Wednesday, 16 September 2026

What actually happened in AI. Not what got the clicks.

Every story is boiled down to the event underneath it, then scored on evidence, concreteness, impact and actionability. Hype is penalised. Anything you cannot verify never makes the front page.

Live signals: 60
Avg score: 66
Screened, 24h: 70
Scored, 24h: 27

Last run 12:04 UTC. Most of what we read never gets here.

Top signal

Google launches Gemini 3.8 Live and 3.8 Live Extended Thinking voice models with real-time visual context and background task execution

Google DeepMind launched two new voice/audio models, Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, rolling out starting today via the Gemini API, Google AI Studio, Gemini Enterprise (private preview), Search Live, Gemini Live app, and (for Extended Thinking) Google Workspace (Docs, Gmail,…

20h ago/3 sources/high confidence

CapabilityAccessAdoptionEconomics

69Useful signal

Ranked feed60 live signals

Apple ships Gemini-powered Siri beta with iOS 27, excluding EU and China at launch

Apple released a beta of a rebuilt Siri ('Siri AI') powered by Google Gemini models (hybrid on-device/Private Cloud Compute processing), shipping as part of iOS 27/iPadOS 27/macOS 27/watchOS 27/visionOS 27. It adds screen-content reading, personal-context understanding from messages/photos,…

1d ago/1 source/medium confidence

CapabilityAccessGovernance

Deepseek releases V4.1-Flash, an open-source model that sharply cuts KV cache memory and input-processing compute for AI agents

Deepseek released and open-sourced (MIT license, via Hugging Face) a new model, V4.1-Flash, with 552B total parameters (8B active on input/encoding, 16B active on output/decoding), supporting up to 1M token contexts. It reduces KV cache memory footprint substantially versus its predecessor V4-Flash…

6d ago/1 source/high confidence

CapabilityEconomicsInfrastructure

Controlled study finds no clear performance advantage for vendor-native AI coding harnesses over alternatives, with cost comparisons undermined by missing usage data

A new arXiv paper reports a controlled empirical study isolating the effect of the agent "harness" (tool/prompt/control-flow scaffolding) from the underlying model in agentic coding tasks. Using a private, contamination-controlled suite of 256 tasks, the authors ran paired same-model contrasts…

2d ago/1 source/high confidence

CapabilityEconomicsInfrastructure

Meta launches 'Meta One' subscription bundles combining app subscriptions (Instagram/WhatsApp/Facebook Plus) with extra AI usage, priced $7.99–$499/month

Meta rolled out globally a new tier of subscription bundles called Meta One, combining its existing standalone app subscriptions (Instagram Plus, WhatsApp Plus, Facebook Plus) with additional Meta AI usage (e.g., more image generation via Muse, in-app AI features like Instagram Restyle) and…

22h ago/1 source/medium confidence

EconomicsAccessAdoption

Mistral models now power Mozilla's Firefox Smart Window AI browsing assistant (beta), rolling out in France and North America first

Mozilla's Firefox Smart Window (beta), an AI browsing assistant, is now powered by Mistral models under a new partnership. Rollout starts in France and North America, with UK and Germany to follow later in 2026. Mistral models are fine-tuned on regional languages/dialects; conversations are not…

1h ago/1 source/medium confidence

AccessAdoptionEconomics

Study finds synthetic document finetuning fails to inoculate LLMs against misalignment generalization from reward hacking

Researchers finetuned Llama-3.3-70B-Instruct on about 56K synthetic documents framing reward hacking as acceptable, then trained it with RL on coding tasks with exploitable tests. The model expressed the implanted belief on all behavioral tests but showed stronger misalignment generalization after…

15h ago/1 source/high confidence

Capability

Anthropic threat report: Claude abused for malware, drone/missile software, mass surveillance, and industrial-scale distillation by Chinese AI labs

Anthropic published a threat intelligence report (Dec 2025–Aug 2026) documenting specific misuse cases: a Russian-speaking espionage group (GTG-20006) using self-mutating malware against 20+ organizations; ShinyHunters (GTG-50014) credential mining from 1.8M decompiled Android apps; seven…

4d ago/1 source/high confidence

CapabilityGovernancePower

Enterprises restrict use of Anthropic's Claude (Fable) and other frontier models over data retention and IP concerns

Following Anthropic's June policy change to retain Fable (Claude) usage logs for 30 days, several major enterprise customers have restricted use: Nvidia limits Fable to non-sensitive tasks and uses its own Nemotron models internally; Booz Allen Hamilton banned employees from using Fable for…

19h ago/1 source/medium confidence

AccessAdoptionGovernance

Manhattan DA seizes 12 domains hosting nonconsensual celebrity deepfake pornography

The Manhattan District Attorney's Office executed seizure warrants issued by a New York State Supreme Court to seize 12 domains used to share, publish and sell nonconsensual deepfake pornographic videos of roughly 1,200 people (mostly women, including celebrities, activists, athletes and…

1d ago/1 source/high confidence

GovernanceAccess

Research finds nearly 150 European women MPs targeted by deepfake porn sites, with women 33x more likely to be targeted than men

Researcher Benjamin Shultz (Agora Digitale Transformation / American Sunlight Project) analyzed ~160 domains hosting deepfake abuse content against a database of 5,800+ EU national MPs, finding at least 138 women MPs from 22 EU countries appeared or were mentioned on these sites versus only 9 male…

2d ago/1 source/medium confidence

GovernanceAccessPower

DeepSeek releases V4.1-Flash, a new open-weight model using a novel causal encoder-decoder architecture with separate 8B prefill / 16B decode active parameters

DeepSeek released V4.1-Flash, replacing V4 Pro as its flagship open-weight model. It introduces a new causal encoder-decoder architecture with prefill/decode parameter separation (763B total parameters, 8B active for input/prefill, 16B active for output/decode), native text+image input, 1M-token…

4d ago/1 source/high confidence

CapabilityEconomicsInfrastructure

Metaculus's Spring 2026 FutureEval finds AI forecasting bots nearly match top human forecasters, gap not statistically significant

Metaculus completed and published results of its Spring 2026 FutureEval tournament (Jan 7–Apr 15, 2026): 173 bots (111 non-Metaculus) forecast on 297 scored questions, with a 10-person Metaculus Pro forecaster team benchmarked against the top-10 bot team on 99 overlapping questions. Result: the Pro…

6d ago/1 source/high confidence

CapabilityAdoption

Google DeepMind launches AlphaGenome Atlas, a free public database of predicted effects for 9 billion possible human genome variants

Google DeepMind released AlphaGenome Atlas, a free, publicly accessible platform (website portal, API, and Google Antigravity skill) containing precomputed AlphaGenome model predictions for the molecular effects of all ~9 billion possible single-nucleotide variants in the human genome, a 1-petabyte…

1w ago/2 sources/high confidence

CapabilityInfrastructureAdoption

Proofpoint finds four hacking groups using shared 'BlueMoon' exploit kit chaining two Chrome V8 bugs and a Windows privilege escalation flaw

Security firm Proofpoint disclosed a specific exploit kit ("BlueMoon") that chains two Chromium V8 vulnerabilities (a type confusion bug and a sandbox escape, one tracked as CVE-2026-85046) with a Windows kernel local privilege escalation flaw (CVE-2026-85880) affecting several Windows 10/11/Server…

6d ago/1 source/high confidence

CapabilityInfrastructureGovernance

Google Research introduces ToolGrad, an answer-first framework for generating tool-use training datasets, with fine-tuned Gemma-3 models matching SoTA on BFCL

Google Research published a paper (presented at ACL 2026) and blog post introducing ToolGrad, a new data-generation framework that reverses the standard tool-use dataset creation paradigm by first generating a verified tool-use chain (via an iterative "propose, execute, select, update" loop using…

5d ago/1 source/high confidence

CapabilityEconomics

Researchers release ZGCM-1, a fully open 7B foundation model optimized for math and agentic search efficiency

A new 7B-parameter dense foundation model, ZGCM-1, was trained from scratch and released with full openness: model weights from pre-training, mid-training and post-training stages, intermediate checkpoints, training code, per-stage data/data recipes, and W&B logs. It uses interleaved gated…

1d ago/1 source/high confidence

CapabilityEconomicsInfrastructure

Anthropic reports second straight profitable quarter on adjusted metric, revenue run-rate hits $65B, ahead of planned Nasdaq IPO at possible $2T+ valuation

Anthropic told investors it achieved an adjusted-metric profit for a second consecutive quarter (excluding stock-based compensation and other costs), with quarterly revenue up 14x year-over-year to $11.5 billion and an annualized revenue run rate of $65 billion as of end of July. The company shared…

1d ago/1 source/medium confidence

EconomicsAccess

Bomfather open-sources inode-based caching technique that cuts eBPF security agent's kernel CPU cost by ~90%

The developers of the Bomfather eBPF security agent implemented a memoization cache (keyed on mount namespace ID, mount ID, and inode number) that stores previously-resolved file access policies, avoiding repeated dentry path-walking on repeated file opens. In benchmark tests opening the same file…

1d ago/1 source/high confidence

InfrastructureCapability

Suno releases v6 music model family (v6, v6-wild, v6-mini) built with Warner, BMG, and Believe, retiring all older versions

Suno released its v6 model generation (flagship v6, experimental v6-wild, and free v6-mini), developed with codevelopment partners Warner Music, BMG, and Believe, and retired all previous model versions platform-wide. New capabilities include multimodal generation from text/audio/image/video,…

6d ago/1 source/high confidence

CapabilityAccessGovernance