AI's Pepsi Challenge
DeepSeek dropped the AI bombshell of the year (so far).?
?
The company's latest AI model, developed in China for a fraction of the cost of comparable offerings from leaders like OpenAI and Google, soared in popularity last month. Its ascendance unleashed a torrent of speculation and debate. Tech-heavy stock markets flinched.
?
Not only did DeepSeek’s R1 include a host of clever engineering hacks, its makers released it as an open-weights model with a matching preprint paper on the arXiv. While many labs are keeping their sauce secret, DeepSeek allowed users a peek under the hood. Last week, the company publicly released even more detail, including code.
?
The headline cost of $6 million has been fiercely debated—here’s a great in-depth analysis—but the model has challenged a few core assumptions in AI.?
?
First, AI is getting efficient fast. This isn’t a new trend, but DeepSeek underscored it. Stock in Nvidia, maker of the industry’s most popular chips, took it on the chin. Investors questioned whether AI companies would need as many Nvidia chips as they’d once thought if the cost of developing frontier models drops an order of magnitude. Utilities stocks also nosedived. More efficient training would translate to less demand for power.
?
Second, it seems competitors can catch up to frontier “reasoning” models—the first of which, OpenAI’s o1, was released late last year—faster than older frontier models. Whereas OpenAI had a clear lead for a year or more after GPT-4’s release, its lead in “reasoning” models evaporated in months.
?
So far, the first point appears to have been overly reactionary. In quarterly earnings calls—after DeepSeek’s release—Alphabet, Microsoft, Amazon, and Meta said they were still committed to $300 billion in capital expenditures, largely related to AI infrastructure, this year alone. That’s in addition to OpenAI and Softbank’s previously announced $500-billion, four-year Stargate data-center project. Unsurprisingly, Nvidia recently reported another quarter of blockbuster growth and projected more to come.
?
The rationale? Even if fewer chips are used to train AI models, more chips will be needed to run them. Reasoning models, like DeepSeek’s R1 and OpenAI’s o1 and o3, use more compute during inference than vanilla models like GPT-4. Further, as models across the spectrum become more efficient—and it’s likely we’ll see a similar trend in reasoning models too—people will use them more, not less. Finally, despite reports suggesting scaling laws are beginning to show diminishing returns in performance, leaders like OpenAI and Anthropic say they still plan to scale models further at the frontier.
?
The second point, however, may be more salient.
?
DeepSeek does seem to have rattled OpenAI. The company pushed out a slimmed down version of its top reasoning model, o3, and even made it available on a limited basis to its free tier of users. On a Reddit “Ask Me Anything” thread, CEO Sam Altman said the company has been on the “wrong side of history” when it comes to open-source AI.?
?
The company also previewed GPT-4.5, a dramatically scaled-up version of GPT-4. OpenAI says the model has better general knowledge, hallucinates less, and is more personable in conversation. But it underperforms o3 on math and coding—which reasoning models excel at—only shows subtle improvement on other benchmarks, and is a lot more expensive. Altman also said the company would release GPT-5 later this year. The model would be OpenAI’s first to blend traditional and reasoning models.
?
At a glance, OpenAI is no longer in rarified air. On the heels of DeepSeek, researchers at Stanford and the University of Washington showed they could match o1 and R1 on the cheap with a base model and a process called distillation, where a more advanced model trains a less advanced one. Meanwhile, GPT-5 won’t be the first blended model. That distinction goes to Anthropic, with the recently released Claude 3.7. Anthropic also beat OpenAI to the punch on computer-controlling agents last year.
?
It’s possible GPT-5 ups the ante some other way. But if not, it will be joining an increasingly crowded space in which AI products are converging.
?
Decades ago, Pepsi ran a marketing campaign called the “Pepsi Challenge.” Customers would take a blindfolded taste of Coke and Pepsi side by side and say which they preferred. Many were surprised to find they’d chosen Pepsi, and their reactions were turned into an influential ad campaign. Pepsi took it from a battle of brand loyalty—where Coke was dominant—to a battle of product quality, even though the difference between the two products was fairly subtle.
?
The commodification of AI is a trend worth watching this year. Until the next big AI advance arrives, it seems likely the battle of brands, bells, and whistles will reign. OpenAI is still in a strong position with ChatGPT, but there are those who prefer Anthropic’s Claude, and the benefits of open models are luring others.
For obvious reasons, investors are also interested in finding and funding the next big thing. OpenAI is said to be closing in on a round of $40 billion at an eye-popping $300 billion valuation. Meanwhile, new startups founded by OpenAI alumni Ilya Sutskever and Mira Murati are working to close $1 billion investment rounds at $30 billion and $9 billion valuations respectively. Neither has a product to show investors: The money is a bet on their ability to attract talent and push next-level AI. Indeed, Sutskever’s Safe Superintelligence claims it won’t release any product until it reaches its goal.
MORE NEWS
From the Future
?
?
Quasi QC. For years, Microsoft has been pursuing an exotic form of quantum computing they hope will vault them ahead of the competition. Topological quantum computers using qubits made of quasiparticles—these are collections of particles that together exhibit quantum behavior—should be far more stable than competing approaches. This month, the company announced a new chip, Majorana 1, with eight “topological qubits.” Meanwhile, Amazon got into the quantum race too with a nine-qubit chip, Ocelot. The chip employs error-correction involving two kinds of qubits—what they call a “cat” qubit and a standard qubit.
?
The zoo. The two announcements join a diverse crowd of quantum computers—including those employing neutral atoms, ions, photons, and superconducting loops of wire—and error-correction strategies. All have a unified goal: To realize practical quantum computing with error-corrected qubits. Microsoft is hoping to jump the queue with topological qubits stable enough to mitigate the errors plaguing competitors. Amazon believes its approach could achieve error-correction with far fewer than the million-plus qubits competitors will need for useful calculations. If they’re right, they could build a useful machine sooner.
?
Dead or alive? The Microsoft news received a good bit of attention, but there are caveats. The paper describing the work didn’t definitively show the team had created quasiparticles, just that they’d demonstrated a reliable way to detect them. Researchers said they'd observed quasiparticles in a 2018 paper that was retracted in 2021 and again in a 2023 paper. The team told MIT Technology Review they have unpublished results showing quasiparticles more definitively, but some physicists are skeptical. Amazon’s chip, meanwhile, is a useful proof-of-concept but needs significant improvements before it can scale. Still, for a problem this hard, the more approaches researchers pursue, the better.
?
Emerging from stealth this month, Inception brings a new approach to the land of large language models. Standard LLMs use an algorithm called a transformer to generate replies one word at a time. Inception, however, was inspired by image-generation AIs that use diffusion models. Inception’s AI resolves whole responses all at once. The startup says this dramatically reduces cost and increases speed without sacrificing performance. Whether the elevator pitch bears fruit has yet to be seen, but it does show innovation in AI isn’t dead. Another startup, Liquid AI, raised $250 million late last year for an efficient LLM-alternative called a liquid neural network.
?
Many people use smartwatches or rings to monitor their heart rates or steps, but MIT researchers say there’s a wealth of information beyond these single-point sensors. They believe computerized clothes could record even more health data. To that end, the team has developed conductive threads and flexible electronics—including memory, processing, and connectivity—that can be stitched into clothing and can run a simple neural network. With the threads working together, the computerized clothing identified exercises performed by wearers with 95 percent accuracy. It’s now being tested by members of the US Army and Navy during a month-long mission to the Arctic.
?
In a recent essay, Wired cofounder and senior maverick Kevin Kelly argues AI automation is well timed. Declining fertility suggests global population growth will peak and begin to fall for the first time in almost a thousand years this century. Because modern economic growth is closely linked to population growth, a declining population would be problematic. But Kelly says AI may take up the slack: We’re about to create “millions of AIs and robots and agents, who could potentially not only generate new and old things, but also consume them as well, and to grow the economy in a new and different way.”
?
?
CALENDAR
Upcoming Events
?
MAY 7-9 The Future of Biotech
The next trillion dollar opportunity is here. The Future of Biotech is your deep dive into the exponential potential of biotech. Over three days, discover how biotech will disrupt entire markets (far beyond healthcare) and what leaders, entrepreneurs and investors should be doing now to capitalize on this opportunity.
?
MARCH 9?
Exponential Mastery
Most people are drastically underestimating the pace of technological change that will permanently transform everything about our daily life in the next 24 months. This is your on-ramp which unveils the coming opportunities and the mindsets you need to capitalize on them. You’ll also get exclusive replay access to Peter Diamandis’ Abundance Summit.
Thanks for reading. We hope you enjoyed this month's updates and found something to inspire you on your exponential journey.
?
See you next month!
The Singularity Team