MIT Technology Review surveys transformer alternatives and Nvidia's $500 billion AI infrastructure bet
MIT Technology Review's August 11, 2026 edition of its weekday newsletter The Download leads with a feature on the search for a successor to the transformer. Nine years after Google researchers introduced the architecture, it powers every major large language model, but its dense attention mechanism grows more expensive as the amount of text it processes increases, and it has trouble keeping track of large amounts of information at once. The story, part of the outlet's What's Next series and bylined to Will Douglas Heaven, presents four separate ideas researchers and startups are pursuing to fix that problem, changes the piece says could make LLMs faster, far more efficient, and possibly smarter. The digest itself does not name or describe what the four ideas are; it points readers to the full story for that detail.
A second feature, from The Algorithm newsletter and bylined to Grace Huckins, covers a convening of the Schmidt Sciences AI2050 program at a hotel in Mountain View, California, that Huckins attended the previous week to host roundtable interviews and speak at a media training session. AI2050, funded by Eric and Wendy Schmidt, supports academics whose work involves AI, and its fellows list is described as a who's who of AI researchers. The newsletter calls it a weird time for university AI researchers, who make up most of the AI2050 group, and points to Huckins's full story for why and for what could come next, without spelling out the specifics itself.
The must-read roundup that follows covers ten same-day stories, each credited to an outside outlet. Nvidia has secured $500 billion from Wall Street for AI infrastructure, with deals struck with BlackRock, Goldman Sachs, and four other unnamed firms, according to the BBC, Reuters, and CNBC. Mark Zuckerberg published a manifesto arguing open-source AI can save the US and outlining a vision of personalized superintelligence, released the same day as a new open-source Meta model; Zuckerberg said he plans to launch more such models, per the Guardian, the New York Times, the Wall Street Journal, and the South China Morning Post. Senator Bernie Sanders called on Silicon Valley to pause AI development, noting that AI companies have already pledged to do so if necessary for safety, and warned that lawmakers will step in if no action follows; the newsletter's quote of the day is Sanders's line, urging Sam Altman, Dario Amodei, and Mark Zuckerberg by letter to stop building machines humans cannot control. House Democrats are separately pressing AI leaders over rogue models, per the Washington Post, and MIT Technology Review notes a populist backlash building against AI more broadly. A US court will allow thousands of social media lawsuits against Meta, TikTok, Google, and Snapchat to proceed, with plaintiffs claiming the platforms are designed to hook young users, according to Axios and Reuters; no total damages figure, plaintiff count, or court name is given. Unitree's Shanghai IPO was oversubscribed more than 8,000 times by retail investors after the Chinese humanoid robotics firm raised $900 million ahead of the listing, with pricing that values the company at $9 billion, per Reuters and the Financial Times. Flock's car tracking camera network is facing bipartisan backlash after spreading rapidly across the US, and the company also plans to use drones to chase shoplifters, according to the New York Times. Beijing has introduced new rules for emotionally interactive AI as part of an effort to break up AI relationships, per Rest of World, though the digest does not detail what the rules require. An AI tool claims to identify the best 1% of scientific papers, but researchers doubt AI can reliably judge scientific quality, according to Nature. Wired reports that a backlash against AI slop is pushing platforms to restrict AI generated content. Fortune reports that an 82-year-old woman rejected $26 million to turn her farm into a data center, citing the environmental impacts of data centers, though the digest does not give her name, her location, or the buyer's identity.
The newsletter's One More Thing section, bylined to Julia Black, covers embryo genetic screening: a new field uses genetic sequencing to predict what kind of person an embryo might become. Most parents who use these tests are trying to avoid devastating genetic disorders, while a smaller group pays tens of thousands of dollars to try to optimize for intelligence, appearance, and personality. The piece notes customers may not get what they are paying for, since genetics experts have flagged the testing's potential deficiencies for years and its underlying assumptions have made the companies a political lightning rod.
Key facts
- Nine years after Google introduced the transformer, MIT Technology Review's What's Next series presents four unnamed ideas to fix the architecture's attention bottleneck.
- Nvidia secured $500 billion in AI infrastructure financing from Wall Street, with deals including BlackRock, Goldman Sachs, and four other unnamed firms.
- Senator Bernie Sanders wrote to Sam Altman, Dario Amodei, and Mark Zuckerberg urging them to pause AI development, warning that lawmakers will step in otherwise.
- Unitree's Shanghai IPO was oversubscribed more than 8,000 times by retail investors after the firm raised $900 million at a $9 billion valuation.
- An 82-year-old woman rejected $26 million to turn her farm into a data center, citing the environmental impacts of data centers.
Why it matters
Nine years after Google researchers introduced the transformer, the architecture that still powers every major large language model is showing strain: its dense attention mechanism grows more expensive as input length increases, and it struggles to track large amounts of information at once. MIT Technology Review's What's Next series counts four separate efforts under way to address that, framed as ideas that could make future LLMs faster, more efficient, and possibly smarter, though this digest does not describe what the four approaches actually are. The same edition tracks how AI research itself is being reshaped: the Schmidt Sciences AI2050 program, backed by Eric and Wendy Schmidt, gathered what the newsletter calls a who's who of AI researchers in Mountain View, and the newsletter frames university AI researchers, who make up most of that group, as living through a weird time. Around those two features, the day's headlines show how much money and political pressure now surrounds AI: Nvidia lined up $500 billion in Wall Street financing for AI infrastructure, and Senator Bernie Sanders published a letter urging Sam Altman, Dario Amodei, and Mark Zuckerberg to pause AI development entirely.
Who it affects
AI research teams and startups working on alternatives to the transformer architecture; university based AI researchers such as the Schmidt Sciences AI2050 fellows navigating a changed academic environment; Wall Street institutions, including BlackRock, Goldman Sachs, and four other firms, now financing Nvidia's AI infrastructure buildout; retail investors who oversubscribed Unitree's Shanghai IPO more than 8,000 times; Meta, TikTok, Google, and Snapchat, named in social media addiction lawsuits a US court allowed to proceed; users of emotionally interactive AI in China, now subject to new government rules; scientific publishing, as an AI tool claims to identify the top 1% of papers; and an 82-year-old farm owner who turned down $26 million to keep her land out of data center development.
How to use it
There is no product here to adopt; the newsletter functions as a set of pointers to fuller reporting. Readers who want the substance behind the headline claims need to follow through: MIT Technology Review's full What's Next piece for what the four transformer alternatives actually are, Grace Huckins's companion story in The Algorithm for what is unsettling university AI researchers, and the linked outlets, including the BBC, Reuters, the Guardian, the New York Times, the Wall Street Journal, the South China Morning Post, Axios, the Financial Times, Rest of World, Nature, Wired, and Fortune, for the reporting behind each must-read item.
How solid is it
The two feature items are MIT Technology Review's own reporting, bylined to Will Douglas Heaven and Grace Huckins. The must-read roundup and the quote of the day are secondhand: each line credits and links to an outside outlet, the BBC and Reuters and CNBC for the Nvidia financing, the Guardian, the New York Times, the Wall Street Journal, and the South China Morning Post for the Zuckerberg manifesto, the Guardian and the Washington Post for the Sanders story, Axios and Reuters for the lawsuit ruling, Reuters and the Financial Times for the Unitree IPO, the New York Times for Flock, Rest of World for China's chatbot rules, Nature for the paper ranking tool, Wired for the AI slop backlash, and Fortune for the farm owner, so their numbers and framing rest on those outlets' own reporting rather than on MIT Technology Review's.
Risks and caveats
The digest withholds specifics at several points: it does not name or describe the four proposed fixes to the transformer, does not say what makes this moment weird for university researchers beyond pointing to a separate story, does not name the four Wall Street firms beyond BlackRock and Goldman Sachs, gives no detail on what China's new rules for emotionally interactive AI require, and omits the farm owner's name and location, the data center buyer's identity, and the total dollar figure, plaintiff count, or court name behind the social media addiction lawsuits. Treat each must-read line as a pointer to its source outlet rather than a complete account.
“It is not too late to avoid disaster. Stop building machines that humans cannot control.”
— Senator Bernie Sanders, in a letter urging Sam Altman, Dario Amodei, and Mark Zuckerberg to pause AI development