The Download: the next big thing in LLMs and how AI academic research is shifting

Plus: Nvidia has secured $500 billion from Wall Street for AI infrastructure.
This is today's edition of The Download, our weekday newsletter that provides a daily dose of what's going on in the world of technology.
These startups are chasing the next big thing in LLMs
Nine years after Google researchers introduced the transformer, this family of neural networks has become the engine inside every major large language model. But transformers are starting to show their age.
As LLMs get bigger and better, transformers have become a bottleneck. Their dense attention mechanism becomes increasingly expensive as the amount of text grows, and they’re not great at keeping track of a lot of information at once.
Here are four new ideas for how to solve the transformer problem—innovations that could change LLMs for good, making them faster, far more efficient, and (maybe) even smarter.
—Will Douglas Heaven
This story is from MIT Technology Review’s What’s Next series, which looks across industries, trends, and technologies to give you a first look at the future. You can read the rest of them here.
AI professors are negotiating the new realities of academic research
—Grace Huckins
Last week, I headed to a hotel in Mountain View, California, to join some of the most accomplished, and some of the most promising, AI researchers in the world. I was hosting roundtable interviews and speaking at a media training for a convening of the Schmidt Sciences AI2050 program, an initiative funded by Eric and Wendy Schmidt that supports academics whose work involves AI.
The fellows list is a who’s who of AI luminaries, and though not all of them made it out to the Bay, every time I turned a corner I saw a scientist whom I’d interviewed previously or whose research I admired. It’s a weird time for university AI researchers, who make up most of the AI2050 group. Read Grace’s story to find out why, and what could be coming next.
This story is from The Algorithm, our weekly AI newsletter. Sign up to receive it in your inbox every Monday.
The must-reads
I’ve combed the internet to find you today’s most fun/important/scary/fascinating stories about technology.
1 Nvidia has secured $500 billion from Wall Street for AI infrastructure
It’s struck deals with BlackRock, Goldman Sachs, and four others. (BBC)
+ Showing the pull of AI compute for institutional investors. (Reuters $)
+ And that AI infrastructure is becoming a new asset class. (CNBC)
2 Mark Zuckerberg's new manifesto says open-source AI can save the US
It presents a utopian vision of personalized “superintelligence.” (Guardian)
+ And arrived the same day as Meta’s new, open-source model. (NYT $)
+ Zuckerberg said he plans to launch more of these models. (WSJ $)
+ And pit Meta against Chinese open-weight developers. (SCMP)
3 Bernie Sanders has called on Silicon Valley to “pause AI development”
He noted that AI giants have pledged to do this if necessary for safety.
+ And warned that lawmakers will step in if no action is taken. (Guardian)
+ House Democrats are already pressing AI leaders over rogue models. (WP $)
+ A populist backlash is building against AI. (MIT Technology Review)
4 A US court will allow thousands of social media lawsuits to proceed
The suits target addictive mechanisms used by Meta, TikTok, Google, and Snapchat. (Axios)
+ They claim the platforms are designed to hook young users. (Reuters $)
+ Can we repair the internet? (MIT Technology Review)
5 Unitree's IPO is more than 8,000 times oversubscribed by retail
The Chinese humanoid firm raised $900 million ahead of its listing. (Reuters $)
+ Its pricing for the Shanghai IPO values the company at $9 billion. (FT $)
6 Flock’s car-tracking cameras are facing a bipartisan backlash
The surveillance network has spread rapidly across the US. (NYT $)
+ Flock also plans to chase shoplifters with drones. (MIT Technology Review)
7 China is breaking up AI relationships
Beijing has introduced new rules for emotionally interactive AI. (Rest of World)
+ It’s surprisingly easy to fall for a chatbot. (MIT Technology Review)
8 An AI tool claims to pick the best 1% of scientific papers
But researchers doubt that AI can reliably judge scientific quality. (Nature)
9 The AI slop backlash is working
It’s pushing platforms to restrict AI-generated content. (Wired $)
10 An 82-year-old rejected $26 million to turn her farm into a data center
She criticised the environmental impacts of data centers. (Fortune)
Quote of the day
“It is not too late to avoid disaster. Stop building machines that humans cannot control.”
—Senator Bernie Sanders urges Sam Altman, Dario Amodei, and Mark Zuckerberg to pause all AI development in a letter.
One More Thing
The race to make the perfect baby is creating an ethical mess
A new field of science is using genetic sequencing to predict what kind of person an embryo might become. Some parents turn to these tests to avoid devastating genetic disorders, while a much smaller group are willing to pay tens of thousands of dollars to optimize for intelligence, appearance, and personality.
Customers, however, may not be getting what they’re paying for. Genetics experts have highlighted the potential deficiencies of this testing for years, while its underlying assumptions have made these companies a political lightning rod.
As this technology edges toward the mainstream, scientists and ethicists are racing to confront the implications—for our social contract, for future generations, and for our very understanding of what it means to be human. Read the full story.
—Julia Black
We can still have nice things
A place for comfort, fun, and distraction to brighten up your day. (Got any ideas? Drop me a line.)
+ An eagle-eyed border collie is taking the game of fetch into new waters.
+ musicForprogramming has made a valiant attempt to produce the perfect tunes for sustained concentration.
+ When kids design playgrounds, they create a cheerful mix of giant chess, pink basketball courts—and lava.
+ This power metal version of the “Back to the Future” music is an epic reinvention of the film’s classic theme.
Deep Dive
The Download
The Download: Claude’s inner workings and OpenAI’s “super app”
Plus: OpenAI has unveiled its long-awaited "super app."
The Download: Claude’s inner workings, and the future of world models
Plus: New York has become the first state to enact a data center moratorium.
The Download: the future of chipmaking and Anthropic’s government clash
Plus: Meta is pausing an AI training program that tracks workers’ keystrokes.
The Download: a reality check for geoengineering and the science of interoception
Plus: SpaceX is now valued higher than Amazon.
Stay connected
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.

Facts Only

* Nvidia secured $500 billion from Wall Street for AI infrastructure deals with BlackRock, Goldman Sachs, and four others.
* Researchers are developing new ideas to solve the transformer problem related to LLMs being a bottleneck.
* Academic researchers are participating in initiatives like the Schmidt Sciences AI2050 program.
* Lawsuits target addictive mechanisms used by Meta, TikTok, Google, and Snapchat.
* Unitree’s IPO was oversubscribed by retail investors.
* Beijing introduced new rules for emotionally interactive AI.
* An AI tool claiming to pick the best 1% of scientific papers faces doubt from researchers regarding quality judgment.
* A court is set to allow thousands of social media lawsuits to proceed targeting addictive mechanisms.

Executive Summary

The technology landscape is seeing shifts in Large Language Models (LLMs) driven by the limitations of the transformer architecture, which are proving to be a bottleneck as models grow larger. Researchers are exploring new concepts to overcome these limitations, aiming for LLMs that are faster, more efficient, and better at retaining information. Concurrently, academic AI research is being reshaped by prominent figures in the field, with researchers engaging in discussions about the realities of current AI development. In the broader technology sphere, significant financial activity involves infrastructure, as Nvidia secured substantial investment from Wall Street for AI compute. Public discourse reflects tension between rapid AI advancement and calls for safety, with some leaders urging a pause in development while others see open-source models as a path forward. Furthermore, societal concerns are manifesting in legal actions targeting social media platforms regarding addictive design, and there is emerging debate around the ethical implications of genetic sequencing for human optimization.

Full Take

The narrative highlights a tension between exponential technological capability and the lagging ethical, infrastructural, and regulatory frameworks attempting to govern it. The focus on overcoming the transformer bottleneck signals a fundamental architectural limit in current LLM scaling, suggesting that future advances may require paradigm shifts beyond dense attention mechanisms. Simultaneously, the convergence of high-stakes financial investment into AI infrastructure and public safety concerns regarding autonomous systems (social media tracking, humanoid robotics) reveals a societal struggle to establish control over powerful emergent technologies. The engagement of academic leaders alongside activist calls for pauses in development demonstrates that the debate is not confined to technical solutions but extends deeply into political and philosophical domains concerning human agency and the definition of intelligence. The pattern involves an acceleration in capability outpacing consensus on governance, where financial incentives and public anxiety simultaneously push for speed and caution, often creating a space for self-regulation or reactive legislation rather than proactive alignment with long-term ethical goals.
Bridge Questions: If architectural solutions to the transformer bottleneck prove insufficient, what non-computational constraints will define the next stage of intelligence development? How can frameworks be built that incentivize safety and efficiency across competing corporate and academic interests, rather than relying on reactive public backlash or government intervention? What are the long-term societal costs associated with optimizing for traits like intelligence and appearance through genetic manipulation?

Sentinel — Human

Confidence

The text reads like a curated summary or newsletter compilation, employing a distinct journalistic voice that synthesizes disparate technology and ethical stories.

Signals Detected

Moderate sentence length variance; uses journalistic style with some inherent rhythm.

The text shifts smoothly between high-level tech trends (LLMs, transformers) and specific current events (Nvidia deals, political calls), indicating a curated editorial voice.

Structured presentation of multiple linked news items suggests editorial selection rather than raw data aggregation.

References to specific sources (BBC, Reuters, Guardian, MIT Tech Review) and named figures appear grounded, though the framing is clearly editorialized.

Human Indicators

The inclusion of newsletter references ('The Download', 'The Algorithm') and quoted segments from specific authors (Douglas Heaven, Grace Huckins, Julia Black) strongly suggest human curation and narrative framing.

The juxtaposition of highly technical topics with socio-political commentary (AI safety, genetic testing ethics) points toward a synthesized editorial perspective rather than pure data dumping.