#Data/Training

AIDatabricks has launched AI Extract, a capability for extracting fields from PDFs that it says reaches 95% accuracy versus 87% for other tools, at very low cost. The post notes that LLMs' next-token training makes them "autocorrect" content they should preserve, which this approach is designed to avoid. The function can be called directly from SQL and used across the Databricks platform.

AIDatabricks says its in-house AI Extract model, paired with a custom agent harness, achieves a new frontier on complex document processing tasks. The system handles documents over 500 pages and more than 1M tokens, plus nested schemas with 1k+ objects. It decomposes large jobs, runs smaller tasks in parallel, and reconciles them into one structured output.

AIGeneralist says its GEN-1.5 model shows a new level of generality when pretrained on physical interaction data at a scale few thought feasible without shortcuts. The company says it does not yet see where performance gains level off.

AIThe model sometimes uses both hands to rotate a jar lid, employing a fundamentally different contact and motion strategy than in the fine-tuning demonstrations. The post does not specify further details about the setup or results.

AIA fine-tuned robot can clear obstacles, such as a paper covering a bowl, to complete a block-placing task. It does this even though such obstacle-clearing behavior was absent from its training demonstrations.

AIRobot behaviors fine-tuned or prompted on specific demonstrations can generalize beyond them and improvise fundamentally different manipulation strategies toward the same goal. In one example, a model fine-tuned to sweep a block into a bowl with a brush accomplished the task with a dustpan instead, using a very different approach.

AIGeneralist AI reports a robot that, when a Lego brick unexpectedly sticks to its fingertips, uses its other hand to remove it. The post shows this recovery behavior but gives no details on the model, hardware, or benchmarks.

AIGeneralist reports that its robot model can adapt to new physical tasks in 1–10 gradient steps using 1–5 minutes of data, roughly 10–50 demonstrations. The post describes this as test-time training in a low-data regime, and says the results were obtained without tuning the procedure or sweeping hyperparameters.

AIGeneralist's model reached 59% average success on 10 diverse physical tasks with one-shot prompting straight from pretraining. With few-shot learning, using 10 gradient steps on 5 minutes of data per task, performance rose to 83%. The post calls it the first model it knows of that learns a wide range of dexterous closed-loop physical tasks from one or few demonstrations.

AIGeneralist reports that with GEN-1.5, in-context learning can transfer across the embodiment gap entirely in some cases. A human demonstrates a task with their own hands, observed through the robot's cameras, and the robot can reproduce it immediately afterward.

AIPrompts built entirely from simulated experience, such as from a scripted policy, an RL agent, or a teleoperated simulated robot, can produce behaviors on a real robot. The model was not trained on the task in either the simulator or the real world, so the transfer is zero-shot.

AIGoogle has released TIPS g/14 low-res (v1) on Hugging Face, a Text-Image Pre-training with Spatial awareness vision-language model with 1.1B vision parameters and 389M text parameters. The model produces spatially rich image features aligned with text embeddings at 224 resolution, under the Apache 2.0 license. It supports image encoding, text encoding, and zero-shot classification via the transformers library.

AIGoogle has released the original TIPS g/14 (v1) vision-language model on Hugging Face under Apache 2.0, with 1.1B vision parameters and 389M text parameters at 448 resolution. The TIPS family, presented at ICLR 2025, produces spatially rich image features aligned with text embeddings, and the release includes a low-res 224 variant.

AILiquid AI released 4-bit Q4_0 GGUF checkpoints for LFM2.5-230M, LFM2.5-350M, LFM2.5-1.2B-Instruct, and LFM2.5-2.6B, trained with Quantization-Aware Distillation. The company says the checkpoints recover most accuracy lost to quantization, reaching roughly 97% of their BF16 averages while keeping Q4_0 memory footprint and throughput. Benchmarks compare them against post-training quantized Q4_0 GGUFs and against Q5_K_M, Q4_K_M, and Unsloth's UD-Q4_K_XL.

Why it matters: The post shows how quantization-aware distillation recovers accuracy lost in Q4_0 checkpoints, with throughput measured across four hardware backends for deployment tradeoffs.

AISequoia Capital argues that companies increasingly should build and own their AI capabilities, citing open-weight models such as Kimi K3 and GLM 5.2 that now approach frontier performance. Owning intelligence can protect margins as inference costs scale, speed up small distilled models, and keep proprietary data in-house, the article says.

AIStability AI released a Stable Audio plugin that runs Stable Audio 3.0 generation inside DAWs as an instrument, available as a macOS AU and VST3 with Apple Silicon and Intel support. The enhanced StableAudio.com web app adds iterative prompting, audio-to-audio variations, per-track mixing controls, and export, and both tools are in beta and powered by commercially-safe models that users can distribute freely.

AIIan Johnson used UMAP to visualize a video dataset, adding vision latents extracted from Marlin-2B for each clip alongside the included embeddings. He built the interactive map to render smoothly in the browser, with a writeup linked in the post. The quoted post by Daniel van Strien describes indexing 370 hours of Prelinger Archives films into 23,148 timestamped searchable moments.

AIReplit announced enterprise governance updates including more than 50 audit log events that can stream to SIEM tools like Datadog and Splunk. It also launched a beta Admin API for pulling usage, workspace, member, and project data, with workspace settings for company-wide policies and team-level exceptions. Some features are available now, while workspace settings roll out at the end of the week and the Compliance API at the end of August.

AIMathForm-8B is an open-source autoformalization model from OpenBMB that translates natural-language mathematical statements into Lean 4. It was trained on FormalVerse through supervised fine-tuning, then reinforcement learning using Lean compilation and semantic-consistency feedback. The model is available on Hugging Face under Apache License 2.0 and can be served with Transformers, vLLM, or SGLang, using a recommended max_new_tokens of 16384.

AIThe article argues that scientific papers omit the failed experiments and rejected branches that could train AI systems to develop scientific judgment. It describes Alasdair Russell's Cambridge group logging discovery paths as graphs of ideas, and proposes recording six fields per decision, including candidates and outcomes, to test whether this taste transfers to unfamiliar projects.

AIMiniMax introduces Music 3.0, a music generation model that composes, arranges, performs, and produces a complete song from a creative concept and optional lyrics. The post describes an eight-layer RVQ tokenizer, a Hybrid-LM pairing an 8B Global LLM with a 0.6B Local LLM, and a flow-matching and Flow-VAE audio renderer. It says songs can run up to five minutes and that the model focuses on creative intent, arrangement, and vocal naturalness.

Why it matters: The post explains how the model's pipeline targets structure, acoustic detail, and vocal realism, which helps readers judge where open-weights music generation stands.

AIDatabricks has acquired ElectricSQL, the team behind PGlite, a WebAssembly implementation of Postgres that runs in the browser and syncs asynchronously with Postgres instances. Databricks says this suits fast AI agents and plans to use the capabilities to strengthen its Lakebase Postgres offering.

AIAWS has opened registration for Trainium Frontier, a competition where participants train language models from scratch on purpose-built AI chips for NeurIPS 2026. The top prize is $25K, with co-publication alongside Annapurna Labs researchers and a presentation in Sydney. The deadline for entries is September 30.

AINovo Nordisk has chosen AWS as its preferred cloud provider and strategic AI partner to accelerate drug discovery. The collaboration will combine Novo Nordisk's scientific expertise with AWS AI tools, including Amazon Bio Discovery and Bedrock AgentCore, and establish a co-innovation hub in London. The partnership already spans AWS, Amazon Pharmacy, and One Medical.

AIMatei Zaharia argues AI tokens are now a resource to optimize in software engineering, with companies routing all AI usage through an AI Gateway. The approach enables centralized analysis, which found settings on Claude Code and Codex that can substantially lower cost, plus smart routing and per-task budgets for engineers.

AIShanghai AI Lab's InternLM released Intern-MemDec-4B, a 4B-parameter memory decoder that runs alongside an Intern-S2 backbone and a token-level router to add biology knowledge. On all 21 Biology-Instructions tasks, the average score rose from 56.92 to 60.32 when paired with Intern-S2-Preview-397B. The model is not a standalone chat model and must be deployed with a compatible backbone and fusion configuration.

AIAmazon Research Awards named 34 recipients of its Build on Trainium program, a $110M credit initiative supporting AI research at 30 universities on AWS Trainium. The post frames the program as support for academic AI research on Amazon's custom Trainium chips.

AIVoyage AI by MongoDB's full lineup, including the Voyage 4 family, voyage-multimodal-3.5, and rerank-2.5, now runs natively on the Fireworks inference platform. The partnership lets teams run embedding, retrieval, reranking, and generation on one platform and one API. Fireworks says Voyage 4 Large outperforms Voyage 4, Voyage 4 Lite, Gemini Embedding 001, Cohere Embed v4, and OpenAI v3 Large on average retrieval quality.

AIOpenAI's Fidji Simo recommended a post by insitro CEO Daphne Koller on the same topic. The quoted a16z background says the data needed to understand human biology is about 1000x what has been collected so far.

AIKimi Slides handles the full slide-building process, from structure and research powered by Kimi K3 to cohesive design with polished charts and SmartArts. The resulting slides are editable and ready to download. This is the first tutorial in the Kimi Work series.

AIThinking Machines says Inkling-Small began training after its larger counterpart and benefits from the lessons learned. The post cites an improved pre-training data mix, a refined ML recipe, on-policy distillation using Inkling as the teacher, and two additional weeks of agentic coding RL.

AIFireworks AI ran controlled SFT experiments on Qwen3.5-9B comparing LoRA with full parameter fine-tuning across three synthetic verifiable tasks. The post argues that a FullFT advantage can come from data coverage, learning-rate tuning, or adapter rank, and it recommends testing these in that order before switching methods. Under a fixed multi-task budget, FullFT kept a 4.29-point lead over the best LoRA recipe tested, while matched data exposure favored LoRA.

AIAhmad Al-Dahle argues that AI infrastructure faces both a compute shortage and overbuilding, with the four largest hyperscalers planning roughly $725 billion of capex in 2026, up 77 percent from last year. He describes a "mutually assured construction" dynamic in which every well-capitalized player buys the same insurance against falling behind, so the industry overbuilds by construction.

AIPoolside released Laguna S 2.1, an open-weights agentic coding model with 118 billion total parameters and about 8 billion active per token, supporting up to a million tokens of context. Quantized, it fits on one NVIDIA DGX Spark, and Poolside reports 70.2% on Terminal-Bench 2.1 with thinking enabled, with its evaluation trajectories published online. The same week it shipped the Poolside Desktop Assistant for macOS, which runs Laguna locally or alongside Claude Code, Codex, and Gemini agents.

AILiquid AI doubled the LFM2 tokenizer's vocabulary from 65K to 128K without retraining from scratch, extending the original BPE merges and initializing new embeddings as the mean of their sub-tokens. The expanded tokenizer needs 4.0× fewer tokens for Thai, 2.6× fewer for Vietnamese, and 2.4× fewer for Hindi, which the source says yields roughly 2.2–3.7× faster on-device decoding for these languages with no reported quality loss on previously supported languages. LFM2.5-8B-A1B and the expanded tokenizer are available on Hugging Face with open weights.

AIBerkeley AI Research extended the K-Search evolutionary kernel framework with an MLX backend and a CUDA-to-MLX translation layer, letting it adapt existing CUDA kernels for Apple Silicon. The team reports a 0.97x speedup relative to the native MLX Attention kernel and up to a 20x prefill speedup over the community mlx-lm implementation on the Mamba SSM kernel. The method uses Gemini 3.5 Pro Preview to both reason about optimizations and write candidate kernels.

AIFireworks AI describes fine-tuning Qwen3-Embedding-8B on private (query, positive) pairs using bidirectional InfoNCE loss through its Training SDK, then serving the model via an OpenAI-compatible embeddings endpoint. The post reports that around 150 training steps was enough, that rank-32 LoRA landed within about one point of full-parameter fine-tuning, and that gains were largest where the base model struggled, while tasks like CoSQA and FiQA2018 showed flat results.

AIWorld Labs, following its SceniX acquisition, says spatial intelligence must cover interacting with worlds as well as perceiving and generating them. It is sharing early results on building generated worlds that can be used to train robots. The post itself gives no further technical details or figures.

AIIntern Large Models introduces Visual Pretraining, a pretraining paradigm for foundation models that learns directly from visual documents. The post says it outperforms text-only pretraining across backbones and benchmarks, and links the arXiv paper 2607.09657 along with Intern-S2-Preview (35B) and Intern-S2-Preview-397B on Hugging Face, the latter presented as a multimodal foundation model trained with this recipe.

AIMoonshot AI's Kimi has released PerceptionBench, a benchmark that evaluates visual perception as atomic capabilities derived from frontier-model failures across 42 benchmarks. It contains 3,000 verified questions, each isolating a single capability and answerable by looking alone, without reasoning or external knowledge.