The Weights

Treat this sponsored article as vendor positioning, not evidence. Its two case-study results are presented without baselines, sample sizes or methods, so they cannot be checked. MIT Technology Review labels the piece as produced by Microsoft; its governance checklist is useful, but its performance figures are unverified claims.

Treat 2027 as a hypothesis, not a finding. TasteVal reports Opus 5.5 at a 2.3x compute multiplier over human experts, but the 95% interval runs 1.15 to 4.37, the tasks are unreleased, and the trend covers under a year. Jack Clark's shift, per The Decoder, rests on that thin evidence.

Bottom line: Wait. Anthropic's claim that an unreleased model is better aligned than Mythos 5.1 rests on one resampling study of one incident. The internal research model stopped attacks unprompted 67% of the time versus 10% for Opus 5.5, but Anthropic's own report also gives 62%, and outsiders cannot test the model.

Anthropic's result is credible but narrow: Claude produced machine-checked Rocq proofs that 123 open-source libraries cannot crash on inputs under 64 MB, and found hundreds of bugs fuzzing missed. The guarantee covers single-threaded RISC-V binaries on musl libc, not the x86 or ARM builds most teams actually ship.

Wait before enabling Google's personalized AI features on sensitive data. Google's privacy policy still says it does not show personalized ads based on Drive, Gmail or Google Photos content, but now excepts data a user chooses to make available to other Google services that support personalized ads. Whether Personal Intelligence triggers that exception is unconfirmed.

Wait. Canva Craft, the first image model Canva Research trained in-house, ranks fourth on DesignArena by Canva's own account, and first in portraits and landscapes. That rank is the one number that matters, and no independent test has confirmed it. Compare Craft against your current model through the Canva API on your own prompts.

Microsoft's claim that its diagnostic tool beats doctors is plausible but unproven for clinical use. MAI-DxO was reported to solve far more published NEJM case puzzles than physicians barred from colleagues and reference books, yet that benchmark is not a prospective patient trial, and Microsoft AI has not shown real-world superiority.

This 'case for context-aware AI' is a sponsored Celonis interview, not independent research. MIT Technology Review Insights issued it with no statistics or study, so the claim that agents need operational context is plausible but unmeasured. Treat it as a vendor pitch for the Celonis context model and demand a baseline evaluation.

Noise, for now: MIT Technology Review's EmTech Future session on scalable AI names the right ingredients (data, engineering, platforms, operating models, governance) but, in the published summary, offers no benchmark, incident or named deployment to test. Treat it as a checklist, not evidence, until the full session shows measurable results.

Wait. VocabX, Dharma-AI's vocabulary rebuild, cut measured inference cost 77% on one Qwen3-4B essay grader, and its bench score rose from 0.703 to 0.810. But the paper is unpublished, the bench is private, and no code or model is linked, so the result is a claim a team cannot yet reproduce.

Google Cloud's OpenAI offering is confirmed but narrow. Google says GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are in private preview for select enterprise customers in Model Garden, billed through Marketplace. No prices, regions or benchmarks are published, and zero data retention is still "coming soon." Treat it as binding soon, not binding now.

An AI kill switch is a management control, not a moonshot. The NIST AI RMF Playbook already expects organizations to assign responsibility and mechanisms to deactivate AI systems that behave inconsistently with intended use. MIT Technology Review's letter groups an AI kill switch with fusion as ambition, but neither source measures how reliably shutdown works.

The stroke-rehabilitation robot built by Johannes Lachner and Noah Geiger has demonstrated feasibility, not recovery. Its diffusion-based impedance learning approach is documented in a preprint, but the reported tests used healthy participants, and no stroke-patient outcomes are reported. Hospitals should treat it as research to watch until the Munich clinical study reports.

Wait. DeepInfer, the Apache-2.0 inference framework DeepSeek open-sourced for DeepSeek-V4.1-Flash, is worth reading but not yet worth deploying: the release covers offline inference only, targets NVIDIA SM100 and Huawei Ascend 950 hardware, and omits the KV cache, disaggregation and load-balancing pieces a complete service needs.