Contrary Research Rundown #102
The cutting-edge of supercomputers, plus new memos on Benchling, Vercel, and more
Announcements
Contrary Research is growing our team, hiring a Research Analyst. Come help us build the best starting place to understand any private tech company. To apply, email research@contrary.com.
Research Rundown
In the latest essay from Foundations & Frontiers, we dive into the history and evolution of supercomputers.
From the earliest days of computing to the cutting-edge machines of today, supercomputers have stood as monuments to human ingenuity and our relentless pursuit of computational power. These machines have been at the forefront of scientific discovery, national security, and technological innovation for decades.
The journey of supercomputers follows two parallel paths: the drive for miniaturization that has given us powerful smartphones, and the quest to build ever-larger and more powerful computing systems. This dual approach has led to remarkable advancements in both personal computing and supercomputing, each serving different but equally important purposes:
"The ENIAC, the first general-purpose computer, occupied 300 square feet, consumed 150 kilowatts, and processed about 500 floating point operations per second. Stories say that when the ENIAC was switched on at the University of Pennsylvania, the lights in Philadelphia dimmed."
Today, we have supercomputers capable of performing more than a quintillion calculations per second. The Frontier supercomputer at Oak Ridge National Laboratory represents this new frontier:
"Oak Ridge National Lab, where Frontier is headquartered, likes to say that if every person on earth could do a single operation, like an addition or a multiplication, every second — it would take them four years to do what Frontier can do in just one second."
Yet, even as traditional supercomputers grow in power, new approaches are emerging that could revolutionize the field. Companies like Cerebras are developing solutions that challenge the conventional supercomputing paradigm:
"Rather than wiring up tens of thousands of GPUs together through elaborate cabling, Cerebras etches this type of interconnection right into the silicon."
Side note, you can look forward to our massive memo on Cerebras next week! The ongoing evolution of supercomputing technology reflects a broader drive to push the boundaries of what's computationally possible. As our world becomes increasingly data-driven and interconnected, the work in advancing supercomputing capabilities becomes ever more crucial in solving some of humanity's most pressing challenges.
However, one thing is certain: the era of supercomputers is far from over. It's only just beginning. Read the full essay here.
Benchling offers a cloud-based platform for life sciences R&D with digital tools to standardize data from lab work across an organization. To learn more, read our full memo here and check out some open roles below:
IT DevOps Engineer (High Seniority) - San Francisco, CA
Software Engineer, Infrastructure (Release Engineering, High Seniority) - San Francisco, CA
Vercel is positioning itself at the forefront of this evolving landscape as a cloud-based platform for building, previewing, and deploying user-facing applications. To learn more, read our full memo here and check out some open roles below:
Software Engineer, Deployment Experience - Remote (US or Canada)
Software Engineer, Frontend Platform - Remote (US)
Check out some standout roles from this week.
Ramp | New York or San Francisco - Software Engineer (Backend), Software Engineer (Applied AI), Staff Software Engineer (Forward Deployed)
Railway | Remote - Senior Infrastructure Engineer, Senior Full-Stack Engineer (Product), Community Engineer
Nomic | New York, NY - Front End/Web Engineer, Senior Devops Engineer, Senior Software Engineer
Dual Entry | Remote - Senior/Staff Backend Engineer, Senior/Staff Frontend Engineer, Senior/Staff Quality Assurance Engineer
OpenAI introduced o1 — a new series of reasoning models. “We've developed a new series of AI models designed to spend more time thinking before they respond. They can reason through complex tasks and solve harder problems than previous models in science, coding, and math.”
OpenAI’s released of Strawberry (o1) introduces a shift towards inference-time scaling, focusing on reasoning over large models, leveraging tools like browser and code verifier to factor out knowledge, while optimizing compute during inference for better performance and continuous improvement, similar to AlphaGo's MCTS method.
OpenAI’s new models can effectively fact-check itself, but it comes at the cost of being much slower than existing models.
SpaceX's Polaris Dawn mission completed the world's first commercial spacewalk. The mission also tested new spacesuit technology developed by SpaceX for future private and commercial space missions.
Anduril revealed Barracuda, a family of autonomous missiles. Barracuda missiles can be mass-produced at a scale to beat China, with 50% fewer parts and 95% fewer production tools.
Kodiak Robotics is getting its driverless truck technology into broader commercial operation by going off-road, where it can take advantage of simpler environments with fewer potential hazards compared to highways.
Cohost, a would-be rival to Twitter founded with an anti-Big Tech manifesto, is shutting down due to lack of funding and burnout, despite its ambition to disrupt the tech giants.
Klaviyo's quietly launched a native server-side tracking feature which allows them to potentially track events that may have been missed with their previous API-based approach, such as abandoned checkouts. This could provide Klaviyo customers with more comprehensive data on user behavior.
Okta’s CFO stated the company uses “AI to help our developers write code, and we’ve seen productivity gains between 10% and 30%.”
The relationship between coverage (fraction of problems solved) and the number of samples is often log-linear and can be modeled with an exponentiated power law, suggesting the existence of inference-time scaling laws. This suggests that there may be fundamental scaling laws governing the performance of large language models as the number of inference samples is increased.
Journalist Evan Ratliff conducted an experiment for his podcast "Shell Game," gradually replacing himself with an AI voice clone in personal and professional settings to explore how convincingly the AI could mimic him. Take a listen.
The UK regulator designated Microsoft's acquihire of Inflection AI as a "quasi-merger" that falls under its merger control jurisdiction, even though Microsoft did not fully acquire Inflection.
Art requires making countless small-scale choices during the creative process, which is fundamentally different from the way AI language models generate text or images based on limited prompts. Perhaps, human artists are safe from their AI replacements (for now).
Robert Yang, CEO and Founder of Altera, introduced Project Sid — the first simulations of 1000-plus autonomous agents collaborating in a virtual world (i.e. Minecraft) with emergent economy, culture, religion, and government.
Researchers were able to predict the correct letters someone typed on the Apple Vision Pro's virtual keyboard with up to 92% accuracy, just by analyzing the eye movements of the user's 3D avatar during video calls.
Revolut is now growing faster than Facebook and Snapchat in Europe and is downloaded nearly as often as Instagram, WhatsApp, and TikTok.
Oura has announced that it has acquired Veri, a Helsinki-based metabolic health startup.
Klarna CEO Sebastian Siemiatkowski claims his company is shutting Salesforce and Workday because they can replicate the functionality using AI.
More than half of Wiz's users aren't even in infosec, indicating the company is expanding beyond its core security audience.
Apple's Indian workforce will comprise 200K direct jobs by March-end, with women making up 70% of the roles.
China cannot yet replicate ASML's advanced lithography systems, which are crucial for producing the most advanced computer chips, so the Netherlands' decision to limit ASML's ability to service and provide spare parts in China is a potentially painful blow to Beijing's efforts to develop a world-class chip industry.
Telegram has quietly removed language from its FAQ page that previously stated private chats were protected and the company would not process any requests related to them, indicating a shift in its moderation policies after the arrest of its CEO.
Wildfires in Southern California forced Flexport to evacuate their e-commerce fulfillment facility in San Bernardino.
Waymo's adoption is accelerating and it could achieve a low-to-mid single-digit percentage share of Phoenix and San Francisco rideshare markets by 2025 which could impact partnerships with Uber and Lyft.
Andrej Karpathy, one of OpenAI’s cofounders, suggests that AI, particularly transformers, could soon surpass the human brain in efficiency and memory.
We’re thrilled to be hosting our first NYC Tech Talk of the year featuring eng leads and founders from Ramp, Warp, Railway, Together AI & Moment. It's an evening built by engineers for engineers — each company will live demo their latest product innovations for leading builders in NYC. Register here for a chance to join.
At Contrary Research, we’re building the best starting place to understand private tech companies. We can't do it alone, nor would we want to. We focus on bringing together a variety of different perspectives.
That's why we're opening applications for our Research Fellowship. In the past, we've worked with software engineers, product managers, investors, and more. If you're interested in researching and writing about tech companies, apply here!