AI news for builders and product teamsUpdated Nov 17, 2026, 13:01 UTC
The Decoder
Reporting and perspectives from The Decoder. Headlines and excerpts link to the original articles.
Latest stories
Newest first
Google DeepMind released AlphaGenome 2, a genomics model that predicts variant effects for individual cell types rather than bulk tissue, trained on over 107 million cell profiles from 68 tissues. It found roughly twice as many disease-relevant variants as its predecessor and is available for noncommercial research via API, with a closed commercial preview.

Australia's government, led by Prime Minister Anthony Albanese, has announced a law that would impose mandatory safety rules on frontier AI companies for the first time, including independent audits, ongoing safety checks, published safety frameworks and immediate incident reporting, enforced by a new oversight body with steep penalties.

Google DeepMind has launched the Gemini Decisions API, powered by Gemini Decision 1, which evaluates up to 1,000 closed-ended questions in one pass across text, images, audio, video and PDFs, returning probabilities and confidence scores in 150 to 250 milliseconds at $0.10 per million input tokens.

US corporate AI spending reached a new high in October 2026 but rose only 1.65% month over month, according to the Ramp AI Index, which draws on payment data from more than 85,000 US companies.

Google DeepMind's AlphaProof Code formally verified the cryptographic core of Signal's encryption protocol in four weeks, producing over six million lines of Lean proof code with AI agents built on Gemini Carbon. DeepMind says such verification work typically takes years.

Helsing unveiled Clarity, an agent-based platform that fuses sensor and intelligence data into target recommendations for weapons systems, with a human required to authorize every handoff. It runs on customer infrastructure and is positioned as a European alternative to Palantir's Maven.

Anthropic says its in-house model Mythos ported Claude to AMD MI355X GPUs in three days with almost no human help, against a usual six to twelve months for onboarding new accelerator hardware.

A ChinaTalk investigation found at least nine US data firms, including Mercor, Surge AI, AfterQuery and Turing, selling roughly $500 million of AI training data and reinforcement learning environments to Chinese AI labs, with export controls barely applying.

DeepL has introduced Aure, a model for DeepL Voice that carries tone, emotion, emphasis and speaking pace from the original speaker into the translated speech, instead of relying on a single voice sample taken at the start of a conversation.

England's Office for Students says AI has made fair grading of coursework nearly impossible and has warned universities they could face sanctions unless they adopt AI-free assessment formats such as traditional exams and in-person presentations.

OpenAI is beta-testing "Businesses in ChatGPT," letting companies claim profiles in a Business Hub with product feeds and analytics, and enabling eligible US e-commerce sellers to build "Business Agents" that answer customer questions inside ChatGPT.

Nvidia's GR00T 2 model learns robot tasks by transferring human hand movements from video instead of relying on teleoperation, trained on roughly 400,000 hours of footage, though success rates range from 10 to 80 percent depending on the task.

Google has made Gemini 4 Carbon generally available to paying developers and business customers after a two-week testing phase, priced at $2 per million input tokens and $10 per million output tokens as an introductory rate. It also previewed a "/swarm" feature in Antigravity that turns one prompt into a coordinated team of agents.

Anthropic's Fable 5.2 finished 41.3 percent of real freelance jobs at professional quality on the Remote Labor Index, more than double its closest competitors Opus 5.5 and GPT 6 Astra, versus a 2.5 percent top score a year earlier.

Eight countries and representatives from OpenAI, Anthropic, Microsoft and Mistral AI signed the Paris Call for AI Safety and Security, committing to shutdown mechanisms, independent audits and risk disclosure, though without binding sanctions or deadlines.

Anthropic released Claude Fable 5.2 and Mythos 5.2, a shared model foundation split by access tier, ahead of its IPO; Fable 5.2 tops the Artificial Analysis Intelligence Index v5 at 65 points, five ahead of Gemini 4 Carbon.

Bytedance Seed has introduced Seedream 5.5, an image model it says keeps edits stable across up to 100 rounds, preserving untouched areas at the pixel level without color shifts. It also outputs native RGBA images and is rolling out gradually via Jimeng, Doubao, Doubao Work and Volcengine Ark.

Google has added two systems to its eighth-generation TPU line, the TPU 8tx for confidential computing and the TPU 8i Max for memory-bound workloads, bringing it to four TPU systems shipped in a single year and nearly 100 external customers.

Nvidia's board has launched an independent investigation after CEO Jensen Huang admitted in a company-wide memo to an extramarital relationship with a woman US intelligence agencies suspect is a Chinese informant. The disclosure undercuts Nvidia's lobbying against tighter AI chip export controls as a newly elected Democratic Congress prepares to take committee control in January.

On Alibaba's new Co-Design Bench, built from real user requests that previously defeated coding agents, the best model, GPT-6 Astra, fully resolves only 26.4 percent of tasks, far below the 99-plus percent scores models post on older benchmarks.