Microsoft says Copilot cut New York Times clicks by up to 93% versus traditional Bing, a concrete sign that AI search can replace the publishers whose work supplies its answers. The figures come from the Times’ brief, while the underlying exhibits remain sealed.
Qwen-Image-2.1 is slated to become an open model on September 20, 2026, giving creators one workflow for generating images and refining selected details. ModelScope describes it as compact and built for fast inference, with support for diverse inputs and outputs plus localized edits, so users can change part of a composition without starting over. The page does not yet provide benchmarks or pricing.
StepFun’s Step 5 Preview is available through an API, but the BF16 model is not yet something you can run yourself. Its Hugging Face repository contains only a .gitattributes file, with no weights, license, model card, or configuration; StepFun’s materials put the open-weights release on October 15, 2026. For now, self-hosting plans have no checkpoint behind them.
Program-as-Weights lets developers describe fuzzy tasks such as log alerting or malformed-JSON repair in natural language, then run the resulting function locally instead of calling an LLM API for every input. Its 4B compiler emits an adapter for a frozen 0.6B Qwen3 interpreter, which matches direct Qwen3-32B prompting while using about one fiftieth the inference memory and running at 30 tokens per second on a MacBook M3. The project releases the 10M-example FuzzyBench dataset.
Noam Brown’s interview connects multi-agent systems to the harder question ahead: proving models are aligned before recursive self-improvement begins. It moves from AI agents tackling hard mathematics, including a system described as 10,000 agents using 130 billion tokens over 88 hours, to the gap between a model’s internal reasoning and its outward behavior. The practical test is whether we can trust models as they help automate AI research.
CXMT says its fifth-generation DRAM platform is now in mass production, giving phone and device makers another memory supplier with chips designed to cost less and use less power. The platform stores structures closer together, fitting 50% more data into new mobile-memory chips and producing at least 50% more dies per wafer, according to CXMT. It uses quadruple patterning to create 11.95-nanometre features.
Google’s CC can now coordinate a household for up to six members, turning chosen emails and calendars into a shared daily brief instead of leaving logistics in one person’s inbox. Each member decides what CC can see, while the agent can add events, maintain task lists, fill permission slips or registration PDFs, and draft meal plans with permission. It uses its own verified Google Account and only responds to group members.
Microsoft’s own data says Copilot’s answer engine reduced click-throughs to The New York Times domain by as much as 93% versus traditional Bing, giving publishers a concrete measure of the substitution risk in AI search. An internal presentation called the effect a “doom loop”: the system threatens the content suppliers it relies on, while mass scraping and paywall bypassing feed the models. The figures come from the Times’ brief, while the underlying exhibits remain sealed.
Mustafa Suleyman says AI models should not be treated as having feelings, preferences, rights, or any claim on human welfare, because doing so could make containment and alignment harder. His argument starts with consciousness: since it underpins human ethical, legal, and political systems, extending comparable rights to another entity is not justified by the evidence, he says.
Agora lets research agents build on one another’s experiments without sharing a conversation or workspace, giving teams a searchable record of what worked, failed, and was verified. Each contribution becomes an immutable Git commit, while evidence scores and an analyze() API point workers toward promising or neglected branches. In a 12-day run, 13 workers published 1,703 contributions and improved the target from a 3.3923 bits-per-byte baseline to 1.899.
You’re caught up.
That was all 9 stories in the 20 September issue.