Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Lockheed Martin and the U.S. Navy demonstrated the SensorMAX AI/ML sonar system on two MH-60R helicopters during RIMPAC 2026 off Hawaii, showing that ground operators could label unfamiliar acoustic signatures, retrain m…
A 2024 study of nearly 9 million U.S. death certificates found that taxi and ambulance drivers had the lowest risk of dying from Alzheimer's disease among 443 occupations, with about 1 in 100 dying from it compared with …
Google DeepMind retrofitted Gemma 4 into a text diffusion model, DiffusionGemma, using less than 10 percent of the original training budget, and it generates 256 tokens in parallel at about 1,500 tokens per second. Quali…
DeepMind's WeatherNext model achieves a breakthrough in forecasting cyclones, reducing prediction errors by 20% compared to NOAA's operational baseline and cutting false alarms by 30%. This enables earlier evacuations an…
DeepSeek's V4 Flash 0731 model scored 82.7% (±1.79 SE) on Terminal-Bench 2.1, ranking second overall behind Anthropic's Claude Code with Fable 5 at 83.8% (±1.16 SE), according to Ante's public benchmark harness. The mode…
A tutorial by an unnamed author demonstrates an end-to-end sentiment analysis workflow on the Stanford NLP IMDb Large Movie Review Dataset, comparing a TF-IDF and Logistic Regression baseline with a DistilBERT model fine…
Hugging Face's LeRobot documentation recommends at least 50 demonstration episodes, roughly 10 per location, for training an ACT model, but the original ACT/ALOHA results show success rates ranging from 64% to 96% across…
Jane Street's three-player auction game MegaGem was used to train a 4B specialist model, with SFT raising Qwen3-4B-Instruct's benchmark rating from 551 to 1158, ahead of Claude Sonnet 4.6, Claude Opus 4.8, and GPT-5.5, a…
Makerfabs released the MaTouch ESP32-S3 MaUWB development board, which combines an ESP32-S3 wireless MCU, a 3.95-inch touchscreen, and a UWB module for indoor positioning and ranging. The board uses a trilateration algor…
A 3.16M-parameter transformer running on a $250 FPGA achieves 44 tokens per second by moving the CPU out of the per-token loop, according to a technical blog post detailing the on-chip LLM design. The hardware uses a tra…
A new compiler called Torchwright, created by researcher Rob, compiles fixed computation graphs into transformer weights, and its first dedicated digit-level multiplier achieves 100% accuracy on all 3,000,000 valid three…
An agent that passes 75 out of 100 test runs has a 75% pass@1 rate, but reliability depends on the product: pass@k (at least one success in k tries) yields 98.4% for k=3, while pass^k (all k succeed) yields only 42.2%, a…
Dwarkesh Patel has been advocating for continual learning in AI, arguing that frozen models are limited and that continual learning is necessary for long-running, open-ended problem solving. The author agrees with Patel'…
Teddy Warner's Avian Visitors project uses a Raspberry Pi and USB microphone with Cornell's BirdNET deep learning classifier to identify and track bird species by sound, displaying recent visitors as a collage on a web i…
Pmalfa31 developed a tomato grading system using an ESP32 and machine learning that sorts standard and cherry tomatoes by color and size, with a web simulator available for testing. The system uses two learned datasets, …
Intel's Core 5 320 'Wildcat Lake' processor, built on Intel 18A in Intel's own fabs, achieved 6.21 Gflops per watt in Jeff Geerling's HPL Linpack testing, outperforming Apple's A18 Pro in the $599 MacBook Neo at 5.38 Gfl…
Google DeepMind open-sourced its WeatherNext cyclone model on August 6, publishing a Nature paper and releasing code, weights, and ensemble forecast feeds, with three-day forecasts matching the accuracy of previous two-d…
Researchers at Stanford University and the Arc Institute used the Evo 1 and Evo 2 genome models to design variants of the ΦX174 bacteriophage, and of 285 synthesized candidates, 16 produced viable phages that infected an…
At YC Paper Club's robotics edition, researchers presented findings on embodied memory, self-supervised reasoning, dexterous sim-to-real tool use, and world action models, while Francois Chaubard noted that despite a dec…
A mechanistic interpretability study found that ablating specific MLP neurons in Google's Gemma 4 model causes it to revert to its dormant Bard identity, with the model answering 'My name is Bard' with p=0.9923 after neu…