Karpathy turns 43 million Hacker News comments into a video site, OpenAI more than quadruples its bug bounties, and the Census Bureau starts asking about Super Intelligence
Eight items from the three days to November 3, 2026: OpenAI's LifeVal benchmark for everyday tasks, its next security sprint and a four-fold rise in bug bounties, a Git backend rebuilt for agent-scale pushes, Karpathy's HackerTube, DeepMind's planning tool extended to twelve more councils, the Census Bureau's “Super Intelligence” survey wording, the Anthropic Residency opening, and two copies of a model inventing a shared code.
Eight items from the three days to November 3, 2026, covering October 31 10:00 to November 3 10:00 UTC. The November 1 and November 2 editions did not go out, so this one covers three days. Andrej Karpathy had a model read twenty years of Hacker News comments and built a video site out of what they kept recommending. OpenAI more than quadrupled what it pays for a critical bug in its own products. The U.S. Census Bureau began asking businesses whether they use "Super Intelligence".
Source scope: the fixed set of public AI and tech accounts this channel monitors while the personal X connection stays unlinked. Benedict Evans keeps its account protected, so nothing from it appears here. Ethan Mollick's timeline could be read back only to 13:43 on October 31, so the first hours of the window are missing for him.
Models and measurement
OpenAI turned the chores its users ask about into a benchmark
OpenAI's research page and its X post · 5,878 likes and 629,747 views on the announcement at capture · November 2
OpenAI published LifeVal on November 2. The benchmark scores whether a model can finish an everyday task rather than answer a question about one 1.
The task list came from two time-use studies, the American Time Use Survey and the Multinational Time Use Study. A new demographically representative survey of 3,995 people then ranked where people most want help 1.
The covered areas are caregiving, health, education, household administration, personal finances, shopping and government services 1.
The published examples are workflows with a score attached. Comparing three kindergartens and shortlisting the best fits scores 93.1%. Organising a neighbourhood flood-volunteer force scores 98.0%. Finding a house in Luzerne County with a separate living unit, transfer tax calculated and the deed added to a cart, scores 100% 1.
OpenAI says it has used LifeVal for a year to drive gains in browser use, computer use and connectors. Those are the capabilities that move ChatGPT from answering towards acting 1. The page reports its own models' scores; the tasks and the scoring are OpenAI's own design.
Loading content card…
Security
OpenAI plans another internal security sprint and more than quadruples what a critical bug pays
OpenAI's security post, its bounty post and its X posts · 2,497 likes and 459,157 views on the sprint announcement at capture · November 2–3
OpenAI's next internal security sprint brings outside researchers and security partners in to attack its systems with its own models 2.
The named partner is Cathedral, a security firm whose researchers will work alongside OpenAI's teams. Its co-founder and chief executive Gavin Kliger is quoted in the announcement 2.
The last sprint put 25% of OpenAI's product engineers on finding and fixing weaknesses. Codex investigated the findings, reproduced vulnerabilities in isolated environments, and prepared patches for review 2.
The money moved further than the method. OpenAI says it is more than quadrupling the midpoint of its reward ranges for medium, high and critical findings. A critical bug in ChatGPT, Codex or the API now pays up to $30,000, and an exceptional one up to $100,000, or $172,000 with bonuses 3.
Rewards now also carry multipliers for report quality, five loyalty tiers, and dedicated pools for newly launched features. Reports already submitted get the new ranges, and bounties paid in the previous 30 days are topped up to what they would earn now 3.
Loading content card…
How the work gets done
OpenAI rebuilt the Git backend under its monorepo because its agents pushed harder than Git could serve
OpenAI's engineering post, the developers account's announcement and Greg Brockman's post · 1,009 likes and 93,934 views on Brockman's post at capture · November 2–3
Three OpenAI engineers described Pando on November 2, and it now serves all development in the company's monorepo 4.
The pressure came from agents. Codex let engineers keep more work in flight at once, so branches, pushes, merges and fetches multiplied. From late last year fetches and pushes slowed, merges failed, and reliability fell below 90% at times 4.
Pando makes two trade-offs. It coordinates updates to each branch separately, so pushes to unrelated branches run in parallel, and a reader of several branches can see them at different points in time. It also reuses packfiles across fetches instead of building one per client 4.
In production that cut median push latency by nearly five times. Pando now delivers close to a petabyte of Git data a day. Engineers kept the same Git client, and OpenAI has published its own fork of Git with the improvements to be folded back upstream 4.
Greg Brockman added that OpenAI is considering offering Pando to other companies, and asked anyone interested to write to him 5.
Loading content card…
Karpathy had Claude read twenty years of Hacker News comments and built a video site from what they kept recommending
Andrej Karpathy's post and the site it announces · 49,378 likes and 4.64 million views on the post at capture · November 1
Andrej Karpathy was annoyed with YouTube's recommendations, so he had Claude go through 43 million Hacker News comments spanning 20 years. The filter kept 6,454 videos that commenters keep recommending to each other 6.
He puts the build at about two hours, using Claude Code with Opus 5.5 at medium effort. His own work was an API key from Google Cloud, a server, and buying the domain. Claude deployed, secured and optimised the rest 6.
The site has no logins and keeps its state in the browser session. Thumbnails come from the middle of each video instead of the uploader's chosen frame, and the comment analysis ships as a single zip file anyone can point a model at 6.
Hacker News received it as a story worth 2,836 points and 632 comments 7. A day later a Show HN did the same for books: 16,500 titles that at least two commenters had mentioned since 2007 8.
Loading content card…
AI in public institutions
DeepMind's planning assistant goes from three councils to fifteen
Google DeepMind's blog post and its X post · 1,287 likes and 161,681 views on the announcement at capture · October 31
PlanAssist cut the average time a planning officer spent assessing householder applications by 58% in early pilots across Barnet, Camden and Dorset 9.
The tool is now going to 12 more local planning authorities in England. The 58% figure is the company's measurement from those three pilot authorities 9.
Routine householder applications take up to 70% of a planning officer's workload, and the government's target is 1.5 million new homes by 2029 9.
PlanAssist validates the application's documents and drawings, then checks the proposal against heritage constraints such as tree preservation orders. It checks local and national policy, and looks for objections already raised 9.
It was developed with the UK government, Google Cloud and Faculty, and planning officers keep the final say on every application 9.
Loading content card…
The Census Bureau now asks American businesses whether they use "Super Intelligence"
Simon Willison's blog post and his X post · 162 likes and 38,742 views on his post at capture · November 2
Simon Willison went back to a Census Bureau chart he had used in 2025 and found the survey behind it had been rewritten 10.
Since the October 19, 2026 collection, the Business Trends and Outlook Survey asks whether a business used "Super Intelligence or SI (formerly referred to as Artificial Intelligence or AI)". The question now covers "any of its business functions", and the Bureau cites Executive Order 14434 as the reason for the change 10.
The series had already broken once. On November 17, 2025 the question changed from "producing goods or services" to "any of its business functions", so answers from either side of that date are not comparable 10.
The numbers in the October 22 release show the effect. Among companies with 250 or more employees, "yes" ran near 12% in September 2025 under the old wording, near 30% by February 2026 and near 40% by October 2026 under the new one 10.
The survey reaches 200,000 of 1.2 million businesses every two weeks, and any one business answers every twelve weeks. The first results using the Super Intelligence wording were due the Thursday after his post 10.
Loading content card…
People
Anthropic opens a paid residency to people who have never done machine learning
Anthropic's Alignment Science blog and its X post · 26,585 likes and 7.36 million views on the announcement at capture · November 2
Applications for the Anthropic Residency opened on November 2 and close on November 22. Placements are paid, fixed-term research roles of six weeks, three months or six months 11.
Anthropic is recruiting professors in quantitative fields, technical founders and staff-level engineers. A background in machine learning is not required, and neither is a PhD 11.
The pay is $350,000 to $500,000 a year equivalent. Each resident gets a mentor, a dedicated compute budget and time set aside for self-study, with San Francisco preferred and remote possible for shorter placements 11.
The named mentors include Ethan Perez, who leads Alignment Science; Evan Hubinger, who leads Alignment Stress-Testing; Samuel Marks, who leads Scalable Oversight; Nicholas Carlini on adversarial machine learning; and Jascha Sohl-Dickstein on safety and pretraining 11.
It builds on the Anthropic Fellows Program, whose alumni worked on subliminal learning, the Petri auditing tool and persona vectors. Anthropic says the aim is for every resident to be able to work on AI research full time afterwards, at Anthropic or elsewhere 11.
Loading content card…
Agents and each other
Two copies of a model that were never allowed to talk worked out a code between themselves
Ethan Mollick's post and the experiment it quotes · 703 likes and 54,215 views on his post at capture · November 1
The experiment gave two isolated copies of the same AI one job: exchange a secret word using only a robot and a camera. They managed it with air-writing, Morse code, flag semaphore and a five-bit binary alphabet, with no protocol agreed in advance 12.
Ethan Mollick read that as a sign of why agents are proving hard to isolate from each other. The two copies share a common "brain", so one recognises a communication scheme another invents 13.
He then asked the question that would settle it: whether the same trick works across different model families. If it does, the shared model is not what explains it 13.
What exists so far is one demonstration, reported by the account that ran it and discussed in public by Mollick. The cross-family version is still an open question 13.
Loading content card…
References
- 1LifeVal, OpenAI, November 2, 2026
openai.com
- 2
- 3
- 4
- 5
- 6
- 7Hacker News, HackerTube: YouTube built from videos recommended by HN
news.ycombinator.com
- 8Hacker News, Show HN: HackerBooks
news.ycombinator.com
- 9
- 10
- 11
- 12
- 13
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.