Topic
Determining the origin and authorship of web text used to build model training and retrieval corpora, and the contamination risk from synthetic pages.
OpenAI's new mathematicians' panel must first help release scores more results the company says an unreleased model produced. Researchers call the panel a good first step, though many are unsure how much say it will have over how OpenAI announces what its models find.
Reality
Authors suing OpenAI filed a 162-page statement of facts drawn from its Slack threads and paper drafts, including a 2022 project to delete LibGen data. Any team training on outside data should expect its dataset codenames and cleanup channels to be read the same way.
Reality
A detector run over 490,000 Common Crawl pages puts a number on synthetic text in the crawlable web, and the number is uneven enough to change how you build a training set.
Reality
Sony and Warner want evidence about which founder approved which corpus, and that request is aimed at internal records rather than model weights. The complaint is effectively a test case for any lab whose pipeline cannot answer that question per corpus.
Perspective Coverage
3 publishersReality
Buckmaster was preparing to publish when OpenAI announced it had solved the problem and offered him a share of the credit. A second mathematician, Andreas Thom, then asked the company whether his own ChatGPT conversations had reached its model.
Reality
Redactions came off a 92-page plaintiffs' brief in the consolidated OpenAI copyright case on Thursday. Inside are internal Microsoft and OpenAI documents on substitution, memorised training data and collapsing publisher click-throughs.
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
A week and a half after accusing OpenAI of copying his approach to a problem carrying a $1 million bounty, the NYU mathematician is still running the company's coding agent on his own papers.
Reality
Version 2.0 arrives at the same price as version 1.0. The accuracy gain SpaceXAI reports is concentrated on short multilingual commands. The evaluation sets behind the claim are drawn from production support calls.
Reality
Alex Karp told CNBC that frontier labs are angling for government liability cover because their commercial clients would otherwise sue them. Buyers whose data feeds those models should read their contracts with that in mind.
Reality
Anthropic's June decision to keep Fable usage logs for 30 days has turned zero data retention into a procurement gate at Nvidia, Booz Allen and Palantir. The metadata channel stays open under the clause they are demanding.
Perspective Coverage
4 publishersReality
The AI Act's general-purpose AI obligations became enforceable in August, and the parts of them that touch a data pipeline depend on records only the collector can write while it is fetching. Penalties scale with turnover.
Reality
The publishers want $150,000 for every alleged infringement and say Claude will recite lyrics verbatim, which moves output filtering into the column of legal controls rather than polish for anyone shipping a text feature.
Perspective Coverage
6 publishersReality
Tristan Buckmaster used OpenAI's Codex on the Navier-Stokes problem, then asked whether his prompts had fed the system that beat him to a proof, and the company's answer to The Verge got firmer over time.
Reality
Tristan Buckmaster told the ABC that the game is up for mathematicians. In the same week, OpenAI described what his Codex prompts did to its models in two incompatible ways, and a customer has no way to pick between them.
Reality
OpenAI says the run took 88 hours and cost millions, which tells a small research group more about whether it can compete than the proof itself does, and its wording on user data shows the limit of what any vendor can promise.
Reality
The Navier-Stokes result cost more in compute than the Clay Institute's award, and the interesting line in OpenAI's blog post is the one conceding it cannot rule out that customers' de-identified prompts helped.
Reality
A mathematician kept a year of unpublished drafts in Codex, then asked whether they had been used for training and got silence on that half of the question, which is the half every enterprise rollout treats as settled.
Reality
A computer engineer writing in Fast Company says dermatology models latch onto background skin tone instead of the lesion, which turns training-set composition into a spec question for anyone shipping skin imaging to patients.
Reality
Sony Music and Warner Chappell want up to $150,000 for each of thousands, maybe tens of thousands, of songs, and they argue in the filing that the $1.5bn Anthropic already paid was too small to deter anyone valued at $2 trillion.
Reality