AI and infrastructure news
Filter by type: Post307
AI R&D agents need to revisit training strategy
A study published on August 22, 2026, analyzed 1,338 post-training trajectories. The overall strategy changed in 74 of 3,557 adjacent experiments.
- 19 days ago x.com
Training terminal agents in Dockerized RL environments
A post published on June 24, 2026 describes 14,600 Dockerized reinforcement-learning environments and an outcome-only DPPO recipe. A 9B model reached 27% on Terminal-Bench 2.0.
Agents that update their model of the environment during inference
A post published on June 30, 2026, about ARC-AGI-3 says that learning during inference leads agents to build models of rules and mechanics, updating them with new evidence.
- 19 days ago x.com
ARGUS diagnoses slowdowns in large-scale training
A paper published on arXiv on June 19, 2026 describes a system combining CPU, framework, and GPU data to investigate training slowdowns, with continuous overhead below 2%.
- 19 days ago x.com
Hybrid architectures beyond standard attention
A post published on July 8, 2026, compares the architectures of Qwen 3.5 and Nemotron 3 and describes linear attention and state-space models as examples of convergent evolution.
- 19 days ago x.com
Sparse attention can optimize KV block selection
A Tencent publication dated September 14, 2026 describes a method that uses language modeling loss to learn which KV blocks each query should access.
- 19 days ago x.com
Dynamic linear attention adapts memory to token information
A post published on June 15, 2026, describes Dynamic Linear Attention adapting memory states to token information while keeping a fixed cache, and reports gains in benchmarks, throughput, and memory use.
- 19 days ago x.com
Linear and sparse attention: evaluating long-context models
A post dated August 26, 2026, claims GLM-5.3 uses three linear-attention layers for each sparse-attention layer, highlighting compute-efficiency trade-offs for long contexts.
Autodata structures agents to create and evaluate datasets
Published on June 25, 2026, the material presents Autodata, a staged method in which AI agents create and analyze training and evaluation data, then apply meta-optimization.
- 19 days ago x.com
AutoDecompiler applies RL to iterative decompilation
A post dated June 25, 2026 describes a model using reinforcement learning and direct feedback across multiple stages of decompilation.
AutoMem explores agent-managed memory
A post published on July 6, 2026 describes a paper on agents that write, search, and organize notes, with reported 2x to 4x gains in long games.
Autoresearch agent beats a manual harness after eight days
A post published on July 14, 2026, reports that an autoresearch agent improved over eight days and outperformed a manually tuned harness on benchmarks excluded from training.
- 19 days ago x.com
BITCOS reduces storage for ternary LLMs
In a post published on September 20, 2026, Intel describes BITCOS, a method that stores exact ternary LLM weights at 1.485 bits per weight and reports decode throughput gains.
- 19 days ago x.com
Breeze TTS 2: Multilingual voice generation in streaming
A publication dated August 25, 2026 describes a speech synthesis model supporting 50 languages, text-prompt voice generation, and streaming output.
Local hybrid search for agents: what zg offers
On September 2, 2026, Alibaba's Zvec team announced zg, a local search tool combining semantic search, BM25, hybrid search, and rg, with support for popular agents.
C++26 standardizes Hazard Pointers for concurrent code
Published on June 30, 2026, the post says C++26 standardizes Hazard Pointers to defer node deletion until concurrent readers have finished using them.
- 19 days ago x.com
A 64-token cache for linear attention: what to evaluate
A post about HOLA, published on July 5, 2026, describes a parameter-free token-surprise metric and a 64-token cache to address catastrophic forgetting in linear-attention models.
- 19 days ago x.com
CodeMidas creates RL training tasks from code
In a post dated September 21, 2026, Xiaomi describes CodeMidas, which uses agents to generate 5,545 executable tasks in 23 languages from open-source codebases.
Iterative code improves scientific forecasts
A study published in Nature assessed combinations of forecasting methods across six fields: 24 of 55 combinations outperformed both methods from which they originated.
Evaluating rewards for coding agents
A July 4, 2026 post describes four reward signals for coding agents and discusses failures and corrections that can help detect reward hacking.
- 19 days ago x.com
How to assess KV cache reductions
A post reports an alleged reduction in KV cache use from 389,000 to 890 bytes per token over three years. The claim points to an opportunity to measure memory and cost in long-context systems.
- 19 days ago x.com
How to measure RL gains on software engineering tasks
A post dated September 21, 2026 reports that MiMo Pro's DeepSWE score rose from 58.41 to 72.57 after reinforcement learning runs.
- 19 days ago x.com
How RL Could Expand Skills Learned Through SFT
A post published on July 11, 2026, presents the hypothesis that RL can separate useful SFT components into reusable skills and routing rules.
- 19 days ago x.com
How vLLM schedules requests and allocates the KV cache
The September 26, 2026 article traces a request through vLLM and explains per-iteration scheduling and KV cache block allocation.
How to organize and evaluate compiler optimizations
A 1971 paper by Frances Allen catalogs compiler optimization transformations and serves as an introductory reference for engineers evaluating passes in their toolchains.
- 19 days ago x.com
Making model training more observable
On August 21, 2026, a publication about Marin described training data and metrics visualizations that can support model inspection.
Validating results generated by research agents
An arXiv paper published on September 1, 2026, reports that checking manuscript claims against execution logs reduced the reported rate of serious errors to 4%.
Crescent Island listed with 32 Xe3P cores and 480 GB LP5X
A publication dated August 24, 2026, lists Intel Crescent Island with 32 Xe3P cores, 480 GB of LP5X memory, and 350 W power consumption.
- 19 days ago x.com
CUDA C++: GPU execution, memory and correctness
Published on September 19, 2026, NVIDIA's CUDA C++ programming guide covers kernels, threads, memory, synchronization, errors, and validating GPU results against the CPU.
Dell Precision T7910: assessing a used multi-GPU platform
A publication dated July 17, 2026, lists the T7910 configuration and highlights PCIe slots, ECC memory, and power limits as factors in assessing multi-GPU computing.
- 19 days ago x.com
On-policy distillation combines specialist models
A post published on October 6, 2026 describes MOPD, a method that trains a student model on its own rollouts with token-level supervision from domain specialists.
- 19 days ago x.com
Token-level distillation blends teacher and model supervision
A post dated July 8, 2026, describes a paper on dual on-policy distillation, which dynamically selects a supervision source for each token.
- 19 days ago x.com
Speaker diarization: evaluating label correction
A Google article published on October 4, 2026 describes a 4-billion-parameter model fine-tuned to correct speaker labels and reduce word-level diarization errors across four benchmarks.
- 19 days ago x.com
Diffusing Blame proposes training networks without backpropagation
A post published on July 17, 2026 describes an error-signal routing method and reports results in image recognition, locomotion, and Craftax.
- 19 days ago x.com
Model distillation: the direction of per-token KL
A systematic study published on October 3, 2026, reports that the direction of per-token KL divergence matters more than rollout policy in on-policy versus off-policy distillation.
- 19 days ago x.com
On-policy distillation for consolidating specialist models
MOPD trains a student model on its own completions and uses token-level probabilities from multiple teachers to transfer specialist capabilities.
A shell prompt mismatch increased latency in DSH
A post published on August 17, 2026 reports that different prompts in two DSH shell tools triggered a 3.5-second timeout; a change reduced one command's time from 3,600 ms to 158 ms.
- 19 days ago x.com
From REINFORCE to GRPO: policy gradient methods for LLMs
An article published on September 24, 2026, presents a progression of policy gradient methods used to train LLMs, from vanilla policy gradient and REINFORCE to PPO, GRPO, and GRPO variants.
- 19 days ago x.com
Dust explores gradient-free transformer training
Published on October 5, 2026, the paper presents Dust, a zeroth-order method using activation perturbations, and reports scaling experiments and comparisons with backpropagation.
ESP32 blocks 537,000 domains with about 50 KB of RAM
A post published on July 20, 2026 describes ESP32 firmware that stores its blocklist in flash and takes about 10 milliseconds to block requests.
Let's assess your project
I reply with scope and a proposal.