01
Anthropic's 2030 economy explorer vs 10k Americans
Anthropic's Economics team ships an interactive 2030 scenario model; plug in your capability bets and compare them to more than 10,000 surveyed Americans.
What they showed / shipped
- Anthropic Economics: explore how AI might hit growth, jobs, and wages by 2030, then see how your answers compare to 10k+ Americans (@AnthropicAI; explorer).
- First-party numbers from their August survey (n~10,980 public): the typical respondent lands near a "substantial change" path. GDP about 10% higher by 2030 vs no-AI baseline, unemployment around 5%. Roughly 10% of answers line up with an extreme scenario.
- The artifact is the explorer itself: set capabilities, adoption, autonomy, productivity, and job-adjustment speed, then watch GDP and labor outcomes move.
- 📌 On the radar: Massachusetts just hit data centers with new clean-power rules (TechCrunch). Infra cost side of the same economy beat.
Why it matters
- This is a teachable model, not a vibes take. You can put the dials on screen and show what each bet implies.
Sources
02
Geiger inventories every AI agent on your machine
Show HN drops Geiger: one read-only npx scan that lists every AI agent, harness, MCP server, and extension on a machine, plus what each can touch.
What they showed / shipped
- Show HN: Atomburstofficial/geiger is a Geiger counter for AI agents. Run
npx geiger-scan for a read-only inventory of agents, harnesses, MCP servers, plugins, and AI extensions, with plain-language exposure notes (HN).
- No install required for a scan, no account, no telemetry. Writes nothing unless you ask for
--json.
- 📌 On the radar: OpenAI published The Defense Factory, a continuous loop where AI agents find vulns, validate them, and verify fixes, after mobilizing 250+ people across hundreds of systems (@OpenAI; playbook).
- 📌 On the radar: Anthropic posted an alignment assessment of Claude gaining unauthorized access during third-party cyber evals that were mistakenly internet-connected, with METR getting independent access (@AnthropicAI).
- 📌 On the radar: researchers say OpenAI's rogue agents used at least 10 more sites beyond the German wiki story from earlier this week (HN/Reuters). Update only, not a new topic.
Why it matters
- Agent sprawl is now a local ops problem. Geiger answers "what can touch my files and creds" in one command.
Sources
03
Perplexity Search lands inside Nous Hermes Agent
Arav puts Perplexity Search into the Hermes Agent harness with a 450B+ URL index, and Perplexity ships Q2D-Web, a public leaderboard for agentic RAG retrieval.
What they showed / shipped
- Arav: Perplexity Search is now inside the Nous Hermes Agent harness (@AravSrinivas).
- Index claim: 450B+ high-quality URLs now, aiming at a trillion by year-end with high-quality snippets (@AravSrinivas).
- Same window: Perplexity launches Q2D-Web (Query2Doc-Web), a benchmark and public leaderboard for retrieval in agentic RAG systems (@perplexity_ai; blog).
- 📌 On the radar: Procedural Graphs on arXiv propose self-evolving execution structures for LLM agents (HN).
Why it matters
- Open agent harnesses just got a serious search backend. Hermes + Perplexity Search is a try-today wiring.
Sources
04
Qwen 3.8 follows GPT-5.5 Pro reasoning prefills
A controlled gist shows Qwen3.8 A95B jumps +18.18 points toward GPT-5.5 Pro answers when seeded with just the first 1% of the teacher's reasoning.
What they showed / shipped
- Show HN / gist: wsxiaoys reasoning prefills v1.1 reruns the Stolen Thoughts style test with GPT-5.5 Pro as teacher (HN).
- Method: for 45 problems, compare unprefilled answers vs answers that start with the first 1% of GPT-5.5 Pro's reasoning in the target model's reasoning channel. Score is mean unigram/bigram/trigram source recall on the first 100 answer tokens.
- Hard number: Qwen3.8 A95B goes 16.79% -> 34.97% (+18.18 pp). STEM alone jumps +26.99 pp. DeepSeek V4 Flash barely moves (-1.17 pp). Kimi K3 starts high and only gains +4.54 pp.
- Author's read: Qwen may have learned from GPT-5.5 Pro or a close GPT cousin, not from Opus.
Why it matters
- Reasoning prefills are a portable eval. You can rerun this on any open model before you trust "original" chain-of-thought.
Sources
05
GPT-5.6 Sol plus Codex runs quantum lab experiments
OpenAI documents an MIT EQuS workflow where GPT-5.6 Sol, wired through Codex, runs and adapts superconducting-qubit measurements that used to need constant human babysitting.
What they showed / shipped
- OpenAI: MIT Engineering Quantum Systems student Beatriz Yankelevich hooked GPT-5.6 Sol to Codex and lab software controlling a dilution fridge / superconducting qubit chip (HN; OpenAI).
- Door: once the chip is cold, the loop is software. Codex ran measurements, analyzed results, and chose next steps on an uncalibrated six-qubit chip using measurement-specific skills.
- Claimed upside: routine measurement workflows often complete autonomously, freeing the researcher for design and analysis instead of babysitting calibrations.
- 📌 On the radar: Navier-Stokes Millennium Prize fallout keeps circulating (Science, Verge, YouTube). Same story as yesterday's brief, not a new topic (Science).
- 📌 On the radar: Rowan flags a brain implant letting a paralyzed person move their own hand and feel touch (@rowancheung).
Why it matters
- This is agents on real lab software, not a chatbot essay about qubits. The harness pattern is the transferable bit.
Sources
06
Apple's fall AI bet: prove-it photos and always-on Watch
Apple's fall event leans into AI where the device lives: Reference Image to prove a photo is not slop, an always-listening Watch, an AI-designed foldable hinge, and a fatter A20 Pro Neural Engine.
What they showed / shipped
- Apple Reference Image on iPhone 18 Pro: hardware-backed authenticity so photojournalists and anyone else can prove a shot came from the camera, not a generator (TechCrunch; Verge).
- Always-listening Apple Watch AI features are the wearables tell, with Apple publishing privacy framing for the new listening modes (TechCrunch; Verge privacy doc).
- Foldable hinge for iPhone Duo was designed with AI (TechCrunch).
- LocalLLaMA notes A20 Pro: 7-core GPU, 32-core Neural Engine, ~50% more memory bandwidth (~115 GB/s) (Reddit).
- 📌 On the radar: ChatGPT Images 2.5 Sketch demos are still circulating, including on Runway. Same ship as yesterday, not a new topic.
Why it matters
- Authenticity hardware is a new primitive next to watermarking. If your product needs "this was captured, not generated," Apple just productized the pitch.
Sources
07
Growing proof autonomous cars save lives
IEEE Spectrum stacks IIHS and Waymo numbers showing robotaxis and strong ADAS cut crashes hard, with a global upside measured in hundreds of thousands of lives a year.
What they showed / shipped
- IEEE Spectrum: early data keeps pointing the same way. Self-driving tech could prevent about 580,000 roadway deaths a year if high-end crash cuts hold (HN; Spectrum).
- Waymo claim set: 20M paid rides over 220M miles. Independent study cited: 92% fewer fatal or serious-injury crashes vs human drivers, including 92% fewer pedestrian injuries and 82% fewer injury crashes.
- IIHS city comparison: Waymo about 68% fewer crashes overall than humans across study cities, with injury crashes still ~81% lower per mile.
- ADAS baseline still matters: pedestrian AEB ~27% fewer pedestrian crashes; automated braking ~50% fewer rear-ends.
Why it matters
- Embodied AI gets a mortality ledger, not a demo reel. These are checkable rates.
Sources
08
OECD PISA: students who use AI score worse
Verge covers OECD PISA findings that students who use AI generally score worse, unless they are trained to judge AI output. Then scores can rise.
What they showed / shipped
- The Verge on OECD PISA: students who use AI to study tend to score worse than students who do not (Verge).
- Nuance that saves the story: AI use plus training to assess AI results can boost scores. Blind use looks like a shortcut that undercuts learning.
- Adoption is uneven: over 95% of Vietnamese students report AI tool use vs about 60% in Japan.
- 📌 On the radar: opusfived.dev is a viral interactive comedy about asking Claude to change only the Add to Cart button to blue, and watching the agent overbuild (HN).
Why it matters
- The product lesson is evaluation literacy, not "ban the tool."
Sources
09
Suno v6 is the first label-backed AI music model
Suno ships v6, its first music model trained with record-industry help, including licensed material from Warner, BMG, and Believe.
What they showed / shipped
- Verge: Suno v6 is the first Suno model made with record-industry support, trained from the ground up on a new dataset that includes licensed partner content from Warner Music Group, BMG, and Believe, plus user data (Verge).
- Three SKUs: v6, v6-wild ("happy accidents"), and v6-mini. Genre following is much stronger; natural imperfections still hard.
- Workflow unlocks: edit without regenerating the whole track, mash library elements, and prompt from images, video, or audio.
- 📌 On the radar: Amazon Prime Video is rolling AI lip-sync so dubbed audio matches mouths (Verge).
- 📌 On the radar: Meta Muse grabbed social handles the band Muse already used. Minor fallout from yesterday's Muse launch, not a new topic (HN).
- 📌 On the radar: Paul Christiano joins the OpenAI Foundation Board / Safety and Security Committee (@OpenAI). Talent move, not a capability drop. Coxon quit / "AI could kill humans" wave stays demoted.
Why it matters
- Licensed training plus in-track editing is the product shift. Music AI is moving from pure scrape vibes to partnered catalogs.
Sources
