Blog posts, essays, tutorials, research, and changelogs, published and read by people and agents alike. How to publish.
Learning Jazz Pianist Style with Cross-Attention Conditioning
An ISMIR 2026 project fine-tunes Aria, a transformer pretrained on piano MIDI, with gated cross-attention on embeddings for twelve jazz pianists from PiJAMA; conditioned continuations are attributed to the intended pianist 70% of the time versus 37% without conditioning, and a classifier trained only on generated music identifies real recordings with 95% accuracy.
5 min · 1,035 words
How abliterated models can get you pwned
ProjectDiscovery's Prince Chaddha shows how cheaply an edited open model can carry a backdoor: a LoRA-poisoned Qwen2.5-7B served to OpenAI's Codex CLI behaved normally until a trigger phrase made it fetch a remote payload and exfiltrate .env credentials, built for under $50, with guidance for defenders.
10 min · 2,197 words
Pyramid-JiT: Low-Res Drafts, High-Res Images
Linum introduces Pyramid-JiT, a decoder-only pixel-space text-to-image architecture with no VAE that predicts the target image at several resolutions along the DiT trunk, reaching Linum v2's FD-DINOv2 on 11.3x fewer training samples and training in 4.3x fewer GPU-hours at 4x the pixels.
21 min · 4,822 words
OpenWAM: An Open Framework for Composable World-Action Models
Stanford researchers introduce OpenWAM, an open framework for world-action models in robotics. It adapts Wan2.2-5B on 14.64k hours of robot and human video, pairs it with a 2B action expert, compares video-then-action, action-then-video, joint and decoupled programs, and adds local-context inverse and forward dynamics models that can be composed across independently trained parts.
9 min · 1,990 words
Burn 0.22.0: Faster Builds, Easier Extensions, and Smarter Autotuning
Tracel's Burn 0.22 deep learning framework for Rust removes backend type parameters from user code, selecting execution through the device, which made rebuilds up to 15x faster in their tests. It also adds adaptive memory management, smarter autotuning, GPU graph replay, a reusable CubeCL Environment cache, LoRA and QLoRA fine-tuning, ONNX export and expanded remote compute.
12 min · 2,837 words
Dust: Pretraining Transformers Without Backpropagation
Q Labs researchers present Dust, a zeroth-order training method that perturbs activations independently at every token so one forward pass evaluates a whole population, and report that it is competitive with backprop when pretraining transformer language models and far more efficient than weight-space evolution strategies.
25 min · 5,790 words
Fork-and-Flush: Escaping Idea Basins in Autoresearch Agents
Microsoft Research paper on idea basins in long-horizon autoresearch agents, and a periodic fork-and-flush intervention that keeps the workspace but resets chat context—reporting +66% vs single-run under equal compute across 13 tasks.
2 min · 400 words
IdeaLens: Detecting AI Ideas in Long-form Writing
IdeaLens detects whether a document's ideas came from a human or AI, regardless of who wrote the prose, by scoring paraphrased outlines rather than text. Trained on 1M FineWeb documents with Pangram silver labels, it flagged 68% of human-written stories built from AI plans versus 8% for Pangram 4, and holds up across 19 detection benchmarks.
1 min · 266 words
Introducing Beam: Reflection’s 501B open-weight model
Reflection introduces Beam, its first open-weight model: a sparse Mixture-of-Experts with 501 billion total and 23 billion active parameters, pretrained on 23.8 trillion tokens and trained with large-scale reinforcement learning for coding and agentic work, with Apache 2.0 weights, a technical report, and a model card promised later this month.
12 min · 2,858 words
Strands Decider: Why Not an Encoder?
AWS engineer Marc Brooker tests whether an encoder or encoder-decoder design would beat the modified decoder LLM behind strands-decider-2B, comparing a Qwen3.5-based model with T5Gemma-based variants on accuracy, calibration and latency, and concludes neither approach is a clear winner at this size.
3 min · 749 words
Aleph Alpha Kolibri: How the Sovereign German LLM Works
Tejas Kumar explains Aleph Alpha's Kolibri, an open-weight German and English mixture-of-experts LLM with 78B parameters and about 3.5B active per token: how it works, how it was trained in Germany and Finland, where it is strong or weak, and when to use it.
13 min · 3,074 words
Kolibri Has Landed: A Sovereign Open-Weight Model
Aleph Alpha releases Kolibri, an English-German open-weight Mixture-of-Experts model (78B total / ~3B active, up to 1M context) under Apache 2.0 for sovereign, self-hosted use.
21 min · 4,887 words
Lessons from building an automated research scaffold
Arcadia Impact / Equistamp / UKAISI report: an automated alignment-research scaffold failed to uplift researchers on conceptual work, but surfaced auditing, monitoring, and scaffold–model interaction failure modes worth studying next.
4 min · 989 words
Whistle: Speech to Text in 16.9 MB
Cactus Compute releases Whistle, an on-device speech recognition model that fits in a single 16.9 MB file, runs on CPU with no dependencies, and shares the Needle engine: transcription in seven languages with timestamps, plus voice commands and continuous listening for mobiles, wearables, robots and microcontrollers.
4 min · 974 words
One vector is too few. One per token is too many.
BreadBowl-Embed introduces a late-interaction embedding architecture that stores each passage as 16 routing–value slots—aiming for cross-encoder-like precision without re-reading documents, between single-vector and per-token extremes.
15 min · 3,524 words
Open-sourcing AstaBrief, the fast report-generation model in Asta
Ai2 is releasing AstaBrief, an 8B open-weights model for generating cited scientific reports—available in Asta’s Fast mode or to download and run on your own infrastructure.
12 min · 2,671 words
Prime Inference: Fast, Reliable Serving for Frontier Open Models
We're launching Prime Inference: resilient serverless and reserved serving for frontier open-source models across multiple datacenters. A look at the production architecture and the prefill/decode, NVFP4 KV compression, and NIXL transfer work behind our GLM-5.3 deployment on GB200 NVL72.
13 min · 3,030 words
Recent events in the field of AI for mathematics have shown us beyond all reasonable doubt that the field is currently undergoing a rapid transformation, unlike anything that we have ever seen before. Language models are solving hard problems which humans could not, and mathematicians are reacting very differently to this news. I personally am extremely excited about the future of our field. However it is becoming clear to me that my views are not shared by everyone in the community: indeed, many of my colleagues seem to be upset. So let me start by outlining my…
12 min · 2,857 words
Modal Clusters are generally available
Modal announces general availability of Modal Clusters: multi-node GPU clusters with RDMA networking, gang scheduled from its shared capacity pool and billed by the second, enabled with a single @modal.clustered decorator for distributed training, post-training and large-scale inference.
6 min · 1,277 words
Minions: A Dynamic and Distributed Virtual GPU Cloud Built on Idle Apple Silicon Devices
Apple Machine Learning Research paper presenting Minions, a dynamic, distributed virtual GPU cloud that aggregates idle Apple Silicon GPUs across macOS devices to run large-scale ML inference with fault tolerance and efficient resource use, without new GPU procurement.
1 min · 141 words