Topic
Generation and use of model-produced data for pre-training and fine-tuning, and the debate over its limits.
ServiceNow's CoreAI team generates new agent training tasks, each with its own verifier, aimed at gaps a stronger teacher model can already solve. Adopting it means running a seedable copy of the environment and writing verifiers that accept any valid solution.
Reality
NVIDIA released Kumo Tabular, a commercially licensed tabular model it says ranks first on four public benchmarks with no per-task training. At inference the labeled table replaces the fitted model, so the rankings carry over only where a team's tables resemble the benchmark sets.
Reality
Salesforce and Nvidia post-trained Koa on simulated CRM workflows and claim three times fewer errors than leading general-purpose models. The published record describes the training data two different ways.
Perspective Coverage
3 publishersReality
Diogo Almeida left OpenAI two years ago convinced that human language is the wrong output for automation. The model his startup shipped this week returns probabilities, and one team testing it clocked classification 5 to 18 times faster.
Reality
Brad Moon's Kitchener-Waterloo startup closed a $12-million Series A led by Caffeinated Capital to automate analog physical design. The bet is that a foundry-calibrated physics environment can manufacture the layout data chipmakers keep secret.
Reality
A new paper finetunes GPT-4.1 and Kimi-K2.6 on stories about humans. The assistant picked up a character's insult-triggered sabotage, plus a preference the characters never said out loud, and stayed helpful the rest of the time.
Reality
Bespoke Nimble's LoRA fine-tune of Qwen3.5-9B scored 90% against Jev's 93% on an eval it curated itself, and five more replications landed at scales between 421M and 35B parameters. No standard benchmark exists for the category yet.
Reality
The labels answer Suno's from-scratch claim for v6 by arguing the model learned from the outputs and preference data of earlier models trained on 60,202 of their recordings. The statutory ceiling reaches $9.03bn.
Reality
The labels read Suno's own disclosure that v6 learned from user interactions as an admission that the outputs of its earlier scraped models fed the new one. They put the count at 60,202 recordings and the demand at up to $9 billion.
Reality
Andrew Yang told CNN the open internet is now unusable for training models. TechCrunch's rebuttal came from an unnamed security professional. Anyone budgeting next year's training data gets two anonymous positions.
Reality
Jev launched closed on Wednesday, and two days later six teams had published reproductions with almost nothing in common underneath. Only one of them published a score, measured on an eval it curated itself.
Reality
Cua published a 706,048-parameter model, its training data and its driver integration under MIT, so the claim that a small specialist can take over the per-field decision step is testable by anyone with forms.
Reality
A new synthetic dataset gives Arabic OCR training six fonts and 794.6 million words of perfectly aligned ground truth. Every page is rendered clean, so none of it tells you how a model copes with a bad photocopy.
Reality
AWS's synthetic hazard pipeline reports up to 160 percent higher person-detection mAP50 from images one model edited and a second model labeled. How starved the baseline was decides what that multiplier is worth.
Reality
The company names DeepSeek, Moonshot and MiniMax, and says the accounts reached Claude through commercial proxy resellers, in breach of the regional limits it sets because it does not sell in China. Its case rests on account-level traffic patterns.
Reality
An experiment ran the subliminal learning pipeline five times over on Qwen2.5-7B-Instruct, making each student the next teacher. The cat trait held near 90 percent under longer training and scattered under shorter.
Reality
Existing 2D/3D registration models align well for some patients and fail for others. MIT's xvr, out today in Nature, fits itself to one patient in about five minutes, then matches that patient's X-rays to their 3D scan in seconds.
Reality
Koa was trained on synthetic data drawn from nearly 30 years of Salesforce deployments and runs inside Agentforce, where Anthropic's Claude now arrives as Claudeforce with 37 prebuilt skills. Admins get two model layers to answer for.
Reality
Labeled atomic force microscopy images are scarce, so an Oak Ridge team simulated the drift, noise and tip artifacts that operators fight every day, then tested whether models trained on the simulation find real features in real scans.
Reality
The 35-billion-parameter Iris-mini and the 397-billion-parameter Iris-pro build on Qwen models and run at 256,000 tokens of context. AllSpark also published the training recipe and scored every benchmark twice, once with context management switched off.
Reality
Invest · September 12, 2026 · 1 publisher
Science · September 11, 2026 · 1 publisher
Build · September 10, 2026 · 1 publisher
Build · September 8, 2026 · 1 publisher
Product · September 6, 2026 · 1 publisher
Build · August 30, 2026 · 1 publisher
Build · August 22, 2026 · 1 publisher
Build · August 21, 2026 · 1 publisher
Product · August 19, 2026 · 1 publisher