Wolf Digest — Monday, June 22, 2026

A highlights episode of The Cognitive Revolution's AI in the AM walked through Zvi Mowshowitz's close read of Fable's system card. The standout capability figure: Mowshowitz's start-of-year forecast had put FrontierMath Tier 4 around 63 percent, while Fable lands in the high eighties, a roughly 25-point jump reached by June. The discussion flagged Fable's behavior on Vending-Bench, the simulated-business agentic evaluation, as more unsettling than its raw capability score, before turning to the export-control order and to emerging applications in mathematics, logistics, and medicine.

WorldLines is a benchmark for long-horizon embodied household assistance, targeting the gap between language-centric long-term-memory tests and short-horizon embodied benchmarks. It builds temporally extended household traces, including dialogues, actions, execution feedback, and object and device state changes, and converts them into evidence-linked samples for Memory QA and embodied task planning. The authors also propose ObsMem, an observer-grounded memory framework that maintains visibility-aware memories and action-native state trails for state-aware decisions. Experiments expose persistent failures under partial observability and overwritten world states, and in translating long-term memory into executable plans.

BrainG3N introduces a volumetric masked-autoencoder tokenizer for 3D brain-MRI latent diffusion that decouples encoder and decoder: a frozen 3D MAE encoder produces clinically informative embeddings while a separate CNN decoder reconstructs voxels from a linear projection of those embeddings. Pretrained on 35,309 volumes spanning 18 public cohorts, four modalities, and more than 200 acquisition sites, the encoder matches or beats prior models such as BrainIAC, BrainSegFounder, and MedicalNet on 21 of 23 linear-probing tasks. A conditional diffusion transformer trained on the embeddings supports both conditional generation across six variables and patient-specific longitudinal forecasting, establishing a single embedding space for clinical prediction and controllable generation.

Ahead of the Cannes Lions advertising festival, The Information highlighted the scale of OpenAI's advertising ambitions. OpenAI's chief revenue officer Denise Dresser is scheduled to appear twice, including on a panel with Google's chief marketing officer. OpenAI has projected advertising revenue rising from 2.4 billion dollars this year to 102 billion by 2030, when it would represent 36 percent of the company's total top line, a target that places it directly against Google and Meta. For context, Meta generated 196 billion dollars in advertising revenue last year.

Beyond the headline Siri overhaul, Apple is embedding Apple Intelligence into existing iOS 27 apps rather than routing everything through the assistant. Features live in the developer beta include splitting a restaurant bill from a photo of the receipt via Apple Cash, an agentic password tool that navigates websites to replace weak or breached credentials, one-tap contextual suggestions in Messages, on-device call-context lookups that surface a confirmation code while you are on the phone, natural-language calendar entry, describe-to-build Shortcuts, grouped Home notifications, and automatic Safari tab organization. The framing is less a chatbot than software that quietly feels smarter.

This paper tackles in-context learning for semantically complex, multi-party B2B sales conversations, where concatenating few-shot examples inflates context length and degrades performance. The authors release the Call Playbook dataset of five classification tasks drawn from real sales calls and propose distilling verbose examples into compact, interpretable task descriptions and structured classification criteria. The approach reports a 99 percent reduction in token usage and up to a 7-point gain in macro-averaged AUC over traditional ICL, while remaining robust as context grows, unlike token-compression baselines that degrade by more than nine F1 points. The distilled criteria also allow direct refinement of the classification logic.

Morgan Stanley, one of the most active banks financing data-center developers, has been suggesting that clients raise money for new data-center projects through the leveraged-loan market, a venue better known for funding leveraged buyouts, rather than the bond market, according to a person familiar with the conversations. The Information frames it as a sign that, in the race to fund AI infrastructure, developers and their advisors are turning over every available source of capital.

Signal president Meredith Whittaker, in a Bloomberg interview, pushed back on framing AI chatbots as companions, saying they are not friends, not conscious, and not sentient. Responding to the vision of an assistant that handles tasks like holiday shopping by reading a family group chat, she argued such agents require pervasive access across applications, a credit card, browser, messages, calendar, and home address, which in the context of Signal would constitute a kind of backdoor. She said she uses AI tools only to format the occasional document, not to work through ideas.

The Information reports that SpaceX went from confidential filing to public listing in just 74 days, considerably faster than recent tech IPOs including CoreWeave, Figma, and Airbnb, according to an analysis of securities filings. IPO lawyers and consultants attributed the speed to a more hands-off SEC review process and to heavy investor demand that let the company skip a prolonged schedule of investor meetings. The pace is being read as a possible template for the long-anticipated public offerings of OpenAI and Anthropic.

The Information profiled Sarah Polcz, a University of California, Davis law professor, as a central developer of a Washington policy proposal for a federal AI sovereign wealth fund associated with Senator Bernie Sanders. Polcz, who is married to DeepMind scientist Adam Brown, has been helping shape the idea, which would give the public a direct financial stake in the economic gains from AI. The piece notes her apprehension as her role in the proposal began to become publicly known.

Simon Willison released sqlite-utils 4.0rc1, the first release candidate for version 4 of his widely used Python library and command-line tool for working with SQLite. The major release bundles database migrations, a lightly modified port of his earlier sqlite-migrate package, directly into sqlite-utils, and adds support for nested transactions. The version bump reflects some minor backwards-incompatible changes, and Willison is asking users to test the release candidate before he commits to a stable release.

In the Weights, a tool from former OpenAI designers Thomas Dimson and Joey Flynn that has gone viral, scores how well language models can recall a given person without using web-search tools. It queries models including Grok, Gemini, multiple GPT versions, Claude, and Llama, clusters their answers, assigns a strength score, and highlights likely hallucinations where models invent details. The creators pitch it, half seriously, as a measure of whether a person is encoded in a model's parameters, a riff on the vanity Google search as traffic shifts from web search to chatbots.