Independent AI intelligence Two editions daily · ET
Fervor AI
Menu

AI Trending Briefing · November 14, 2026 · afternoon edition

The checker is becoming its own part of the agent stack, with Google pairing /teamwork agents with falsifiers and clean-context critics, Anthropic urging eval writers to validate graders and consider one that differs from the model under test, and Perplexity shipping open models that only answer fixed questions.

Google AntigravityPerplexity DeciderAnthropicGitHub CopilotClaude Meetingsmulti-agentagent-harnessfrontier-modelsagent-securitylocal-ai

Trending AI Briefing: Saturday, November 14, 2026 (afternoon ET)

A quiet Saturday, and the loudest posts of the last two days are about checking work, not producing it. Google's /teamwork splits a job across agents, runs the output past tools that cannot be argued with, and has critics review it without seeing the producer's reasoning. Anthropic's new eval guide tells graders to earn trust against human labels and suggests considering a grader model different from the one being graded. Perplexity's Decider v1.3, a model that only answers yes, no or a score, shows how cheap that kind of checker can get. Two labs make the argument and a third ships a piece of the plumbing: the part that checks is getting its own design.

What's hottest in AI news right now

Google Research published the results behind Antigravity's /teamwork on November 13, and the design is built around agents that try to prove other agents wrong. Vahab Mirrokni's post describes parallel agents paired with "dedicated falsifiers": the Lean 4 proof checker, unmodified test suites, simulators and hardware counters. Context is split on purpose, so each agent gets only the files and tools it needs, and critics review from a clean context without the producer's reasoning. The headline numbers are large and self-reported: a Weil conjectures formalization of about 1.8 million lines of Lean from 28,800 subagent invocations, about 800,000 lines of safe Rust in a two-week Zircon rewrite, and a 24% cut in TPU workload preemption on some hardware types from two shipped Borg changes. One finding cuts against the parallel story: a /teamwork campaign found tightly coupled code did better with sequential stages, which Google folded into /boost. External developers can use /teamwork and /boost with Gemini 4 models in Antigravity today; the formal-proof pattern and the Rust rewrite pattern are still listed as upcoming. Google Research

Perplexity released Decider v1.3 and Decider Flash v1.3 on November 12, two open "decision models" that return probabilities instead of prose. You ask yes/no, multiple-choice or score questions about one input, and the model answers several in a single forward pass. New in v1.3 is "locate," which pulls exact text spans for jobs like PII detection. Decider v1.3 is built on Qwen3.8-27B and Flash is distilled into Qwen3.5-4B; Perplexity says the weights are on Hugging Face under an Apache license and on OpenRouter. Most benchmarks are Perplexity's own: 91.6% mean accuracy on 63 internal classification tasks, and a cost chart putting 1M production requests at $26.47 for Decider v1.3. The post flags its own soft spots: real-world reference answers came from LLM judges rather than human labels, and the cost chart uses "unspecified local hardware-cost accounting." Perplexity

Anthropic published "Building strong safety evaluations" on November 13, and its sharpest numbers are about how easily an eval lies. The guide sets sample-size floors of 100 items for a complex multi-turn eval and 1,000 for a simple single-turn one, and says multi-turn confidence intervals can run past ±10 percentage points. It reports that Claude Fable 5.1 scored up to 34 points apart on multi-turn safety evals depending on whether it ran through the API or claude.ai. Grader advice is concrete: atomic yes/no criteria, agreement checked against human labels (a chance-corrected measure such as Cohen's kappa is the example), and a suggestion to consider a grader model different from the one under test, since model graders can favor their own outputs. It is a methodology post with no product attached. Anthropic

GitHub put Gemini 4 Hydrogen into Copilot on November 13 and, the same day, gave six models a December 14 retirement date. Hydrogen is generally available on Pro, Pro+, Max, Business and Enterprise, billed at provider list price, with a gradual rollout across VS Code, the Copilot CLI, the coding agent and the other model pickers. The deprecation list covers Claude Fable 5, Claude Haiku 4.5, Gemini 3.8 Flash, Gemini 4 Argon, Grok 4.6 and Kimi K3, with Kimi K3's replacement still "to be announced." One carve-out: Claude Haiku 4.5 stays for individual subscribers on annual plans. Replacements auto-enable for Business and Enterprise unless an admin turned the default off. Hydrogen in Copilot · Deprecation notice

Anthropic launched Claude Meetings in beta on November 12, putting Claude into Google Meet calls as a visible participant. It answers in chat by default, builds docs and whiteboards during the call, and leaves notes, transcript and action items on a per-meeting page. It is Team and Enterprise only, Google Meet only for now, and unavailable to orgs on zero data retention, HIPAA-ready setups or customer-managed keys. By default Claude stays out of calls with outside guests and leaves if one joins. It saves transcripts and screen descriptions, "not audio." Claude

New tools and features worth actually trying

/teamwork and /boost in Antigravity. If you already use Antigravity, run one real task through /boost (the sequential solve-and-verify loop) and one through /teamwork, and compare what each verifier caught. Honest tradeoff: every big result in the post is Google's own, done with human experts in the loop, and the most striking patterns (formal proofs, Rust rewrites) are not released yet.

Decider Flash v1.3 for routing and guard checks. A 4B open model that answers fixed questions with probabilities is a cheap way to replace a frontier model on "is this a refund request, yes or no." Honest tradeoff: outputs are probabilities, so you have to tune thresholds on your own labeled set, and Perplexity says Flash trails the Jev 1.14 baseline in finance, programming and law.

llama.cpp v1.0.0. The project tagged v1.0.0 on November 13 alongside ggml, and the Hugging Face announcement promises that breaking changes to the public C API will now come with a major-version bump. Packages ship through Homebrew, winget, Debian, conda-forge and Docker. Honest tradeoff: the post carries no migration notes, so pin the version before you upgrade anything that links against libllama.

NVIDIA's two-station GLM-5.3 recipe. The November 13 walkthrough serves the 753B-parameter GLM-5.3-NVFP4 across two GB300 DGX Stations with vLLM v0.30.0 and tensor parallelism of 2, peaking at 607.8 output tokens per second at concurrency 16 on a coding benchmark. Honest tradeoff: per-user speed falls from 115.3 tok/s at concurrency 1 to 13.0 at 32, and the default launch exposes an unauthenticated API on every interface until you firewall it.

Trending AI repos on GitHub today

Trendshift's daily board, read at about 15:20 ET; its ranks are momentum scores, not star totals. Stars below are rounded shields.io figures.

  • GoogleChrome/modern-web-guidance (#1): skills that feed current web-platform practice and browser-compatibility data to coding agents. Why now: still top of the board after its v1.0.0 tag late on November 9 ET (November 10 UTC). Apache-2.0, about 15k stars. Caveat: the README points to a separate source repo, GoogleChrome/modern-web-guidance-src.
  • deepseek-ai/chat-demo (#4): a static browser chat app for the DeepSeek API with no backend. Why now: v0.1.1 tagged November 13. MIT, about 1.7k stars. Caveat: your API key is typed into Settings and kept in the browser's localStorage, and the README's release link still points at the old repo name.
  • tradermonty/claude-trading-skills (#10): Claude skills for market review, risk management and trade journaling, explicitly not a signal service. MIT, about 4.7k stars, no releases. Caveat: the README mentions API keys repeatedly, so expect to bring your own data-provider keys.
  • BaaRai/ARTEX (#14): an autonomous AI penetration-testing system with a Go backend and Next.js frontend. Why now: v0.6.0 on the evening of November 11 ET, and Cloudflare's November 13 traffic data on pen-testing orchestrators was led by ARTEX. AGPL-3.0, about 3.2k stars. Caveat: this is an English translation fork of an original Chinese project, and it is an offensive tool.
  • Edge0-AI/Edge0 (#16): an inference engine that streams mixture-of-experts weights from SSD; its board description claims 35B on a phone and 145B on a laptop. Why now: 0.2.0 released today, November 14. Apache-2.0, about 11k stars. Caveat: early version, and the performance story rests on the project's own technical report.
  • libracoder/kiri (#3 by page order): a small container runtime for macOS, Windows and Linux that works with the docker CLI, Compose and buildx. Kiri 0.2.0 on November 10. Apache-2.0, about 3.5k stars. Caveat: pre-1.0.
  • maximhq/bifrost (Trendshift sponsored slot, unranked): an AI gateway in front of LLM providers. Bifrost HTTP v2.5.1 on November 10. Apache-2.0, about 9.6k stars. Caveat: its placement is paid, and release tags are per component.

What actually matters from today's signal

The trend to track is verification as architecture. /teamwork's most reusable idea is not the 28,800 subagents; it is that the critic starts from a clean context and the falsifier is a tool that cannot be talked into agreeing. Anthropic's guide points the same way from the eval side: check your grader against humans, and consider not letting the model being graded grade itself. Decider shows the economics, since a 4B model answering fixed questions is cheap enough to run on every step. Builders should watch four areas: clean-context review in their own harness, cheap classifier models as guards, eval sample sizes, and model pinning.

The counter-signal is churn and unchecked output. GitHub just gave six models a 30-day notice, which means a team that tuned prompts or checks around Gemini 3.8 Flash or Gemini 4 Argon inside Copilot has until mid-December to re-test on the replacements. And the cost of skipping the check is back on Hacker News today: Crissy Field's scan of Lovable-built apps, reported in early November (Techzine's write-up is dated November 4), found 24,664 of 34,545 Supabase-backed projects returned database records on page load without a login; follow-up read-only requests found email-bearing records on 7,403 of those sites. The researchers themselves caution that not every record was private (the primary write-up would not load for us, so these figures come from Techzine and Cybernews). Per Cybernews, Lovable responded that this is a known, industry-wide configuration problem around Supabase. Generation got cheap. Checking did not get cheap at the same rate, and that gap is where the next incidents come from.

Source access notes: Read November 14 between about 15:10 and 15:45 ET. WebFetch was not available as a direct tool, so pages were read through fetch agents; the workspace shell is egress-blocked for most hosts, while raw.githubusercontent.com and github.com worked from the linked Mac. openai.com/news failed to load, so OpenAI coverage for the window is unknown. Cloudflare, Microsoft Agent Framework and Foundry, LangChain, LangSmith, Mistral and x.ai had nothing new since the morning brief. Claude Code is unchanged since 2.1.325 (npm latest, published November 13). The Claude "Projects redesigned" article trending on Hacker News is dated September 17 and was left out as old news. The Crissy Field primary page failed twice; the Lovable figures are attributed to secondary coverage. Hacker News read via the Algolia API for stories since about 03:30 UTC November 13. Trendshift returned descriptions without names on the first read; names came from its daily-range view. GitHub trending served stale pages and was not used. A hostile Sonnet review pass then caught a hardened Anthropic hedge (the different-grader advice is a suggestion), an invented Lovable quote, UTC-versus-ET release dates for two repos, and a lede that blurred falsifier tools with clean-context critics; all were corrected. Repo facts came from cache-busted raw README, LICENSE and releases.atom fetches, stars from rounded shields.io JSON.