DeepSeek-V3 uses sequential multi-token prediction (MTP) modules, in which predictions at one depth feed into the next depth, according to a technical explainer published via Towards AI. The article d…
A Stanford University paper introduces Self-Organizing Agent Teams (SAT), a framework in which groups of AI agents learn collaboration strategies from as few as 15 AIME 2024 or 25 GPQA Diamond problem…
Researchers introduced RULER (Instance-aware Rubric Rewards for Reinforcement Learning), a method that converts each natural-language instruction into an instance-aware rubric of six items spanning se…
A census of self-published AI training-cost disclosures found 21 figures from 12 organisations published between May 2022 and September 2026, of which only six carry a dollar amount, and none has been…
ByteLex reported that a tokenizer-free coordinate map built from 11 tokenizer vocabularies over 3-byte windows predicted its 237M-parameter byte-level language model's per-word errors with a Spearman …
NVIDIA reported that combining its Transformer Engine library with the JAX Python library raised DeepSeek-V3 MoE training throughput on NVIDIA GB200 GPUs from an unoptimized baseline of 103 TFLOPS/GPU…
A developer explains the evolution of large language model training from supervised fine-tuning to reinforcement learning from human feedback and reinforcement learning with verifiable rewards, noting…
MoneyPrinterTurbo, a widely starred open-source workflow engine, automates short-form video generation by orchestrating LLM narrative planning, multi-provider media retrieval, and programmatic renderi…
DeepSeek-AI released the DeepSeek-V4 series on April 24, 2026, including the MIT-licensed DeepSeek-V4-Flash model with 284B total parameters (13B activated) and a one-million-token context window, and…
A new technical blog series by an unnamed author provides a roofline-to-reality performance analysis of DeepSeek-V3, a mixture-of-experts transformer model recently added to MLPerf 6.0 as a large-scal…
In a benchmark of three large language models for detecting subtle security vulnerabilities in code, Claude 3.5 Sonnet outperformed GPT-4o and DeepSeek-V3 in identifying an Insecure Direct Object Refe…
A developer added a fourth model, Mistral Small 3.2, midway through a field test of AdversarialDebate, changing the experiment's outcome. The addition revealed that maximum diversity can lead to capit…
NVIDIA's Spectrum-X Ethernet architecture, designed for giga-scale AI factories, maintains stable training step times of 668 ms under heavy multi-tenant congestion, while standard Ethernet slows from …
LLM tokenizer variance means identical dollar-per-million token prices can produce materially different bills, because each vendor's tokenizer segments the same input into different token counts. Anth…
A new open-source tool called Model Genome Korea can determine whether a Korean large language model (LLM) or vision-language model (VLM) was trained from scratch or derived from an open-weight base m…
Claude 3.5 Sonnet outperformed GPT-4o, Gemini 1.5 Pro, and DeepSeek-V3 in Spring Boot configuration and code generation tests on a legacy Java codebase, correctly identifying transaction boundaries an…
AI developers are shifting from manual prompt-based workflows to building Model Context Protocol (MCP) servers that connect AI models directly to local databases and file systems, enabling automated c…
In a benchmark comparing AI models on designing a Unix-like file system, Claude 3.5 Sonnet outperformed GPT-4o by providing precise block mapping and compile-ready C code with bounds checking, while G…
ByteDance is training a single AI model with 10 trillion parameters, a scale that requires extreme 3D parallelism and massive high-quality data, according to a report. The move signals an ongoing race…
DeepSeek-V3, an open-weight large language model from Chinese AI company DeepSeek, has leaked and is matching or beating top-tier proprietary US models in coding and math benchmarks, according to a te…