Subscribe
Share
Share
Embed
Hosts: James & Maya
In this episode:
• Welcome to the Pivot 5 daily briefing. Hi, I'm James. Today we start with an AI that isn't writing code — it's designing proteins that actually stick ...
• Hi, I'm Maya. Let's look at what actually happened with Ant
Pivot5 | 5 Headlines & Unprompted
James: Welcome to the Pivot 5 daily briefing. Hi, I'm James. Today we start with an AI that isn't writing code — it's designing proteins that actually stick to their targets in the lab.
Maya: Hi, I'm Maya. Let's look at what actually happened with Anthropic's Claude and those protein binders.
Maya: Contract labs tested 1,320 minibinders that Claude agents designed. Of those, 354 bound their targets — a 26.8 percent success rate. Anthropic compares that to a more typical 10 to 15 percent, which would roughly double the fraction of designs worth carrying forward.
James: Here's why this matters for anyone in drug discovery: a higher hit rate could mean fewer wasted lab runs per candidate, and shorter cycles from idea to a molecule that works. But the caveats are real, so read the number carefully.
Maya: Worth noting, Anthropic ran no human-expert control campaign, so we can't compare Claude directly against skilled designers. And the work hasn't been independently reviewed, so treat 26.8 percent as a claim to test, not a settled benchmark. Now to a very different kind of AI money — chips.
James: Marvell just handed Google a $12.2 billion warrant as part of a custom chip deal. Marvell says it will build AI inference accelerators, storage, networking, memory interface controllers, and near-memory computing for Google.
Maya: Investors liked it. Marvell shares jumped more than 11 percent in premarket trading. The warrant essentially ties Marvell's fortunes to Google's spending.
James: The consequence is concentration. When one customer's warrant moves your stock double digits, your revenue now rises and falls with their order schedule, and a single pause in Google's buildout could reset Marvell's outlook overnight. Speaking of costs, our next story is one nobody's talking about yet.
Maya: 404 Media reports that Amazon is buying rare and obscure books in bulk, then destroying them to scan the pages for AI training.
James: European booksellers, including Ireland's Kennys, say they've received orders for thousands of unusual titles. Some now say they may start refusing those orders outright.
Maya: The direct consequence: if sellers refuse, that supply of scanned text dries up, and rare copies get pulled out of circulation for good in the process. The real story is how physical books are becoming raw material. Next, the traffic numbers behind the AI scare stories.
James: Greg Jarboe at Search Engine Journal checked the figures behind Tony Uphoff's widely shared post on AI-driven traffic loss. Most of them held up, including a roughly 42 percent click drop.
Maya: But one number didn't. A Wall Street Journal chart showing an 85 percent collapse at Business Insider had already been corrected to 43 percent — three weeks before Uphoff cited it.
James: The consequence for leaders: the underlying trend is real, but a stale figure spread anyway and made the decline look nearly twice as steep. Check when a chart was published, and whether it was later corrected, before you plan a budget around it. Now to Washington.
Maya: The FDA published a discussion paper floating competence tests for generative AI in medical devices. It covers foundation models and agentic systems, and comments are due October 19.
James: The FDA stresses this is not draft guidance and not proposed policy. It's early thinking, out loud.
Maya: The consequence is a window. Companies building AI medical tools can shape the rules now, before anything binding arrives. Miss the October 19 deadline and you're reacting to standards someone else wrote.
James: Let's move to our rapid-fire round for the rest of the headlines.
Maya: Pocket, a 12-person Y Combinator startup, says its $129 phone-mounted recorder hit a $100 million run rate in eight months, with over 200,000 units sold and growth near 50 percent a month. A run rate projects the current pace, not money already banked.
James: Pew finds 55 percent of Americans under 30 are now more worried than excited by AI — the first majority since it began asking in 2021. And 73 percent expect AI to mean fewer jobs over the next two decades, up from 61 percent.
Maya: Replit launched Free Mode running on OpenAI's cheaper GPT-5.6 Luna, now the default for its $20 Core and $100 Pro tiers. Replit's AI chief says it only became viable after OpenAI cut Luna prices 80 percent on July 30.
James: Rillet told Fortune it raised $100 million at a $1 billion valuation for AI agents that run inside the general ledger. That takes it past $200 million in funding in 14 months, with more than 600 customers.
Maya: Apple researchers analyzed 21,000 multi-turn conversations across four models. Human evaluators rated self-referential and relationship-building behavior less appropriate from a chatbot than a person, while boundary-setting was judged more appropriate.
James: The through-line: bold AI claims are everywhere, from lab results to traffic charts, and the value is in checking which ones actually hold up.
Maya: That's your Pivot 5 daily briefing.