No. 3 ·

The Oct 31 edition

Window
2026-10-30 07:30 → 2026-10-31 07:30 (America/Chicago, CDT). Older items that you can still act on are kept and labeled.
Method
discovery from aggregators, confirmation from official sources, rubric score >=5, no read-time cap
Rule
every topic, story, and mention is a clickable markdown link
  • 204 primary links
  • 9 sections
  • 20 min read
  • Still open8
  • News21
  • Update4
  • Context3

Labels: News = first seen in the window · Update = older story with a new number or artifact · Context = background you need to act · Still open = not new, but still worth trying

Snapshot#

  • Quiet day for models. Price Per Token model releases shows 0 in the last 24h. The OpenRouter models list has no new IDs; the newest is still particle-1 (10-28). The only new model eval on Artificial Analysis is DeepSeek V4.1 Flash 1017 (Non-reasoning). Today's action is in security, harnesses, and plan economics.
  • News (security): Northflank's status page says to assume all environment variables were decrypted and to rotate every secret. At 18:38 CDT on 10-30 it added that users are reporting unauthorized use of AI API keys they had stored on the platform. Wiz's "Black October" report counts at least nine AI developer platforms with incidents this month, OpenCode (Oct 16–17) among them, and documents stolen keys being resold. HN thread: 448 points.
  • News Claude Code v2.1.316 (10-30 20:01 CDT) deprecates fast mode for Opus 4.8. It will be removed on Nov 30, 2026, and those requests will then fail. It also adds CLAUDE_CODE_SUBAGENT_EFFORT, pooled gateway budgets, and a fix so --restricted sessions no longer upload files from outside the working directories.
  • News Anthropic renamed its consumer plans. Max 20x is now Ultra at $200/mo: 10x Pro usage per week, no five-hour limit for verified accounts, and monthly API credits. claude.com/pricing lists $100 in API credits for Max and $200 for Ultra.
  • News Factory shipped Droid 1.0. Its main new feature is programmatic tool calling: a Script tool runs JS in a sandbox, and it is on by default. Factory's own Terminal-Bench 4 runs put it above Codex and Claude Code at lower cost. These are vendor-run numbers.
  • News (tool): Microsoft's TokenEconomizer is an Apache-2.0 plugin for Claude Code, Codex, Copilot CLI and Pi. Its paper reports 18.4–30.9% lower mean cost on SWE-Bench tasks. Vercel Labs' agent-desktop v1.0.2 does accessibility-tree computer use from a CLI.
  • News opencode v2.2.4 and v2.2.5 shipped the same day. Highlights: persistent desktop terminals, experimental artifacts, guided Claude Agent setup, and a percentage-based compaction buffer.
  • Do this today: if you have any AI provider keys stored on Northflank or another platform named in the Wiz report, revoke and reissue them, then check your provider billing. If you use Opus 4.8 fast mode anywhere, schedule the move to Opus 5.5 before Nov 30 (v2.1.316 notes).

Frontier and open-weight models#

DeepSeek V4.1 Flash 1017 (Non-reasoning) on AA — Update (AA eval 2026-10-30)#

  • What: AA's changelog added a new eval for the non-reasoning mode of DeepSeek V4.1 Flash 1017. AA lists 552B total / 16B active parameters, MIT license, 1M context, and text plus image input.
  • Numbers (AA):
    • Intelligence Index 19 (median 19).
    • 1,486 output tok/s, #1 of 25 on speed.
    • 51M output tokens to run the index.
    • Cost per task is listed as N/A.
  • Price disagreement: AA lists $0.45 in / $1.80 out per 1M tokens with a 98% cache discount. OpenRouter deepseek/deepseek-v4.1-flash-1017 lists $0.15 / $0.60 ($0.003 cache read). Price your own provider route rather than trusting either list.
  • Builder take: Non-reasoning mode is a speed play for bulk extraction, routing and classification, not agent work. For agentic coding, the reasoning "Max" config is the one on the Pareto table (section 3).
  • Local: ISTA-DASLab GSQ GGUF is trending on HF (511 likes). The base repo has 4,060 likes (HF trending).

AA-Omniscience v1.1 — News (2026-10-30, ~19:01 CDT)#

  • AA refreshed its knowledge and hallucination benchmark. It now has 3,000 questions, down from 6,000, with saturated questions removed and ambiguity fixed. A new public set on HF is calibrated to the private set.
  • AA's thread reports 0.98 Spearman rank correlation with v1 and a median Net Score drop of 0.6 points. AA attributes the drop to the dataset change, not model regression.
  • Builder take: If your RAG or agent picks models by hallucination rate, re-read the board. The public set is a usable local factuality check.

Anthropic: automated subtle emergent-misalignment search — News (research, Oct 2026)#

  • Anthropic had a Claude agent search for fine-tuning data that a production-like training-data monitor would not flag but that still made a small research model misaligned. Given enough attempts, it succeeded. Anthropic frames this as a monitor gap to fix, not a finding about deployed Claude.
  • Builder take: If you fine-tune on third-party or synthetic data and rely on per-row content screening, this is evidence that row-level screening alone can miss a dataset-level pattern. Add behavioral evals on the tuned model.

Still open#

Pareto frontier#

Coding Agent Index (Artificial Analysis, v2.0: FrontierCode v1.1 + Terminal-Bench 4.0 + SWE-Atlas-QnA-AA). The data was materialized 2026-10-29 23:46 CDT, so the rows are unchanged since yesterday's read. Cost is mean USD per task. Cache-hit and tokens/task are shown where I read them.

Harness + model Index $/task Cache hit · tokens/task Note
Claude Code + Opus 5.5 (Max) 61.7 $11.42 95.3% · 14.5M Top score
Devin Fusion + Opus 5.5 Max / GPT-6 Luna High 61.5 $8.00
GitHub Copilot + Copilot Fusion (max) 59.9 $5.59 Best $/point near the top
Antigravity + Gemini 4 Carbon (High) 59.4 $7.74 Limited access
Claude Code + Sonnet 5.5 (Max) 59.3 $13.30 · 24.1M Costs more per task than Opus here
Codex + GPT-6 Astra (Max) 54.7 $7.04
Codex + GPT-6.1 Sol (Max) 54.1 $1.50 Strong mid-tier $/point
Opencode + GLM-5.3 (Max) 48.9 $5.23 CN open-weight
Mistral Vibe + Mistral Large 4 (High) 48.3 $1.26 91.2% · 8.5M Open weights
Codex + GPT-6.1 Terra (Xhigh) 48.2 $0.69
Codex + DeepSeek V4.1 Flash (Max) 47.3 $0.39 98.3% · 27.2M CN open-weight
DeepSeek Harness + V4.1 Flash 1017 (Max) 46.3 $0.30 96.8% · 16.7M Cheapest row above 45
Codex + GPT-6.1 Terra (Medium) 42.0 $0.39
Kimi Code CLI + Kimi K3 41.4 $5.17 CN open-weight, expensive per point
Mistral Vibe + Mistral Large 4 (Medium) 35.4 $0.78 92.3% · 5.8M
Codex + GPT-6 Luna (Max) 31.1 $0.16 Floor

Read it like this:

  • Frontier tier (≥59): Copilot Fusion costs about half of Claude Code + Opus per task for 1.8 fewer points.
  • Value tier (~46–48): DeepSeek Harness + V4.1 Flash 1017 at $0.30, Codex + DeepSeek V4.1 Flash at $0.39, and Codex + Terra Xhigh at $0.69. All are within about 2 points of Mistral Large 4 High at $1.26.
  • Session economics: the DeepSeek rows burn far more tokens per task (16.7M–27.2M) but stay cheap because of cache hit rates of 96.8–98.3%. If your provider route doesn't discount cached input, the ranking changes.

Session economics and plan moves (today)

  • Claude plans: see the Ultra support article and claude.com/pricing.
    • Ultra ($200/mo): 10x Pro weekly usage, no 5-hour window for verified accounts, $200/mo in API credits.
    • Max: 5x Pro per 5 hours and per week, $100 in API credits.
    • Both plans allow Fable for up to half the weekly limit.
    • For long unattended agent runs, Ultra's lack of a 5-hour window matters more than its headline multiplier. Check your actual weekly burn before you assume it is an upgrade.
  • Factory Droid 1.0 (post, vendor-run, Oct 23–28, 3 trials per task):
    • On Terminal-Bench 4 with GPT-6.1 Sol: 3.5 points above Codex at 12% lower cost.
    • With Opus 5.5: 3.0–7.6 points above Cursor, OpenCode, Devin and Claude Code at up to 24% lower cost.
    • 91.5% prompt-cache reuse by tokens.
    • Droid is not on AA's Coding Agent Index yet, so there is no independent row to compare.
  • TokenEconomizer (repo, paper):
    • 18.4–30.9% lower mean cost across five agent–model pairs on 277 SWE-Bench tasks, with Pass@1 up in 4 of 5.
    • It is a session-cost lever that works on top of any row in the table.
  • OpenRouter usage (rankings, data through Oct 30):
    • Top apps by tokens: Hermes Agent 1.84T, Claude Code 1.57T, Kilo Code 714B, OpenClaw 708B, Codex 698B.
    • DeepSeek Harness (286B) and Strix (298B) are flagged as new entrants.

Index disagreement:

  • The DeepSeek V4.1 Flash 1017 price: AA lists $0.45/$1.80 and OpenRouter lists $0.15/$0.60.
  • Vendor claims vs AA: Factory's harness claims come from Terminal-Bench 4 runs it did itself, while AA's index combines three benchmarks.

Coding CLIs, IDEs and harnesses#

Claude Code v2.1.316 — News (10-30 20:01 CDT)#

  • Deprecation: fast mode for Opus 4.8 is deprecated as of Oct 30. Removal is Nov 30, 2026, with no fallback: requests will error. Migrate to Opus 5.5.
  • New env vars:
    • CLAUDE_CODE_SUBAGENT_EFFORT sets the effort level for subagents.
    • CLAUDE_CODE_DISABLE_ULTRATHINK.
    • CLAUDE_CODE_DISABLE_LOOPED_CROSS_CERT_DROP.
  • Gateway:
    • Pooled group budgets (pooled_budget).
    • managed.policy_mode: layered.
    • Bedrock guardrail.input_tagging.
    • Every model named in a request is now checked against allow/deny lists (400 on violation).
  • Fixes:
    • Shell commands left running at exit.
    • --restricted sessions could upload files from outside the working directories when a Read allow rule covered them.
  • Note: npm also shows 2.1.315 (published 10-30 16:05 CDT) with no GitHub release notes. Go straight to 2.1.316.

Factory Droid 1.0 — News (2026-10-29)#

  • What: A frontier-model harness. Programmatic tool calling is the default: the model writes JS that runs in a sandboxed Script tool, instead of making one tool call per step.
  • Vendor numbers:
    • On Opus 5.5: 26% less time per passed task than Claude Code, using 17% fewer tokens.
    • On Sol: 12% fewer tokens than Codex, but 10% longer.
    • Compared versions: codex@0.163.0, claude-code@2.1.306, cursor-cli 2026.10.19, devin 3000.11.92 and opencode 2.1.2. All of these are older than today's releases.
  • Try it:
  • Builder take: Programmatic tool calling is the idea to steal even if you don't switch harnesses. Batching tool calls into one script cuts round-trips and tokens. Verify the claims on your own tasks.

opencode v2.2.4 + v2.2.5 — News (10-30 11:59 and 18:17 CDT)#

  • v2.2.4:
    • Persistent desktop terminals.
    • Experimental artifacts ("experimental": {"artifacts": true}).
    • Guided Claude Agent setup.
    • Schedules run under opencode serve.
    • Higher OpenCode Go limits for DeepSeek V4.1 Flash.
    • Local service requests no longer go through env proxies.
  • v2.2.5:
    • Percentage compaction buffer ("compaction": {"buffer": "20%"}).
    • MCP stdio messages over 10 MiB no longer disconnect.
    • LiteLLM context-window errors now trigger compaction.
    • A tool named bash maps to shell.
  • Context: Wiz lists OpenCode among this month's platform incidents (Oct 16–17). If you put provider keys into OpenCode's hosted services before then, rotate them.

Vercel AI SDK fx harness adapter — News (2026-10-29)#

  • @ai-sdk/harness-fx now uses libfx. The harness loop and model requests run on the host, while filesystem and shell operations stay in the sandbox, so model credentials never enter the sandbox. You no longer need fx or a bridge inside the sandbox.
  • It also adds tool filtering, mid-turn steering, structured output, image tool results, and save/resume sessions.
  • Given this week's credential-theft news, this is the right shape for any sandboxed agent you build.

No change#

  • Codex CLI is still at 0.166.0.
  • gemini-cli shipped only a nightly (v0.67.0-nightly.20261031); stable is v0.66.0 (10-27).
  • pi is still at v1.4.0.
  • aider is still at v0.86.0.
  • DeepSeek Harness' latest is still rc.5 (10-29).

Update: Cursor-marketplace trojanized Claude Code extension#

Usable tools and open projects#

Repo What Stars / +24h Verdict Why
microsoft/TokenEconomizer Plugin that learns per-repo "micro-skills", prunes unused tools, and nudges stuck loops; Claude Code, Codex, Copilot CLI, Pi (opencode experimental). Apache-2.0 2,534 / n/a Runnable Measured 18.4–30.9% cost cut on SWE-Bench (paper); state stays local in ~/.tokeneconomizer
vercel-labs/agent-desktop CLI + C library for desktop computer use through the OS accessibility tree; JSON snapshots with refs; headless by default. Apache-2.0 3,251 / n/a Runnable npm i -g agent-desktop; macOS, Windows, Linux (at-spi2); v1.0.2 10-30
spinfinc/spinf-serve One-GPU (24 GB, fp8) server that reads content once and scores many questions as probabilities; OpenAI-compatible. Apache-2.0 19 / n/a Runnable Decision-model serving for bulk classification. Show HN, 22 points
dbreunig/nesformer NES 6502 CPU compiled (not trained) into stock Qwen3.5 weights; Gated DeltaNet state used as memory 212 / n/a Watch (learn-from) Simon Willison; HN, 257 points; builds on Qwenputer
Orkas-AI/Orkas Local-first desktop "AI team": Commander + specialists, delegates to Claude Code/Codex/OpenCode/OpenClaw/Hermes. MIT 5,871 / +1,004 Watch New on trending; bring-your-own-keys or local endpoint
mainplane-ai/mainplane Self-hosted agent harness: server holds secrets, workers are your machines, shared drives, computer use. MIT 582 / n/a Watch Keeps secrets on a server you control; nominated by AlphaSignal
aditi-p31/quantcheck Functional acceptance test for quantized GGUF/Ollama artifacts 0 / n/a Runnable Paper found 7 of 305 official Ollama code tags (2.3%) non-functional (2609.05881)
thedotmack/claude-mem Persistent cross-session memory for agents 107,048 / +1,723 Still open · Runnable Covered 10-30
pingdotgg/t3code Coding-agent GUI 33,619 / +1,148 Still open · Watch Covered 10-30
milind-soni/OpenMausBot Open agent with a VM 9,214 / +1,304 Still open · Watch Covered 10-30
lightpanda-io/browser Headless browser for agents (AGPL-3.0) 38,801 / +929 Still open · Runnable Covered 10-30
Donchitos/Claude-Code-Game-Studios, mattpocock/skills, obra/superpowers, pbakaus/impeccable Skill/prompt packs +1,347 / +882 / +663 / +1,339 Skip Prompt packs, not tools

Stars are GitHub API totals and "+24h" values from GitHub trending daily, read 2026-10-31 morning CDT. "n/a" means the repo was not on today's trending page.

Agent builder stack#

  • News: AgentCore Runtime Session endpoints (AWS, 10-30). Each agent session gets an authenticated URL that forwards HTTP and WebSocket traffic to the ports you allow (allowedPorts). You can watch a dev server (Vite in the example) update live while the agent edits it. The URL stops working when the session ends.
  • News: NVIDIA NemoClaw blueprints (repo, quickstart) (10-30).
    • Open blueprints that combine a harness (OpenClaw, Hermes Agent, or LangChain Deep Agents) with Nemotron models and the OpenShell runtime. OpenShell enforces policy on file, network, credential and tool access.
    • NVIDIA says the Deep Agents harness tuned for Nemotron 3 Ultra reaches near-frontier performance at about 10x lower cost per run. That is vendor-reported.
  • News: MCP vs CLI, matched (MICA) (code).
    • The study exposes the same operations through MCP and through a CLI. Six models in three environments show substantial differences in task success and cost, and 94.3% of 2,004 paired episodes diverge structurally.
    • Its DART repair step improves scores in all six settings by up to 9.0 points.
    • Take: interface rendering is a tuning knob. A/B your tool surface; don't assume MCP and CLI are equivalent.
  • Context: the Wiz Black October defensive lesson. In one of the two cases Wiz describes, the key's policy allowed only regional inference profiles. The first wave of abuse failed and much of the rest was throttled by regional quotas, though over a million calls still succeeded. Scope cloud model keys tightly and set quotas and billing alerts. Keep provider keys out of the platforms where your agents run where you can, as with the fx host/sandbox split above.
  • Context: MoEngage put 400+ APIs behind AgentCore Gateway (AWS case study, 10-29). A reference for turning a large internal API surface into agent tools.

Infrastructure dash#

  • News (incident): Northflank. All times CDT, converted from UTC on the status page.
    • 10-30 06:43: Wiz threat alert.
    • 10-30 08:40: told users to rotate environment variables.
    • 10-30 08:58: "Assume all environment variables have been decrypted."
    • 10-30 10:12: secret rotation enabled for Postgres and MongoDB addons; disable public addon exposure.
    • 10-30 18:38: reports of unauthorized use of AI API keys.
    • 10-30 20:04: audit logs enabled for all customers, including free.
    • The status page still showed a major outage across networking, services, addons, jobs and builds when read this morning.
  • News: TensorRT Edge-LLM on DGX Spark (NVIDIA, 10-30; repo, docs).
    • pip install "tensorrt-edgellm[server]==0.11.1", then trtedge chat nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4. trtedge serve exposes an OpenAI-compatible endpoint.
    • NVIDIA reports DSpark speculative decoding (3 draft tokens) raising mean speed from 73.4 to 126.9 tok/s on HumanEval prompts, with 27.9 GB peak memory.
    • The same runtime targets Jetson/DRIVE/IGX Thor. Model: Nemotron-3.5 Lightning 30B-A3B NVFP4.
  • News: Dynamo 1.4.2 via NVIDIA AI Cluster Runtime (AICR) (10-30). Pinned, validated recipes for Dynamo on Kubernetes (DRA, gang scheduling, health signals, Prometheus). It routes using KV-cache awareness through either the native Dynamo Frontend router or the Gateway API Inference Extension/EPP path. See the Dynamo repo.
  • News (research): Soft-MTP, "Multi-Token Prediction Done Right" (code, MIT). A one-line change to multi-token-prediction training. It keeps only the KL self-distillation signal, which helps speculative decoding, with no extra parameters or overhead. It is tested from 167M to 1.2B parameters and supports both DeepSeek-style and Meta-style MTP. Relevant if you train or post-train draft heads.
  • News (research): Broken on Arrival. Covered in section 5 (quantcheck). Low-bit quants of small (≤4B) coding models are the risk zone. Adding an importance matrix fixed three Qwen2.5-Coder-3B builds.
  • Learn-from: SkyPilot RL "flying roadster" (example). A Qwen3-4B RL-tuned pilot landed 347 of 384 storm kickflips against GPT-6 Astra's 286 (21% more). Training, rollouts and evals run as one SkyPilot Job Group. The writeup is clear about real-time latency limits.
  • Engine blogs (vLLM, SGLang, TRT-LLM, InferenceX): not nominated by HN. The HN inference slice surfaced only spinf-serve.

Builder actions#

  1. Rotate exposed AI keys. If any provider keys live on Northflank or a platform listed by Wiz, revoke them at the provider, reissue, and update workloads. Then review provider usage and billing for the last 30 days. Scope cloud model keys to the regions and models you actually use, and set budget alerts.
  2. Upgrade Claude Code and plan the Opus 4.8 fast-mode exit.
    • npm i -g @anthropic-ai/claude-code@2.1.316 (notes).
    • Search your Claude Code settings, scripts and CI config for Opus 4.8 fast-mode use and move it to Opus 5.5 before Nov 30.
    • Try CLAUDE_CODE_SUBAGENT_EFFORT to run subagents at lower effort.
  3. Measure TokenEconomizer on a repo you use daily.
    • git clone https://github.com/microsoft/TokenEconomizer.git && claude plugin marketplace add "$PWD/TokenEconomizer" && claude plugin install tokeneconomizer@tokeneconomizer. Codex uses codex plugin marketplace add + codex plugin add, and Pi uses pi install git:github.com/microsoft/TokenEconomizer; see the README.
    • Compare cost over 5+ sessions against your baseline.
  4. Give an agent desktop eyes. npm i -g agent-desktop && agent-desktop snapshot --app Finder -i, then run agent-desktop skills get desktop to get the agent-side contract (repo). Keep snapshots out of public logs, since they can contain private screen content.
  5. Tune opencode compaction. Upgrade to v2.2.5 and set "compaction": {"buffer": "20%"}. Turn on "experimental": {"artifacts": true} to try artifacts (v2.2.4).
  6. Steal programmatic tool calling. Try Droid 1.0 on three of your own tasks against your current harness, with the same model, and record tokens and wall time. Factory's numbers are vendor-run (post).
  7. Move credentials out of your sandbox. If you build on the AI SDK, update @ai-sdk/harness-fx so model calls run on the host and only filesystem and shell run in the sandbox (changelog).
  8. Bulk-classify on one GPU. On a 24 GB fp8-capable card: pip install "spinf-serve[server]" && spinf-serve serve spinf/gemma-4-12b-fp8, then POST to /v1/score (README).
  9. Screen your local quants. Before trusting a ≤4B low-bit GGUF for code, run quantcheck against it.
  10. Re-check Claude plan fit. If you're on Max 20x (now Ultra), compare your weekly usage with "10x Pro weekly, no 5-hour limit" on claude.com/pricing. Remember the included API credits ($200 Ultra / $100 Max).
  11. Edge box? If you have a DGX Spark, run trtedge chat nvidia/NVIDIA-Nemotron-3.5-Lightning-30B-A3B-NVFP4 and compare the DSpark draft variant (NVIDIA post, docs).
  12. Still open export from Jules before Nov 11 (read-only after that).

Checkmarks stay in your browser only.

Source log#

Show the full source logEvery discovery source checked, hits and misses, and every confirm open.

Primary discovery

Source Result Note
Artificial Analysis home Hit Changelog: DeepSeek V4.1 Flash 1017 (Non-reasoning) eval, 30 Oct
AA articles Partial No new articles since 10-28
AA Coding Agent Index Hit Materialized 10-29 23:46 CDT; full table parsed, unchanged vs yesterday
OpenRouter rankings Hit Top apps through Oct 30 (Hermes Agent, Claude Code, Kilo Code…); benchmarks tab read
OpenRouter models API Miss No new IDs in the window (newest particle-1, 10-28)
HN front page Hit Claude Code extension backdoor (carryover), Northflank breach, nesformer, Scribd breach
HN Algolia last 24h (AI/LLM/agent keywords) Hit Claude plan rename, SkyPilot RL, Wiz Black October, GitLab Orbit, Google AI-training control
HN Algolia inference slice (vllm, sglang, trt-llm, tensorrt-llm, mooncake, "kv cache", inference) Partial Only Show HN spinf-serve
GitHub trending daily Hit Orkas new; star totals checked against GitHub API
HF Daily Papers 2026-10-31 Miss Saturday; API returned an empty list
HF Daily Papers 2026-10-30 Hit 39 papers; MICA, Soft-MTP, Broken on Arrival, Investor's Last Exam
HF models trending Hit No major new release; DeepSeek V4.1 Flash 1017, Mistral-Large-4, Nanbeige4.3-33B, LFM3
Price Per Token news + /agents, /coding, /open-source, /model-releases Hit Model releases last 24h: 0. Nominated AgentCore Session endpoints, Bedrock FinOps, Nemotron on EKS
LearnPrompt/ai-news-radar data JSON (data/latest-24h.json, master) Hit Generated 2026-10-31 05:55 CDT, 247 items
RadarAI Hit NVIDIA Edge-LLM, Cursor backdoor (Ars), Claude Ultra, Google PageBreak
The Agentic Digest Partial No Saturday issue yet; latest is 10-30 (StepSecurity analysis, MoEngage AgentCore Gateway)
Simon Willison 2026-10-30 Hit One post: nesformer
X (narrow: AA, karpathy, swyx, simonw, cursor_ai, claudeai, OpenAI, AIatMeta, huggingface) Hit AA-Omniscience v1.1 thread; rest chatter

Dailies in rank order

  1. Latent Space AINews: Partial. Titles only (latest: Google AI Mode / Search revenue); market narrative, nothing printed.
  2. AlphaSignal: Partial. Several items are Pro-gated. Nominated the Anthropic monitor research, AA-Omniscience v1.1, Mainplane, and AutoTrust JEV-9B.
  3. Code4U AI Digest: Partial. Latest date 2026-10-30, 8 items, mostly industry or already covered (Copilot Fusion GA, Gemini 4 Carbon).
  4. TLDR AI archives: Hit. Opened the 2026-10-30 issue. Nominated Factory Droid 1.0, the Vercel fx adapter, OpenAI's Model Spec update, and Raindrop.
  5. TLDR flagship: Not checked. Only used if TLDR AI is blocked, and it wasn't.
  6. The Neuron: Partial. Latest is 10-30 (Mistral 1T open model), a carryover.
  7. AI Daily Brief archive: Partial. Latest is 2026-10-30 ("How the Most AI-Pilled Company in the World Works"); nothing printed.

Confirm opens: nominated → confirmed

Dropped

  • AutoTrust JEV-9B (AlphaSignal): the HF repo was created 2026-09-23. It's a stale recap and scored below 5.
  • Xiaomi-OCR-0 (LearnPrompt via X): the HF repo is from 2026-09-29. Stale, dropped.
  • DSH Go (DeepSeek Harness on Android; LearnPrompt via waytoagi): nominated, no primary, dropped.
  • Replit sandbox whitepaper (LearnPrompt via X): nominated, no primary, dropped.
  • "Anthropic cracking down on video game rewrites" (HN via X): nominated, no primary, dropped.
  • Fish Audio Drama 3 (TLDR AI, X only): nominated, no primary, dropped.
  • Qwen + Apple Intelligence testing (LearnPrompt via WeChat): consumer; nominated, no primary, dropped.
  • Nemotron-halloween (PPT → HF blog): the page had no substantive content. Dropped.
  • Google shared AI-training opt-out (Search Central blog), AI scraping pledge, Cisco unlimited tokens, Anthropic/OpenAI finance, Wired Pentagon contractor, Scribd breach, Gitee credential exposure: scored below 5 (policy, finance, or not builder-actionable).
  • "We took over ChatGPT accounts" (HN): a security write-up. Not printed, and the method is not described.
  • Epoch patch-gap report, Meta JiTTesting, Copilot Fusion GA, Gemini 4 Carbon (repeated in TLDR AI and Code4U): already covered 10-29/10-30. They appear only as Still open or as Pareto rows.
  • Raindrop agent-swarm leadership, OpenAI Model Spec update, GitLab Orbit, OpenRouter always-on agent tutorial, Investor's Last Exam: opened or seen, but scored below 5 for builders today.

Engine blogs: vLLM, SGLang, TRT-LLM and InferenceX were not nominated. The NVIDIA TensorRT Edge-LLM and Dynamo posts were nominated by RadarAI and LearnPrompt and confirmed on the NVIDIA developer blog (section 7).