AI and infrastructure news
Filter by type: Post307
- 18 days ago x.com
PufferLib 4.0 reports multi-GPU scaling
Published on June 21, 2026, the report on PufferLib 4.0 describes 6.6× scaling across eight GPUs in a worst-case test and throughput above 100 million steps per second on eight RTX 5090 GPUs.
- 18 days ago x.com
Unsloth reports 1-bit quantization reducing model to 217 GB
In a post published on August 28, 2026, Unsloth reports dynamically quantizing GLM-5.3 to 1 bit: 217 GB versus 1.5 TB in BF16, with about 76% top-1 accuracy.
- 18 days ago x.com
Local branching to scale LLM inference
A paper published on July 5, 2026 proposes generating and selecting token branches during inference; the post reports better mathematical reasoning than three comparison approaches.
- 18 days ago x.com
Token retention may reduce reliance on the KV cache
A publication dated September 13, 2026, claims that keeping four initial tokens and a 64-token window without training matches or exceeds distilled linear-attention models on nearly all benchmarks.
- 18 days ago x.com
Post-training RL with fewer parameters changed
A post published on July 5, 2026 describes an approach that reinforces high-contribution intermediate layers during post-training RL. It may match or outperform RL across all parameters while changing fewer parameters.
- 18 days ago x.com
RubricEM uses step-by-step criteria to train research agents
Published on July 5, 2026, the work describes agents that create evaluation criteria before researching and reuse lessons from assessed attempts.
- 18 days ago x.com
Safety tuning may affect models’ attribution of minds
A study reports that tuning to prevent consciousness claims also reduces the attribution of minds to other entities in three models, and that mechanistic interventions restore these tendencies.
- 18 days ago x.com
Sampling for SFT: evaluating post-training
A post published on October 2, 2026 presents a method that adds sampling to the post-training stack and claims it may improve generalization and reduce forgetting in some cases.
- 18 days ago x.com
Sandbox infrastructure: a scale reference
A publication dated September 22, 2026 reports that a production-scale compute unit spans about 160 nodes and supports more than 380,000 concurrent sandboxes.
- 18 days ago x.com
Data-free self-play: how to evaluate the approach
A paper dated September 25, 2026 explores self-play pretraining without real data and reports predictable reductions in zero-shot validation loss across several modalities.
Agent serving: measure bottlenecks beyond inference
A paper published on August 21, 2026 presents AgentSysBench, covering 10 agentic applications, and reports that inference is not always the main serving bottleneck.
- 18 days ago x.com
Serving models across GPU, RAM and NVMe
A September 12, 2026 post reports running a model on an RTX 5090 with 125.7 GiB of RAM and part of the model file on NVMe.
- 18 days ago x.com
JEPA SIGReg: concepts and a training loop
Published on July 24, 2026, the guide covers complex numbers, Euler’s formula, the Cramér–Wold theorem, and a functional training loop related to JEPA SIGReg.
- 18 days ago x.com
SIGReg and pairwise repulsion in JEPA training
A post dated September 28, 2026 describes how the SIGReg regularizer approximates a sliced MMD and introduces repulsion between batch samples during JEPA training.
Survey maps self-improvement in AI agents
Published on July 19, 2026, the survey reviews 239 papers on improving agents through model updates or changes to components such as prompts, memory, and tools.
- 18 days ago x.com
SwiLA proposes switching between linear attention maps
A paper presented as work at COLM 2026 and dated October 5, 2026 describes an approach that switches among multiple linear maps and highlights the trade-off between expressiveness and memory.
ThetaEvolve: test-time learning
Published on July 11, 2026, the open-source ThetaEvolve framework combines in-context learning and test-time reinforcement learning to explore open-ended optimization problems.
- 18 days ago x.com
TinyLoRA tests reinforcement fine-tuning with 13 parameters
A post published on September 17, 2026 describes a technique using 13 trainable parameters and reports gains on three mathematics benchmarks.
Windows tokens and privilege escalation risks
Published on June 27, 2026, the text explains how the Windows kernel’s _TOKEN structure represents identity and security attributes, and how token replacement has historically been used to gain SYSTEM privileges.
- 18 days ago x.com
TR-DPO limits deviations in DPO training
A September 21, 2026 publication describes adding a KL-divergence penalty to the DPO loss to limit deviation from the reference model and stabilize offline training.
- 18 days ago x.com
Speech transcription: how to evaluate a new model
In an article published on September 3, 2026, Microsoft presents MAI-Transcribe-2, supporting 60 languages, speaker diarization, and word-level timestamps.
- 18 days ago x.com
Streaming transcription and diarization in one model
A September 3, 2026 post describes a streaming model that combines speech transcription and diarization—identifying who is speaking—in a single model.
- 18 days ago x.com
Training decision models with calibrated probabilities
A publication dated October 4, 2026 describes a reinforcement-learning setup that scores options in separate branches and combines outcome prediction with confidence calibration.
- 18 days ago x.com
CPU-based TTS: synthetic speech in 31 languages
A publication dated September 17, 2026, says Supersonic TTS runs on CPU and supports 31 languages. Its author plans to use it to generate synthetic data.
uv can block malicious packages before installation
A note dated July 17, 2026, says setting UV_MALWARE_CHECK=1 makes uv query OSV before installing packages from remote registries and block those reported as malware.
- 18 days ago x.com
Voice of Reason: reasoning over spoken math
In a September 21, 2026 post, Kyutai Labs says its model solves spoken math problems without an intermediate transcript and scores 77.1% on GSM8K.
Xring O3: report cites 15,221 Geekbench 6 points
Published on August 24, 2026, the report attributes scores of 3,945 single-core and 15,221 multi-core to Xiaomi’s Xring O3, with the latter exceeding the Apple M4 result.
Let's assess your project
I reply with scope and a proposal.