AI and infrastructure news
Filter by type: Post307
- 19 days ago x.com
Weight spectra may signal memorization in LLMs
A preliminary analysis published on September 24, 2026, found that the attention matrix K exponent α tracked memorization of random labels during LLM training.
- 19 days ago x.com
Attention factorization and gradient flow: implications for models
Published on September 30, 2026, the publication discusses how factoring the attention matrix as S = WWᵀ may change gradient flow and remove an information-exponent bottleneck in optimization.
- 19 days ago x.com
Recurrent flows refine reasoning on discrete tasks
A post published on September 3, 2026 describes models combining continuous flows and recurrent refinement to solve structured problems such as Sudoku.
- 19 days ago x.com
Fork Enables PCIe P2P on GeForce
A post published on August 27, 2026 reports a graphics-driver change enabling peer-to-peer PCIe transfers on GeForce cards while optimizing GPU serving.
- 19 days ago x.com
FuncAttn brings functional attention to continuous data
A June 25, 2026 post describes a method using structured linear operators for attention between function spaces, with results in PDE solving, 3D segmentation, and regression.
- 19 days ago x.com
Functional Attention proposes linear-cost attention
Published on July 6, 2026, the paper describes representing query and key/value fields with learned bases and a compact matrix, reporting linear cost in the number of points.
FUTEX_PRIVATE_FLAG and the cost of thread locks
A post dated June 28, 2026 explains how FUTEX_PRIVATE_FLAG may reduce lock latency for threads sharing memory within the same process.
- 19 days ago x.com
GEPA adapts a model with prompts and labeled examples
A demonstration published on September 19, 2026 describes using prompts and labeled examples to adapt a model to specific decision criteria.
- 19 days ago x.com
GLM-5.3-Flash: an architecture overview
Published on August 26, 2026, the overview describes a hybrid architecture with 3:1 attention, a smaller sparse MoE backbone, a four-stream mHC residual path, and a native vision encoder.
- 19 days ago x.com
Two-core GPU on a $30 FPGA: what to assess
Published on July 9, 2026, the article describes a two-core GPU implemented on a $30 FPGA, running AI model inference at 81 MHz.
- 19 days ago x.com
HydraHead combines two attention types per head
Published on June 30, 2026, the post about HydraHead describes an architecture that combines Full Attention and Linear Attention at the head level to pursue greater efficiency in long-context models.
Local AI: Mac Studio M5 Ultra or RTX PRO 6000?
A comparison of local AI configurations highlights differences in memory, CUDA support, prefill speed, cost, and noise.
- 19 days ago x.com
iLLaDA trains an 8-billion-parameter diffusion model
A paper published on June 27, 2026 describes training an 8-billion-parameter Transformer with bidirectional masked diffusion on 12 trillion tokens.
- 19 days ago x.com
Index compares speech models across three capabilities
Published on June 23, 2026, the Artificial Analysis index combines three evaluation datasets to compare native Speech-to-Speech models.
Prompt injection in AI-analyzed binaries
Research from the Naval Postgraduate School, published on July 12, 2026, describes how strings in C binaries can influence language-model reverse-engineering agents.
- 19 days ago x.com
Automated interpretability for attention heads
A paper published on June 29, 2026 describes a technique that explains attention heads with Python programs. Replacing about 40% of attention patterns in Llama-3B with program outputs had little effect on performance across tasks.
- 19 days ago x.com
IW-OPD reweights tokens in on-policy distillation
A post published on June 30, 2026 presents IW-OPD, a method that assigns greater weight to early tokens in on-policy distillation and reports a 6.9-point gain over standard OPD on AIME-2025.
- 19 days ago x.com
KV Cache: Why Queries Are Not Stored
A post dated August 30, 2026 distinguishes the temporary use of queries from the storage of keys and values in the KV cache during inference.
- 19 days ago x.com
GPU latency hiding: measure before optimizing
A chapter by Vasily Volkov of UC Berkeley, published July 4, 2026, on how GPUs hide latency and architecture concepts related to performance.
- 19 days ago x.com
Local audio cleanup on Raspberry Pi 5: how to assess the use case
A post dated June 22, 2026 presents a 49,000-parameter model for acoustic echo cancellation, noise suppression, and dereverberation, with claimed performance of 21× real time on one Raspberry Pi 5 core.
- 19 days ago x.com
Memory Attention: reusing keys in V=K+M
A publication dated September 27, 2026 reports preliminary ablation results suggesting V=K+M is an effective architectural choice; validation at larger scales remains necessary.
Meta describes techniques to speed up storage
Published on July 5, 2026, Meta's post presents a storage architecture and performance techniques, including data locality, unified caches, and prefetching.
- 19 days ago x.com
Flow models refine reasoning in stages
A paper dated September 17, 2026 turns a flow model into a recurrent model: predictions are fed back for iterative refinement, with a local loss at each step and no backpropagation through time.
- 19 days ago x.com
NextLat explores predicting the next latent state
A June 16, 2026 publication describes a transformer training method and says it may enable compact world models and up to 3.3× faster inference.
- 19 days ago x.com
NoRA proposes normalizing A at LoRA initialization
A post published on September 2, 2026 describes an initialization that normalizes LoRA's A columns and reports higher average results than several adaptation methods in SFT and RLVR.
- 19 days ago x.com
New OCR models expand document-processing options
A post published on September 18, 2026 presents WeVisDoc 2B and 4B under Apache 2.0, and jina-ocr-v1 under a non-commercial license. It reports a 95.38 score for WeVisDoc-4B on OmniDocBench v1.6 and identifies NaviDC-OCR as the current SOTA.
- 19 days ago x.com
NVIDIA presents diarization for overlapping speech
Announcement dated September 23, 2026: Nemotron 3 Diarization identifies who spoke and when, including in overlapping speech, and supports up to eight speakers.
- 19 days ago x.com
What happens when a CUDA kernel is launched
A post published on July 11, 2026, about the CPU-to-GPU communication needed to launch a CUDA kernel and the work involved in that operation.
- 19 days ago x.com
Ostris AI Toolkit adds LoRA training for MiniMax H3
On August 3, 2026, the Ostris AI Toolkit repository announced LoRA training support for MiniMax H3, limited to text-to-video and image-to-video workflows, using quantized Comfy weights.
- 19 days ago x.com
Packed FlashAttention speeds up a Transformer vocoder
A July 20, 2026 publication reports a 48.64% increase in QPS for a Transformer vocoder after waveform reconstruction optimizations.
Pandora’s Router balances estimation and routing costs
A paper published on August 22, 2026 proposes using cheap, noisy estimates before paying for better evaluations. On MATH, RAG, and EmbedLLM, it achieved or tied for the lowest average combined inspection cost and routing regret.
Reverse-proxy phishing can bypass MFA
A post dated August 9, 2026, describes how phishing kits can relay logins and capture session cookies after multifactor authentication.
Device-code phishing is not limited to Wi-Fi
Published on July 25, 2026, the post warns that device-code authentication phishing can also occur over Ethernet or the internet.
- 19 days ago x.com
Phonon-1: evaluating local audio transcription
A post dated August 28, 2026 presents Phonon-1, a speech recognition model with 782 million parameters, a 415 MB download, and an Apache 2.0 license.
- 19 days ago x.com
Model pruning uses effective sample size to select components
A post published on July 1, 2026 presents Effective Model Pruning, a method that uses effective sample size to calculate how many components to retain in neural networks and language models.
- 19 days ago x.com
Policy gradients: from derivation to practical evaluation
Published on August 30, 2026, the article reviews foundational reinforcement learning work, derives the policy gradient algorithm, and discusses its variants.
- 19 days ago x.com
Post points to efficient designs in frontier Chinese models
Published on August 26, 2026, the post says frontier Chinese models are adopting linear or sparse attention, specialized residual designs, and Muon. It also presents open-source collaboration as a strategy for efficiency.
- 19 days ago x.com
Post describes calibration-guided low-bit compression of Hy4
A post published on August 29, 2026, says Tencent reduced the Hy4 preview from 1.5 TB to 200 GB by setting precision per layer. It reports results on two benchmarks.
- 19 days ago x.com
Synthesized programs could replace some attention heads
A post published July 1, 2026 describes Python programs synthesized from attention maps, reporting 69–79% mean IoU and little QA decline when replacing up to 30–40% of heads.
- 19 days ago x.com
ProRL Agent separates rollouts from LLM policy training
Published on September 30, 2026, ProRL Agent describes open-source rollout infrastructure that separates trajectory generation, environment orchestration, and evaluation from policy optimization for multi-turn agents.
Let's assess your project
I reply with scope and a proposal.