Here’s what’s on our plate today:

• 🧪 The antibiotic lesson AI has not yet heard.

• 📰 Google's price-war shot, half of Americans fear AI job loss, Siri's big WWDC moment.

• 🛠️ Three tools worth trying: Papers with Code, Hugging Face Leaderboard, NSF NAIRR.

• 🗳️ Poll: Is AI research leaving universities a problem?

Let’s dive in. No floaties needed…

The Laboratory

TL;DR

• Industry ate the lab: Nearly 90% of notable AI models now come from private companies, up from 60% the year before, and 71% of new U.S. AI doctorates take industry jobs, leaving academia with one in five.

• Money explains everything: Meta offered some researchers packages worth nearly $100M, Anthropic is valued at $965B, and universities are competing for donated compute time through the NSF’s AI Research Resource.

• Openness is declining fast: Of 95 notable models released in 2025, 80 shipped without training code. Stanford’s transparency index dropped the average score from 58 to 40 out of 100 in a single year.

• The countercase is real: DeepMind’s AlphaFold, which cracked protein structure prediction, was released openly and is now a standard part of the graduate curriculum. Private labs do produce landmark science.

• The stake: If foundational research becomes a trade secret rather than a published paper, the next generation of breakthroughs may never have the open scientific base they need to exist.

The antibiotic lesson AI has not yet heard

For most of the twentieth century, new antibiotics arrived on a reliable schedule. Drug companies and university labs kept finding them, and the assumption was that the supply would continue for as long as there were infections to treat. Then the discoveries slowed, and the cause was not scientific.

According to a review in the Journal of Antibiotics, the business had changed: selling large volumes of existing drugs became far more profitable than searching for new ones, so the companies best equipped to search moved on to more lucrative work. The bacteria did not disappear, and neither did the open scientific questions. What faded was the financial reason to keep asking them, and once that incentive weakened, the labs and budgets built around it gradually drifted elsewhere.

That is the part worth holding on to, because it describes something that can happen in any field: the work keeps going on momentum long after the reason to fund it has quietly disappeared. And a version of this may now be forming in artificial intelligence, even though it looks different on the surface. AI is not short of breakthroughs; new results arrive faster than most people can keep up with. The concern sits beneath the output, where the research happens, who can afford to do it, and how much of it the rest of the world ever gets to see.

A founder’s $100M wager

The clearest sign that industry insiders share this concern came in June 2025. Andy Konwinski, who co-founded the data company Databricks and the AI search company Perplexity, put $100M of his own money into a new nonprofit called the Laude Institute, built to keep foundational AI research rooted in universities. According to Bloomberg, Konwinski, who made his fortune turning academic work into commercial products, has said he wants to fund “species-level” problems and give university researchers an alternative to the constant pull of private labs. It is a telling move: someone who profited from the shift toward industry is now spending his own money to slow it down.

The migration in numbers

The shift he is worried about is evident in the data. According to Stanford’s annual AI Index, nearly 90% of notable AI models came from industry in 2024, up from 60% the year before, and the 2026 edition found that figure surpassed 90% in 2025. In other words, an increasing share of AI research is being conducted inside private companies rather than universities, concentrating talent and resources in organizations whose primary goal is building products rather than advancing open science.

The 2024 Index reported that the share of new American AI doctorates taking industry jobs rose from roughly 41% in 2011 to nearly 71% by 2022, while the share heading into academia dropped to one in five. A study in AI & Society that tracked researchers across millions of papers found an added twist: the academics most likely to leave for industry are the young, highly cited ones at top universities, and their work tends to become less original and less influential once they make the move.

Why universities cannot keep up

The reason for the migration is straightforward, and it echoes the antibiotic story. Building a frontier model now requires computing power on a scale no university can match. The AI Index notes that the computing power needed to train top models has been doubling approximately every 5 months, and that the United States alone hosts 5,427 data centers, more than 10 times as many as any other country. The money has gone the same way. According to CNBC, Meta offered some researchers pay packages reported near $100M, and the biggest labs now carry valuations larger than the economies of most countries, with Anthropic reaching $965B and OpenAI $852B by 2026.

Compared to these figures, public alternatives look small by comparison. The National Science Foundation’s National AI Research Resource, a federal program that provides academics with shared access to computing resources, has supported more than 600 projects and 6k students, but it relies on donated processing time rather than cash, which could severely limit access.

From published papers to trade secrets

Beyond the money and computational gaps, money and hardware explain who gets to build the leading models. And with deep-pocketed interests influencing research, there is another fundamental shift in how research is not just conducted but also shared.

As research has moved indoors, companies have disclosed less about it. Stanford's Foundation Model Transparency Index, which scores developers on how much they reveal about their systems, found the average score fell to 40 out of 100 from 58 the year before, with Meta’s dropping from 60 to 31.

The 2026 AI Index reported that of 95 notable models released in 2025, 80 came out without the training code that would let outsiders see how they were built. The knowledge still exists, but it increasingly takes the form of a trade secret rather than a published paper that any lab in the world can read, repeat, and build on.

The case on the other side

However, the picture is not all bleak, and the strongest counterargument deserves room.

The private labs are doing some of the most important science of the era, and they sometimes share it freely. The 2024 Nobel Prize in Chemistry went to Demis Hassabis and John Jumper of Google DeepMind for AlphaFold, a system that predicted the structures of almost every known protein, a problem biologists had wrestled with for 50 years.

DeepMind released the results openly, and AlphaFold is now a standard tool taught to biology graduate students worldwide. The same AI Index that documents industry’s dominance also notes that universities still produce the most highly cited research, a reminder that volume and influence are not the same thing. There is also a reasonable view that this is simply a cycle, that talent always rushes toward whatever technology is rising, and that the flow back toward open research resumes once the field matures and the salaries cool.

What the medicine case warns

The antibiotic story cannot tell us how AI will evolve, but it does show what to watch for. The breakthroughs that made today’s AI boom possible, including the model architectures, training techniques, and scientific insights now worth billions, largely came from open-ended research, often conducted in universities and research labs long before anyone knew their commercial value. Increasingly, however, that model is being replaced by one focused on shipping products and capturing market share.

So far, it has produced impressive results, but it has also concentrated research capabilities in fewer organizations and reduced the amount of knowledge shared publicly. Whether a field can continue generating the fundamental discoveries it depends on under those conditions remains an open question. The lesson from antibiotics is simply that innovation can continue for years on accumulated momentum, even after the incentives that once sustained it have begun to fade.

Thursday Poll

3 Things Worth Trying

• Papers with Code: Tracks ML research alongside open implementations, a great way to see how much (or how little) cutting-edge work still ships with reproducible code.

• Hugging Face Open LLM Leaderboard: Ranks open models on shared benchmarks, useful for seeing how the open ecosystem stacks up against closed frontier labs.

• NSF NAIRR Pilot: The federal program giving academics shared access to AI compute, worth exploring if you're a researcher hunting for resources outside Big Tech.

Quick Bits, No Fluff

• Google's price-war shot: Google fired a warning shot in the AI subscription price wars, undercutting rivals in a move that could reset consumer expectations for premium AI pricing.

• Half of Americans fear AI job loss: A new Reuters/Ipsos poll finds that half worry AI could put someone in their household out of work, a sign of how deeply economic anxiety has spread.

• Siri's big WWDC moment: Apple used WWDC to unveil a long-promised AI overhaul of Siri in iOS 27, its most serious attempt yet to catch up in the assistant race.

The Toolkit

• Leonardo AI: AI image and video generator with fine-grained creative controls, built for designers, marketers, and game studios who need consistent style at scale.

• Modal: Serverless cloud for running Python and AI workloads, lets you spin up GPUs in seconds without touching infrastructure.

• Quillbot: AI writing assistant that paraphrases, summarizes, and rewrites text on demand, useful for tightening drafts or escaping your own voice.

Rate This Edition