01
OpenAI says it hit automated research intern
First-party numbers: OpenAI claims coding agents now handle multi-day research tasks under human direction, with a March 2028 target for a supervised automated AI researcher.
What they showed / shipped
- OpenAI published Research acceleration: The view inside OpenAI: they say they hit the fall goal of an automated research intern by Sep 2026 (tasks that would take a skilled researcher a few days under human direction).
- Trajectory they state: strong progress toward an automated AI researcher by March 2028. Humans still set priorities, judge results, and decide scale/pause/deploy.
- Internal signal: researchers run coding agents all day (often concurrent sessions), contribute code faster, run more experiments; agent usage on research is outpacing other OpenAI teams.
- Sam Altman amplified it (post pointing at Jakub) and earlier RT'd the data release framing recursive self-improvement as the stakes.
Why it matters
- This is the lab saying agent coding is already changing how frontier research gets done, not just product demos.
Sources
02
Recreating Minecraft is a demo, not a benchmark
A sharp post names the launch-cycle trick: fixed viral demos (Minecraft, pelican SVG, bouncing balls) get overfit on a schedule, so they measure prep more than capability.
What they showed / shipped
- Recreating Minecraft Is Not a Benchmark coins demo-benchmarks: visual, viral, finite targets labs can perfect between releases.
- Evidence beat: Thinking Machines' Inkling Small scored within a point of its flagship on Artificial Analysis with under a third of the parameters, and beat it on HLE / GPQA Diamond / SciCode. Static famous sets leak into training.
- Same-window case study: @minchoi lists Astra demos (anatomy, LEGO, GTA, V8 engines, forests) as people feeling AGI less than 91 hours after launch.
Why it matters
- If a model nails the timeline demos but fails your real workflow, trust the workflow. Prefer rotating/holdout evals (LiveBench, ARC private sets, HLE holdouts).
Sources
03
87% of exploited bugs now get hit on day zero
a16z Charts of the Week: of bugs hackers actually exploit, ~87% are attacked on or before the day the bug goes public, up from 23% in 2020.
What they showed / shipped
- @a16z: the window to patch is collapsing. Share of exploited bugs attacked on/before public disclosure day: ~87% now vs 23% in 2020.
- 📊 Chart attached in the Charts of the Week post.
Why it matters
- Assume zero-day-ish for anything you ship publicly. Patch pipelines and agent tooling that touch prod need same-day response, not weekly cadence.
- 📌 On the radar: related volume beat already ran Sep 6 (critical vulns 100→600/mo). Today is the time-to-exploit angle.
Sources
04
AI buildout is printing electrician and HVAC jobs
Goldman via a16z: 300k+ construction jobs tied to the AI buildout since 2022, ~75k in the past year, with electrician/HVAC trades growing ~2%/yr (about double construction overall).
What they showed / shipped
- @a16z Charts of the Week: Goldman attributes 300k+ construction jobs to the AI buildout since 2022, ~75k in the past year.
- These trades growing ~2% a year, roughly double construction overall. 📊 chart in post.
Why it matters
- Builder/creator lens: the AI jobs story is not only prompt engineers. Capex is hiring trades.
- Teachable counter-narrative to pure white-collar displacement takes: physical infra absorbs labor too.
Sources
05
Astra Max one-shots a wedding site and an S-1 banker deck
Two concrete Astra Max workflows landed in the same window: a full wedding website, and an interactive DCF / sum-of-parts / comps pack from an S-1 (tested on Oura's IPO docs).
What they showed / shipped
- @venturetwins: built a wedding website with GPT-6 Astra Max.
- @venturetwins: drop in an S-1, get interactive DCF, sum-of-parts valuation, and public + private comps. Tested on Oura IPO docs.
Why it matters
- This is the post-GA so-what. Not another demo-benchmark. A repeatable workflow you can try on your own PDF/S-1 today.
- Rank these above the AGI-feeling montage. Use beats teach. Vibes do not.
Sources
06
Authors fight publishers and agents over Anthropic settlement cash
Anthropic's $1.5B copyright settlement is paying ~$3k per pirated title across ~500k works, and authors say publishers and even agents are grabbing shares they should not get.
What they showed / shipped
- TechCrunch: settlement approved; ~500k titles at $3,000 each. In-print traditional: 50/50 author/publisher. Self-pub or reverted rights: author should get 100%.
- Writers Beware / Authors Guild: publishers claiming reverted books or 100% when owed 50%; some literary agents claiming cuts even though agents are not rightsholders.
- Dispute path exists; rights must have reverted before Aug 10, 2022 (settlement download date) for a full author claim.
- 📌 On the radar: Seattle Times and Newsday sued OpenAI and Microsoft, seeking destruction of models trained on their work (joins the long publisher pile).
Why it matters
- Training fair-use won in that case; piracy still costs. The payout fights are the hangover.
- Teachable: $3k × ~500k titles is the scale of the check.
Sources
07
Kalanick's Atoms looks like a robotaxi play
FT via TechCrunch: after a $1.7B a16z-led round, Atoms is prepping hiring and acquisitions aimed at AVs, with talks about Uber using its robotaxi tech (Uber already in for $100M).
What they showed / shipped
- TechCrunch: FT says Atoms is planning a hiring spree and acquisitions to become a major AV player.
- Reportedly talked to Uber about using Atoms robotaxi tech; Uber has invested $100M. Fits Kalanick's unfinished business framing and the Pronto / Levandowski acquisition.
Why it matters
- Embodied-AI beat: another deep-pocketed robotaxi stack forming next to Wayve/Uber London and Tesla Cybercab noise.
- Still rumor-forward (FT sourcing). Frame as trajectory, not confirmed product ship.
Sources
08
One engineer maps every feeling about AI at once
A short HN essay walks surprise, fear, disgust, sadness, anger, and happiness about AI in one page, and lands on: positive technologically, bleak societally.
What they showed / shipped
- How I feel about AI (HN): six emotions in sequence, from emergent reasoning surprise to open-web crawler disgust to artist-slop sadness to discovery-era happiness.
- Closer: tech-level positive, society-level bleak. Author notes Mistral reviewed; they typed every word.
Why it matters
- Not a capability drop. It is teachable public-perception framing for a mixed week.
- 📌 On the radar (repeat): a16z Atlas interview quotes (post) rehash the Sep 2/5 World Labs story. No new ship.
Sources
