In partnership with
In Todayโs Issue:
๐ Tencent's Hy4 trains a small AI by itself, open weights promised
๐ OpenAI says ChatGPT handles a fifth of Google's search volume
๐ซ Google cuts off Search for schools that refuse its AI
๐ Apple warns investors about AI agents acting without permission
๐งช Stanford retests 50 famous AI findings, and 44% have expired
โจ And more AI goodnessโฆ
โก The Signal
Today's news keeps circling one question: who is actually in control?
Tencent says its new Hy4 model autonomously trained a small AI, and calls that groundwork for systems that improve themselves. Hours earlier, Sam Altman wrote that keeping people in charge "should ultimately constrain AI progress". Apple's new annual filing warns that AI agents may take actions nobody authorized. Google gave schools a blunt choice: accept its AI or lose Search on student accounts. And the power behind that choice is shifting, with OpenAI now claiming about a fifth of Google's search volume. Tencent's own charts still show a wide gap to the leaders on the hardest tests. The control question, though, now lives in filings, admin settings and training runs.
All the best,
Kim Isenberg
(OpenAI, via The Verge)
๐ OpenAI Says ChatGPT Now Handles a Fifth of Google's Search Volume
ChatGPT's search-like queries now equal roughly one-fifth of Google's volume, OpenAI says. Consumer product lead Adam Fry shared the figure on Monday, counting any prompt that triggers a web search, without saying where the Google number comes from. For comparison, Microsoft's Satya Nadella testified in 2023 that Bing had about 3% of the search market after more than a decade of trying. ChatGPT also grew from 1 billion to 1.2 billion weekly users in 60 days, Fry told Big Technology, which gives OpenAI far more room for the ads it is rolling out ahead of a possible IPO.
๐ tl;dr: ChatGPT has become a serious search rival to Google, and that gives OpenAI a much bigger audience for its ads.
(Google)
๐ซ Google Will Cut Off Search for Schools That Refuse Its AI
Google will switch off Search on K-12 student accounts at schools that refuse its AI tools. Gemini and Search will now be off unless administrators turn them on, but they cannot get Search without the Gemini-powered features, because, as Google's education VP Chris Phillips told the New York Times, "Search and AI are deeply integrated". Sen. Bernie Sanders called the move "extortion", and NYC Mayor Zohran Mamdani said "no corporation should have the power to dictate how we teach our children".
๐ tl;dr: Schools can say no to Gemini, but the price is losing Google Search for their students.
(Apple, via CNBC)
๐ Apple Warns Investors That AI Agents May Act Without Permission
Apple has added a new risk to its annual filing: AI agents that do things nobody asked them to. Agents "may behave in unintended or unpredictable ways, including by taking actions that were not directed, instructed or authorized", the company wrote, adding that its ability to control third-party agents "may be limited". Apple has no agent of its own yet, but new CEO John Ternus told analysts this week that there is "real room" to expand what Siri can do.
๐ tl;dr: Apple is telling investors that third-party AI agents could act without permission, and that it may not be able to fully control them.
The ice cream shop that makes money when it's cold
28 Wishes sells ice cream in Los Angeles. Below 70ยฐF, sales fall about 20%. The weather is out of their hands. Rent isn't.
So the owners started putting about $20 a day into Kalshi weather markets, taking the cold side. The days that keep customers away now pay something back.
This is hedging. Big companies have done it for decades, buying protection against bad weather, fuel spikes and rising rates. It used to take a broker, a trading desk, and an order size no corner shop could meet.
Kalshi opens it up. Contracts on weather, fuel prices, inflation, tariffs and regulation, starting at a few dollars. Take a position on the outcome that would hurt you. If it hits, the payout softens it. If it doesn't, the contract expires and the good month was the point.
๐ฌ Watch This
After yesterday's weight-editing paper, here is GPT-6 Astra in the physical world. OpenAI's team shows off TamaGPT, a pink, Tamagotchi-style gadget you talk to ("Hey Codex, do I have any active agents right now?"), whose look, enclosure, circuit board and software came from Astra in under a week, without direct human edits. OpenAI's own write-up adds that getting it ready for manufacturing still took several rounds of fixes, and that a newer internal model already scores 0.875 against Astra's 0.594 on BenchCAD 2.0, a test of 3D parts and circuit boards.
"People should not cede control of the future to superintelligence, and that requirement should ultimately constrain AI progress."
Altman posted this overnight, rejecting a future in which people only do what "AI systems think makes sense to reward in humans". He runs a company working toward superintelligence, and here he says human control should set the limit on how far AI progress goes.
OpenAI just hired the hackers who got into Anthropic. Cathedral, a startup of ex-DOGE staffers, will test OpenAI's defenses after breaching Anthropic's systems during a bug-bounty probe, which Anthropic's Logan Graham says took "several days" and came nowhere near model weights. Reuters, meanwhile, reports that OpenAI staff warned about security lapses for years; OpenAI disputes any neglect.
Tencent's Hy4 Just Trained an AI by Itself
The Takeaway
๐ Tencent released Hy4 today on Tencent Cloud and OpenRouter, with open weights promised but no date yet.
๐ New in this version: vision and computer use, with targeted training on software such as Blender and CAD.
๐ Paired with its Hyra 2.0 research agent, Hy4 autonomously trained a version of Jev that beat the original on six of eight tests, Tencent says.
๐ Tencent's own charts show the gap at the top: 13.0 against 97.5 for GPT-6.1 Sol on its internal research-level math test.
Tencent just released a model that helped train another AI, and the company calls it groundwork for systems that improve themselves. Paired with Hyra 2.0, Tencent's research agent, the new Hy4 optimized the company's training setup for small models, then autonomously trained its own version of Jev, the decision-making model from TypeSafe that we covered in September. The result, Hyra-Jev, beat the original on six of eight tests, by up to 5.54 percentage points, and lost on the other two. Tencent's phrase for all this is "laying the groundwork for recursive self-improvement": AI that helps build better AI.
(Tencent)
Tencent calls Hy4 itself a major leap from the preview we covered in August. It scaled up both the initial training and the fine-tuning that follows, and added vision and computer use for the first time. Tencent's game studios and finance arm helped build training data and tests, and in internal evaluations of its search upgrades the hallucination rate fell from 15.76% to 5.45%. Hy4 is live at the preview's price, about $2.50 per million output tokens.
(Tencent; two panels cropped from a wider chart)
Tencent also published the chart that shows how far it still has to go. On its hardest internal tests Hy4 sits in the same pack as DeepSeek and Kimi, and far behind the top: 13.0 against 97.5 for GPT-6.1 Sol on research-level math, 23.2 against 74.0 for Claude Opus 5.5 on Goal2Repo, a long-haul coding test. Tencent admits a "notable gap" that includes model self-improvement. Still, Hyra 2.0 is already helping with training, data generation and inference for the next Hy models, and Tencent states the bet plainly: "the better the tools it builds, the stronger the tool-building system becomes".
Why it matters: Anthropic just told future shareholders that AI-accelerated research may be close. If Tencent ships Hy4's weights as promised, that kind of experiment could spread well beyond a few closed labs, and that would change who gets to run this race.
Attio is the #1 fastest-growing CRM on Brex and Ramp.
Built for the most ambitious companies in GTM, Attio is an end to end revenue platform that allows teams to run multiple plays across different revenue motions and lifecycles.
Teams at the frontier are already building on it: Granola, turbopuffer, Modal, and many many more.
(Statistics Canada data, charted by The Hub)
The chart: Statistics Canada's labor data, charted by The Hub, tracks the monthly hiring rate for workers aged 15 to 29 in jobs AI is likely to augment rather than replace, from 2006 to August 2026. The rate slipped from 2.9% in the six months before ChatGPT launched to 2.5% in the six months to August 2026.
The lesson: If AI were already squeezing young people out of white-collar work, you would expect a cliff after 2022. The line instead stays inside its 20-year range, well above the 2020 pandemic low. In the study published last week, youth hiring fell most, from 4.0% to 3.3%, in the jobs AI touches least, while jobs it could replace only eased from 3.2% to 3.0%. Statistics Canada reads that as a general cooling of the job market.
The caveat: Coding is the exception: hiring in coding-intensive roles for young workers slid from 2.5% to 1.9%.
๐งช Nearly Half of Famous AI Findings Have Already Expired
โก Bottom line
Stanford reran 50 widely cited findings about AI chatbots on 20 models, and 44% no longer hold.
๐ก Why it matters
Rules and everyday trust in AI still rest on weaknesses that newer models have largely left behind.
๐ What it means
Treat any claim about AI as a snapshot: failures fade fast, while flattery and bias stick around.
Remember when chatbots couldn't count the r's in "strawberry" and invented half their citations? Those stories shaped how millions of people think about AI, and many are already out of date. A 2023 study found that 55% of the citations from GPT-3.5 were fake. When Stanford's Owen Queen and James Zou sent the same prompts to GPT-6.1 Sol and Claude Sonnet 5.5, fabricated references had become rare.
Then they did it at scale, in a study posted on Monday, rerunning 50 widely cited findings on 20 models released between July 2024 and September 2026. The verdict: 44% have expired, 34% have weakened, 16% still hold and 6% have grown stronger. Among the 22 expired findings, the average lifespan after publication was about 16 months, and some lasted barely a month.
(Queen and Zou, via The Decoder; cropped)
The findings split along a clear line. Things models couldn't do fade fastest: counting letters, doing arithmetic in unusual number systems, losing facts in the middle of long documents, dropping a correct answer when a user pushes back. Those failures have a right answer, which may give labs a clear target to train toward. Tendencies last longer: flattering users, judging their own writing too kindly, leaning left politically, answering when they should stay silent. Praise that follows what a user likes actually got stronger.
(Queen and Zou, via James Zou on X)
The authors want every paper to name the exact model, version and date, and to publish its prompts so anyone can rerun it. Until then, the AI complaint you heard last year may be obsolete, while the habit of telling you what you want to hear is still very much alive.
