Today's thread
All four stories today are about where AI gets attached and on what terms. Google will start student accounts with AI switched off, but because Search and AI come as one bundle, schools have little room to choose. ChatGPT ads, measured against real revenue, turned out to be worth far less per visit than search ads.
Security and research follow the same pattern. The Binance incident showed how big a door one laptop with access to internal systems can be. Google DeepMind expanded its AI work app by handing chemistry calculations to proven tools rather than to the AI's fluency. Set the scope before switching AI on, and measure it separately once it is on.
ChatGPT ads earn $0.89 per visit — a third of Google Search
According to an October working paper (version 1, not yet peer reviewed) by Maximilian Kaiser of WU Vienna and Christian Schulze of Frankfurt School, a visit from a paid ChatGPT ad in US traffic generated $0.89 in revenue on average. A paid Google Search visit generated $2.45, so ChatGPT was 64% lower. The conversion rate, the share of visitors who actually bought something, was 0.37% versus 1.17%.
The analysis used 60 weeks of website visit data from June 2025 to August 2026, and the channel comparison covered 102 advertisers whose revenue could be tracked. Until now, the ChatGPT ads story was mostly OpenAI's reported ad revenue growth; this study looks at what the buyers actually earned. The authors calculate that matching search's rate of return would require a ChatGPT ad click to cost about $0.26.
For anyone spending an ad budget, this is a reason to recheck, with numbers, the hope that ads next to AI answers will become the new search ads. Rather than moving the search budget wholesale, test with a small separate amount and compare revenue per visit, not clicks. If you pay well above $0.26 per click, the math is hard to make work by this study's standard.
There are limits. The data is US traffic, and this is a first version that has not been reviewed by other researchers. How well it captures people who saw an ad and later bought through another route depends on the measurement method, so it is too early to apply it directly to the Korean market or your own industry.
Google turns student AI off by default — and Search goes with it
On November 3, Google said on its education blog that Gemini, Gemini Notebook, Gemini in Classroom and Google Search will be entirely off for K-12 student accounts until school administrators explicitly turn them on. It set no exact date and said it will share details with schools in the coming months before the change. The same day, it made Guided Learning, which coaches students with step-by-step questions instead of answers, available globally in Google Classroom.
In August, Google opened Gemini in Classroom to students of all ages and left it on by default. This change reverses the direction, making off the starting point. But Chris Phillips, Google's VP for education, told The New York Times that Search and AI are deeply integrated, so students at schools that ban AI will not have access to Search.
For schools, academies and education businesses using Google Workspace for Education, Google's bundle of work tools for schools, one admin setting reshapes the whole set of classroom tools. Make a list of what is currently on for student accounts, and decide in advance how classes will run if Search is switched off too. Company accounts face the same question: should AI be on or off by default, and what else goes off with it?
Much is still unknown. Google has not said whether AI summaries and AI Mode can be turned off separately inside Search for schools, or whether existing school settings will carry over. Guided Learning in Search starts with a pilot at select US schools this month.
Binance's $6.3 billion theft started with one laptop
Crypto exchange Binance said that on October 29 (UTC) about $6.3 billion worth of assets left its wallets. Its initial investigation pointed to a compromised laptop belonging to a DevOps engineer, who handles server operations, which may have let the attacker reach an internal service and submit unauthorized withdrawals. It said it found no evidence that private keys, the equivalent of the keys that open a wallet, had leaked.
CEO Richard Teng later said the attacker did not simply use one engineer's permissions and that it was not a single one-click action, adding that additional points of breach had been found. Binance said it would cover losses from its SAFU user protection fund and company funds, and that about $3 billion had been frozen. Deposits reopened the next day, and as of a November 1 notice, withdrawals are being restored in phases as checks are completed.
Read this as a story about work laptops, not exchanges. The key point is that one PC belonging to someone with operating permissions could reach the systems where company money moves. It is time to check which laptops are signed in to accounts that touch payments, customer data or servers, and whether one person can move money alone.
How the laptop was compromised and what the additional points of breach were have not been disclosed. Some blockchain analytics firms have pointed to a North Korea-linked hacking group, but neither Binance nor any government agency has officially named who was responsible.
Google Antigravity chemistry skills — RDKit does the math
On November 3, Google DeepMind released Cheminformatics Skills, a set of chemistry analysis features for Google Antigravity. Antigravity is Google's work app where agents, AI that takes an instruction and carries out multiple steps on its own, do the work. The new skills are built on RDKit, an open-source chemistry toolkit, connect to the public compound databases ChEMBL and PubChem, and become available once you install the Science plugin (an add-on) in the app.
The key is a division of labor. The language model, the engine behind chat AI such as ChatGPT, plans the steps, searches the data and reasons, while molecular calculations go to chemistry tools that always give the same answer for the same input. DeepMind described a lung cancer case with lorlatinib and crizotinib in which thousands of compounds were narrowed to 568 and new ring-structure candidates with synthetic routes were proposed in about an hour.
The design is useful even if you never touch chemistry. Having the AI hand off work that must not be wrong, such as number crunching or rule checks, to Excel functions or proven programs instead of doing it itself follows the same principle. When building work automation, splitting judgment to the AI and calculation to tools makes results easier to double-check.
DeepMind itself wrote that AI-generated designs are testable hypotheses, not finished drugs, and that every one must be synthesized and validated experimentally. It is also taking applications to give 10,000 scientists the highest usage tier at no cost for six months, subject to eligibility, capacity and selection.
Try this today
If you are testing ChatGPT ads, add a revenue-per-visit column to this week's report instead of clicks, and put it side by side with Google Search ads for the same period.
If your school or academy uses Google Workspace for Education, record in the admin screen which Gemini and Search settings are currently on for student accounts. It becomes your baseline for comparing what changed when the default flips.
List the work laptops signed in to payments, customer data or servers, and if any account lets one person move money alone, check whether it can be switched to two-person approval.
Pick one automation you have handed to AI and check whether the calculations inside it are done by the AI itself or by a function or program. If the AI is doing the math, move just the calculation to a tool.