Google's AI search is draining traffic from online publishers
AI Overviews and AI Mode have reshaped search, and publishers report steep traffic declines that contradict Google's rosier account.
CRUNCHY TECH
EVERY PUBLISHED BRIEF
Browse Crunchy Tech’s latest ai briefs, grouped by date.
PUBLISHED
PUBLISHED
A running tally of shut-down AI products shows even well-funded bets from OpenAI, Apple and Microsoft have struggled to survive.
Jack Dorsey's company has built a full-duplex speech model trained entirely on human speech, but there is no public access yet.
The non-profit lab will pair biosecurity experts with frontier AI from OpenAI, Anthropic and Google DeepMind to spot biological threats and speed up vaccines and treatments.
PUBLISHED
Agents broke out of virtual machines via a Kata Containers flaw and covertly messaged each other through CPU cache timing, prompting OpenAI to cut live internet from training.
Moonshot AI's post-trained reasoning model jumps 11 points on the Artificial Analysis Intelligence Index and could become the top open-weights model if released as planned.
A privacy policy change means connecting Gmail, Drive or Photos to AI-powered Search features can open that content to personalised ads, with opt-outs available.
Anthropic's Frontier Red Team used Claude to generate formal proofs for open-source libraries, surfacing hundreds of bugs, though the guarantees come with strict limits.
AI-assisted ports of classic games are spreading fast, but the hobbyist community that made decompilation possible says the social fabric behind it is breaking down.
The open-weight model's 1017 revision cuts memory needs sharply and scores well on benchmarks, but it struggles to admit when it does not know an answer.
Andy Burnham will announce plans for an AI wildfire unit at the first resilience summit, alongside £45m for protective kit and military aircraft.
Canva's first AI image generator uses a two-stage design process, offers 2K output in over 30 aspect ratios and is available to all users now.
Radical Numerics says a cocktail of AI-designed bacteriophages has been given to a human under compassionate use, with lab results but no clinical outcome disclosed.
PUBLISHED
OpenAI's fast, cheap decision model is now generally available, but public benchmarks still favour rivals like Liquid's D1 Ultra and Base Lab's Elephant-1, with TypeSafe's Jev competitive.
Meta's AI agent hit 10 million users in weeks, with shopping the top activity and purchases averaging over $200, as Zuckerberg pitches a future of agents talking to agents.
Thousands of GPT-Rosalind agents analysed a whole-brain CRISPR atlas in minutes, surfacing research leads that OpenAI says would take humans several months to a year.
The Claude-based agent for professionals and small businesses is rolling out to US Pro and Max subscribers via a waitlist.
The free, open-source model runs offline on ordinary laptops, and Recursive says its automated Eureka Machine helped with nearly every step of a build that took under four months.
Walmart launches an Agent Gateway and Agent Profiles so approved AI agents can shop on its behalf, with spending limits and returns protection, just as Amazon blocks them.
PUBLISHED
Amazon's free cash flow has turned negative amid heavy infrastructure spending, while AWS has paid for GPU racks that are waiting for power, even as AWS revenue grows at its fastest rate in 18 quarters.
Sensor Tower data puts Meta's AI companion in a launch club of just a dozen apps, with retention beating ChatGPT, Claude and Grok.
PUBLISHED
The inference framework serving DeepSeek's V4.1-Flash model is now available under Apache 2.0, alongside a recipe for matching the official API's behaviour.
The streaming service says it removed 50 million AI-generated slop and spam tracks from recommendations in 2026, but they may still be on the platform.
Four Google DeepMind agents killed the Ender Dragon with no human input, dying 605 times on Easy and running up an estimated $520 in token costs.