OpenAI will watermark ChatGPT and Codex text in the EU
Over the coming weeks, text generated by ChatGPT and Codex in the EU will carry an invisible watermark called textGrain: where several words fit equally well, a secret key picks one, and with the key the statistical pattern becomes detectable in longer passages; API customers can opt in worldwide, off by default. This is a compliance move under the EU AI Act’s transparency rules (applicable since August 2, 2026), and I wrote in September about the detection-authority and key-governance problems with this class of scheme (watermarks without verification). OpenAI’s own numbers draw the capability line: detection is about 95% at 400 tokens but falls to 17% once a quarter of the words are swapped for synonyms, so it holds up against copy-paste and degrades sharply once the text is rewritten.
ChatGPT signs fake New Yorker cartoons with real cartoonists’ names
Nieman Lab found that when asked for a “New Yorker-style cartoon,” ChatGPT pastes in the signatures of more than 15 real cartoonists; one fake signed “BLOPER,” Brendan Loper’s pen name, went viral on Facebook and readers asked Loper whether he drew it. OpenAI signed a licensing deal with Condé Nast, The New Yorker’s parent, in 2024, but per Nieman Lab it did not cover training on cartoons. Read this next to the item above: a watermark signals that content came from an AI, while the problem here runs the other way, AI output claiming to be the work of a specific person, which schemes like textGrain are not designed to catch.
Reflection announces Beam, a 501B open-weight model
Beam is a sparse mixture-of-experts model, 501B total and 23B active parameters, built for coding and agent workloads; Reflection says reasoning is on par with GLM-5.2 at 3–4x less inference compute, with weights due later this month under Apache 2.0 (see also TechCrunch). Two caveats: the weights and technical report aren’t out yet, so every number is self-reported, and the headline comparison targets are Chinese open models like GLM and Qwen. Selling enterprises and sovereign buyers “an open-weight option that isn’t Chinese” is the real positioning, and it says more about the open-weight market than the benchmark table does.
OpenAI tests visual ads inside ChatGPT image generation
Starting this month in the US with a first group of advertisers, labeled visual ads can appear while an image generates, placed apart from the result; OpenAI says ads do not influence what the model produces, and it is piloting brand-suitability measurement with DoubleVerify and Integral Ad Science. Bringing in ad-measurement firms on day one signals a real revenue line, not an experiment. Every ad platform promises “ads don’t affect results” at launch; the test comes once the revenue becomes something to defend, so this is one to track.
TikTok launches an AI shopping agent and one-click checkout
At Advertising Week, TikTok announced Shopping Assistant, a conversational shopping agent, and Buy Direct, one-click in-app checkout (TechCrunch), built with Salesforce, Shopify, and Stripe; on the merchant side, Salesforce says Buy Direct checkout runs on the Universal Commerce Protocol, a standard interface that lets agents call inventory, ordering, and payment systems under shared rules. Platforms keep piling into agentic commerce because the contest is the same everywhere: whoever hosts the conversation where the purchase decision happens takes the transaction.
HackerRank’s AI interviewer Chakra goes GA after 500,000 interviews
HackerRank moved Chakra (official product page: chakra.sh) out of a roughly six-month beta in which it ran over 500,000 interviews, with Snowflake and Capgemini among the early users. It folds the recruiter screen, take-home, and engineer follow-up into one session and scores process signals, including what HackerRank calls AI fluency: how a candidate frames a problem for AI, then checks and corrects its output. For engineers on the market the signal is plain: this interview grades how well you use AI instead of banning it, and half a million sessions means it is already process, not pilot.
Dust: pretraining transformers without backpropagation
Dust, from Q Labs, is a zeroth-order method: it adds Gaussian noise to activations independently at every token (perturbing activations, not weights), so one forward pass evaluates a huge virtual population, then weights each perturbation by how much it reduced the loss to estimate weight gradients. At up to 243M parameters it comes close to backprop, and larger models use the population more efficiently, the opposite of the usual expectation for zeroth-order methods. Still small scale, but a forward-only training recipe would loosen two hard constraints at once, memory for stored activations and cross-chip gradient sync, so the next scale-up is worth watching.
An 1809 Napoleonic cipher letter read by AI
Carter Church, a staff AI engineer at SentinelOne, used GPT-6 Astra to read an encrypted 1809 letter sent on Napoleon’s orders by his stepson Eugène to Marshal Marmont: six hours, building on 33 characters earlier cryptographers had solved, using simulated annealing to score candidate plaintexts against French letter statistics; the writeup is from September 18 and drew attention this week after Bruce Schneier picked it up. Church’s own framing is modest: “a patient specialist could have done this in 1970.” Same curve as last month’s Enigma break by the same model (see the September 22 briefing): what changed is cost, not method, and I wrote about where that curve points for real weak cryptosystems in my cryptanalysis post.
Research radar
RealCompanion: benchmarking long-term memory in AI companions
The benchmark is built from ten real human-AI relationships spanning months and tests whether a model remembers user history, recalls it at the right moment, and reasons over it, instead of simulating memory in single turns. Companion products advertise memory far more than anyone measures it, and this is a rare benchmark shaped like actual use. If you build memory systems or companion apps, read it.
Does learning protein folding transfer to broader reasoning?
Trains on protein folding, a task with exact, machine-checkable answers, and asks whether the gains carry over to general reasoning. The appeal is a reward signal that sidesteps answers that are only superficially correct in human text. Researchers thinking about where reasoning training data comes from should check which tasks the transfer actually appears on.
Latent-MOPD: distillation moves from output distributions to hidden states
In on-policy distillation with several teachers, it fuses the teachers’ hidden states alongside their output distributions; the authors say it is the first in this line to fuse at the representation level. On-policy distillation keeps surfacing in recent weeks, and this is a concrete method-level step. Worth checking whether the reported gains cover the training-time cost of running several teachers; the deployed student is still a single model.
One line for today: A watermark signals “an AI probably made this”; a forged signature claims “a human made this.” The first now has a law, keys, and published detection rates; the second, in this week’s case, surfaced only because readers asked the cartoonist himself. That second half is where provenance actually leaks.