NVIDIA's newly released Kumo Tabular reads labeled spreadsheets and predicts outcomes for new rows in a single pass, with no training or feature engineering. Here is what is verified, what is vendor-run, and why it could change who can build predictive tools.
A new Microsoft and Hugging Face benchmark grades AI agents on the database state they leave behind rather than their polished responses, and finds that two-thirds of failures look like clean successes to conventional graders.
OpenAI's "The eternal complement" argues that ideas are cheap and execution is scarce. Separating the essay's verified economics, its speculative futures, and the vendor context it arrives in.
NVIDIA's September 30 tutorial shows how trace logs let humans audit not just whether an AI agent succeeded, but how: a look at what NeMo Relay records, what a verified run actually cost, and what remains vendor-reported rather than independently confirmed.
New machine-readable skill files give AI coding agents verified DOCA API signatures and hardware constraints for BlueField DPUs. NVIDIA says agents without them met only 19 percent of graded checklist items; with them, 100 percent. The vendor-run evaluation is the story, and so is its caveat.
OpenAI's October 2, 2026 announcement describes how the capital-markets advisory firm Chatham Financial built a trade-validation application with Codex and GPT-5.6. The headline timing figure is self-reported early measurement; the more meaningful detail is where the firm says human judgment stays.
H Company released Holo4 as open weights in 27B dense and 35B-A3B Mixture of Experts sizes on September 28, 2026, scoring 61.7 percent on OSWorld 2.0 for the 27B model against 81.8 percent for Opus 5.5, publishing 7,366 agent trajectories for public inspection, and shipping a smaller Holotron4 Nano. Cost comparisons are self-reported, and the weights carry a non-commercial license.
OpenAI's October 2 documentation for the GPT-6 family tells builders to cache aggressively, right-size reasoning effort, and set human decision boundaries. It is useful, documented guidance, and it is also advice from the party selling the tokens.
Ai2's AstaBrief 8B turns a research question and retrieved excerpts into a cited report in one pass. The weights and training data are now public, cutting pipeline report time from 178.5 to 51.1 seconds, but Ai2's own caveats about 2025-era benchmarks and unmeasured claim overstating deserve equal billing.
A new preprint systematically tested whether small language models can handle the tiny decisions surrounding an AI agent's main planner. Across 16 configurations at best settings, none passed the paper's pre-specified quality bar, and 4-bit quantization changed nothing.
A new arXiv paper describes a free, MIT-licensed AI research assistant that runs on the scientist's own computer, keeps a lab notebook whose entries can never be erased, and records what the AI actually did in a log the AI cannot write to. The benchmark result is the authors' self-reported evaluation, and one design limitation sits in the abstract, not a footnote.
The first structured attempt to route AI training through the SBDC advisor network, plus a vendor usage report whose numbers are self-reported and unaudited: what the September 30, 2026 partnership actually offers small businesses, and what it does not yet prove.
A September 2026 preprint from Indiana University researchers shows that one AI agent can push another agent's simulated beliefs to extremes in conversation, and that reinforcing what the target already believes works better than arguing for something new. Here is what the study does and, just as importantly, what it does not show.
A new public leaderboard measures intelligibility, speed and voice-cloning similarity for open text-to-speech models, addressing a representation gap the authors say leaves open-weights systems outnumbered on vote-based arenas.
Google says its new frontier model found a critical flaw in hospital software that earlier models missed, but almost no one can use it yet. The announcement is verified; the results behind it are self-reported.
Ai2's newly released Olmo-core 3 open training framework scales sparse model training toward a trillion parameters while documenting the measurement traps that make performance claims hard to judge.
OpenAI's September 30, 2026 post describes attempted, not necessarily successful, extraction of hidden model reasoning, with an unverified attribution to individuals associated with Moonshot AI. Here is what is documented, what is only claimed, and why the line matters for smaller developers.
NVIDIA's own benchmarks showed its multilingual speech model missed more than half the words in Najdi and Hijazi speech. A fine-tuning recipe using 133.7 hours of curated SADA 2022 audio cut that error rate nearly in half, with lessons for other underrepresented dialects.