on localization and navigation data.
Hugging Face introduces Storage Buckets as a mutable, S3-like object storage layer on the Hub for intermediate machine-learning artifacts such as checkpoints, optimizer states, processed shards, logs, and traces. Unlike versioned Models and Datasets repositories, Buckets support frequent writes, overwrites, directory synchronization, and cleanup while retaining standard Hub permissions and public or private visibility. Built on Xet, they chunk and deduplicate content across files, reducing repeated bytes, bandwidth, transfer time, and, for Enterprise customers, billed storage. The CLI and Python APIs support creation, syncing, dry runs, plans, inspection, individual file operations, and programmatic workflows; HfFileSystem and fsspec also allow compatible libraries to read and write through hf:// paths. Pre-warming can place hot data near AWS or GCP compute, while roadmap work aims to connect Buckets directly with versioned repos.
At-scale reinforcement-learning training is constrained by long, variable data-generation workloads that leave training GPUs idle, especially for reasoning and agentic models. The surveyed architecture separates inference and training onto distinct GPU pools, connects them through a rollout buffer, and synchronizes weights asynchronously so both stages can proceed concurrently. Sixteen open-source libraries are compared across orchestration, buffering, weight synchronization, staleness, partial rollouts, LoRA, and distributed backends; Ray appears in 8 of 16 orchestration implementations, while NCCL broadcast is the default weight-transfer method. The proposed TRL async trainer favors a lightweight bounded queue with per-token model_version metadata, packed NCCL transfers, and experimental support for prefix-resume or abort-and-retry handling of rollouts interrupted by weight updates.
Ulysses Sequence Parallelism addresses the memory and compute pressure of training language models on sequences extending beyond tens of thousands of tokens, where attention remains quadratic in sequence length even with FlashAttention. It shards tokens across GPUs, projects local queries, keys, and values, then uses two all-to-all operations to exchange data so each GPU computes a subset of attention heads before restoring sequence-sharded outputs. The post describes integrations with Accelerate, the Transformers Trainer, and TRL's SFTTrainer, including variable-length handling and weighted loss aggregation for uneven valid-token counts. In reported four-GPU tests, SP=4 reduced per-GPU memory sufficiently for 96K-token training on H100 80GB GPUs, reached 13,396 tokens per second at 64K, and matched non-SP canonical token-normalized loss within 0.000005 maximum absolute difference.
LeRobot v0.5.0 is a major release that expands supported hardware, policies, datasets, simulation, and its software foundation, following more than 200 merged PRs and more than 50 new contributors since v0.4.0. It adds full Unitree G1 humanoid support for locomotion, manipulation, teleoperation, and whole-body control, alongside OpenArm, Earth Rover, CAN-bus motor controllers, and other integrations. The policy additions include Pi0-FAST's autoregressive VLAs with FAST action tokenization, Real-Time Chunking for continuously blended predictions, Wall-X, X-VLA, SARM, and PEFT fine-tuning. For data and simulation workflows, streaming video encoding removes waits between recording episodes, while EnvHub loads environments from the Hugging Face Hub and IsaacLab-Arena supplies GPU-accelerated simulation. The codebase now requires Python 3.12, uses Transformers v5, supports installable third-party policy plugins, and adds remote Rerun visualization plus installation and documentation improvements.
Thumbtack software engineer Kenza Boulisfane develops an artificial intelligence tool called the Marketing Analytics Agent to democratize access to complex cross-channel marketing data. Non-technical staff across Data Science, Growth, Brand, and Product can ask natural language performance queries without requiring SQL expertise or manual reporting. A primary technical challenge involves unifying fragmented data across multiple advertising platforms that define metrics differently while maintaining data reliability. The system interprets natural language queries, analyzes underlying performance metrics, and translates results into actionable, data-backed recommendations for next steps. Boulisfane also notes that virtual work environments enhance collaboration by encouraging clearer documentation, intentional communication, and focused individual productivity.
Deploying Vision–Language–Action policies on embedded robots is constrained by compute, memory, power, and real-time control requirements; synchronous inference can leave the arm idle and cause oscillatory corrections. NXP presents dataset-recording practices for a tea-bag placement task, including fixed cameras, controlled lighting, a gripper-mounted view, calibration backups, and diverse training and validation splits with recovery episodes. The example uses 120 episodes, three 640×480 cameras at 30 fps, batch size 8, and checkpoint selection based on training and validation success rather than loss alone. For deployment, the VLA is decomposed into vision, LLM backbone, and action-expert stages, with selective quantization and asynchronous scheduling constrained by inference time being shorter than action execution. On the i.MX 95, optimized ACT latency reached 0.32 seconds with 0.89 global accuracy, while ONNX FP32 ACT reached 0.96 global accuracy at 2.86 seconds.
Modular Diffusers introduces a composable alternative to Diffusers’ existing DiffusionPipeline for building diffusion workflows from reusable blocks. A ModularPipeline can load a workflow definition, load model components separately with settings such as bfloat16, and preserve the familiar inference API, while blocks for text encoding, image encoding, denoising, and decoding can be inspected, run independently, or rearranged. Custom Python blocks declare components, inputs, outputs, and computation logic; the example adds a Depth Anything V2 depth processor to a Qwen Image ControlNet workflow, with data flowing between compatible blocks. Blocks and complete pipelines can be published to the Hub, and Mellon can generate node interfaces from their consistent definitions. The post presents the system as an early, flexible direction and says Mellon is not ready for production use.
PRX Part 3 reports a 24-hour speedrun for training a text-to-image diffusion model, combining techniques that performed well in earlier experiments under a budget of about $1,500 on 32 H200 GPUs. The recipe uses x-prediction in pixel space with 32-pixel patches and a 256-dimensional initial bottleneck, followed by LPIPS and DINOv2 perceptual losses, TREAD token routing, REPA alignment with DINOv3, and Muon for matrix parameters while Adam handles the rest. Training ran for 100,000 steps at 512px with batch size 1,024, then 20,000 steps at 1024px with batch size 512 and no REPA, across three synthetic datasets. The resulting model is described as usable, with strong prompt following and consistent aesthetics, though texture glitches, anatomy issues, and difficulty with hard prompts remain. The post attributes these limitations mainly to undertraining and limited data diversity and presents the recipe as a foundation for further scaling.
The best engineering reads, once a week.
