Good day . It’s Election Day, and a new audit found AI chatbots botching basic voting questions. Plus OpenAI’s test for everyday chores, an AI that rewrites another AI’s weights, Anthropic’s amended S-1, and why Postgres 19 is late.
Chatbots Botch Voting Rules on Election Day NPR
What happened: In October, the nonprofit Innovating for the Public Good (IFPG) tested major chatbots on election questions, and NPR reports that 13.9% of their answers to voting-logistics questions were outdated or inaccurate. In 5.9% of answers the error could cost a voter: a missed deadline, the wrong day, or a skipped step.
Why it matters: Today is Election Day, Tuesday, November 3, and these tools are where a lot of people now ask how to vote. NPR says accuracy depended more on the chatbot than the question: Claude and Copilot produced three-quarters of the clear errors, and on voting questions Claude was wrong or outdated about one time in three, Copilot one in four, ChatGPT one in 10. Earlier this week we covered ChatGPT trusting fake local news. This is the same weakness, with a deadline on it.
What everyone’s saying: IFPG’s Chambers called the error rates “extremely troubling” and said the “legions of errors” make you wonder how much more there is to find. Anthropic spokesperson Michael Aciman told NPR that because IFPG used an API, some of Claude’s key safeguards may not have been active. IFPG says its own comparison of API and web answers found them comparable.
My read between the lines: The wrong answer is half the problem; the confidence is the other half. NPR also found every chatbot tested commonly pointed to betting sites like Polymarket and Kalshi when asked who’s winning. Betting odds dressed up as polling is a new way to be wrong. Check your state’s official election site before you trust any of them.
📖 Further reading: AI Is a Trust Problem, Not a Tech Problem — My notes from the Chamber of Commerce AI + Work Summit, where the word under every panel was trust, the gap a confidently wrong answer opens.
A chatbot that guesses with total confidence makes a terrible coworker. Viktor is an AI agent that lives in your Slack and connects to 3,000+ tools, then does the work: reports, dashboards, code and campaigns. A coworker, not a chatbot. New readers get $50 off their first month. Hire Viktor →
OpenAI’s LifeVal Grades AI on Chores OpenAI
What happened: On Monday OpenAI published LifeVal, a test of how well AI handles everyday tasks like shortlisting kindergartens after a move, organizing flood-prep volunteers, or building a meal plan for a parent newly diagnosed with diabetes. OpenAI says GPT-6 Astra scored 93.1%, 98% and 90.4% on those three, and CNN reports it scored 80% overall versus 49.3% for GPT-5.2 from December 2025.
Why it matters: Benchmarks steer what labs build. OpenAI says it has used LifeVal for the past year to track and drive progress on everyday tasks, because it wants ChatGPT to “take actions that actually get things done,” not just chat. The tasks came from a 3,995-person survey, and about 100 users wrote their own tasks and grading rubrics, per CNN.
What everyone’s saying: CNN frames it as a move in the race for personal AI agents against Meta, Anthropic and Google, noting Meta’s Muse hit 10 million U.S. downloads in its first 40 days, per Sensor Tower. OpenAI researcher Jenny Yang told CNN the goal is to learn “what are the things that we’re not so good at yet?”
My read between the lines: OpenAI built the test, picked the tasks and graded its own model, and the model did great. The 100% scores in OpenAI’s table are labeled “best recorded,” which is not the same as typical. The number that matters is how an average run handles your Tuesday, and that one isn’t in the chart.
📖 Further reading: OpenAI Made Its New Agent Adorable. That’s the Part That Worries Me. — Dots is the always-on agent running on GPT-6 Astra, the model LifeVal grades, and my read on what happens when a chore-doing agent is cute.
The daily Brief is free and always will be. Members get the paywalled deep-dives behind the headlines, like how I tested whether AI agents can even find my business, plus the full archive. Become a member →
GPT-6 Astra Rewrites Another AI’s Weights MIT CSAIL
What happened: MIT CSAIL researchers report that OpenAI’s GPT-6 Astra can read and edit a neural network’s raw weights directly, with no training loop. Asked to improve Qwen3.5-0.8B, a small open model, Astra edited 0.016% of its weights, and the authors say its MATH-500 score rose from 7.9% to 48.1%.
Why it matters: Weights are the billions of numbers that make up a model, normally adjusted by optimizer software. The authors say the results “might be a big deal for both optimization and interpretability,” and report that Astra could tell what game a network was trained on just from its weights. They call the findings very early.
What everyone’s saying: Hacker News sent it to about 1,959 points, per Hacker Times. One commenter was skeptical that Astra “can only improve weights to its own level,” so it can’t train itself this way. Another posted a single word: “RSI,” short for recursive self-improvement.
My read between the lines: Editing 0.016% of a model sounds surgical until you notice the patient: 0.8 billion parameters, scoring 7.9% on math, stuck in reasoning loops its own model card warns about. The bar was on the floor. Still, an AI fixing a bug in another AI by reading its raw numbers is a new kind of party trick.
📖 Further reading: The AI Pattern That Optimizes Anything Measurable — Overnight — Karpathy’s agent ran 700 experiments on his own training code overnight; this week’s paper takes the same idea down to the weights themselves.
Anthropic Says AI Speedup May Be Imminent SEC filing
What happened: On Monday Anthropic filed an amended S-1 with the SEC that says its “internal estimates and our expectations regarding future models” indicate it “may imminently meet” its Responsible Scaling Policy’s definition of “dramatically accelerated progress attributable to AI.” The filing says no current model is designated as exceeding that threshold.
Why it matters: CellCog’s read is that the bar means a year of progress equal to two at baseline. In the same filing, Anthropic says over 90% of contributions to its codebase in September were written by Claude, and that it does not expect to “safely use Claude to fully autonomously design and build successive generations of frontier models.” Last week we covered Anthropic opening its books. This is the update.
What everyone’s saying: The sentence spread fast on X, with posts from Peter Wildeford and Andrew Curran quoting it, per Techmeme’s roundup. Bloomberg, via Techmeme, reports the update also showed Q3 revenue up 12x from a year earlier to nearly $17 billion, and a $23.3 billion net loss.
My read between the lines: This is a risk-factor line in an IPO filing, so lawyers made it say exactly what it says: future models, not current ones. Still, a company telling investors the AI-building-AI dial may soon hit 2x, while saying it can’t safely let Claude build its successors alone, is a sentence worth reading twice.
Postgres 19 Is Late, and AI Gets the Blame Robert Haas
What happened: Postgres committer Robert Haas writes that the release of Postgres 19 is delayed because “a huge number of people on the Internet are now reporting bugs in PostgreSQL that they found using LLMs.” Between feature freeze and Sept. 25, he says, back-patched commits to the stable branch rose from 724 to 1,169 year over year, while commits to the main branch slightly declined.
Why it matters: Haas says LLMs “can draft fixes, but those fixes aren’t always correct or compliant with PostgreSQL standards,” so humans still review them. He says reviewing the older branches ate time that would have gone to checking what went into 19, which is why late reverts piled up. His verdict: the delay “emerges directly from the project’s concern for quality.”
What everyone’s saying: The post drew 479 comments and 807 points on Hacker News, per Hacker Times. One commenter, TomatoCo, wrote that you “can create a project on vibes, but you cannot maintain one on vibes.”
My read between the lines: AI made finding bugs cheap, and fixing them still costs a human an afternoon. Every open-source project short on maintainers is about to learn that the way Postgres did. Free bug reports are a gift right up until they become a queue.
That’s your AI Brief for Tuesday.
—Nicholas from Artificially Intimidating
