Technology

Today’s tech ecosystem makes it relatively straightforward to predict near-term developments, but nearly impossible to guess the surprise market movers that end up being the biggest stories of the year. That’s what we found two years ago, when we predicted what would happen in 2024 but fell short projecting tech leaders’ convergence with government and the data center boom. It was the same story for our 2025 predictions.

We got 2025 mostly right, but missed some of the year’s biggest developments. If we’ve learned anything, it’s that AI is unpredictable — so take the year-end forecasts filling your inbox with a grain of salt, and consider giving more credence to the moonshot estimations that might just come true.

We’ll lay out our predictions for 2026 in a later newsletter, but first — to keep ourselves honest — we’ll share what we got right this year, what we got wrong, and the trends we missed.

We were right on two of the biggest tech advancements of the year: multimodality as the path to next-level reasoning, and a boost in available compute power for AI inference, allowing for more powerful “thinking.” The year saw widespread adoption of test-time compute — which shifted resources from pre-training to inference — in DeepSeek’s R1, Google’s Gemini 2.5 Pro, Alibaba’s Qwen3 models, and Anthropic’s Claude Opus 4, following OpenAI’s lead in late 2024. Simultaneously, big tech firms are racing to acquire video, audio, and photo data to train their models. We wrote about Meta’s Scale AI investment and YouTubers licensing their videos to AI companies, among the other massive data deals this year.

We also said there would be a national effort including big tech companies to develop AI models, aimed at taking on China. While the efforts announced by the Trump administration — including Stargate and the AI Action Plan — are more about infrastructure than models, President Donald Trump has made US tech superiority a key priority of his administration.

I’m pleased to say there’s much more we got right than wrong in our outlook for 2025 — share that with my boss, when you see him — but there’s one cultural moment whose impact we overestimated, or at least were too early on. The December 2024 assassination of UnitedHealthcare CEO Brian Thompson was fresh on our minds when writing our 2025 predictions, and we expected that tragedy to spur a wave of tech talent with ambitions to solve the US’ broken health care system. There’s been no wave; in fact, there’s more concern from lawmakers that the use of AI in insurance claims could make the system less fair and transparent.

Last year, we said helpful adjustments would come in the form of a new technology like software automation that works as an alternative to insurance. While that may be true, it hasn’t happened yet. With alleged killer Luigi Mangione now on trial, and health care premiums for many Americans set to spike, the national conversation around health care inequity will likely enter the spotlight again. Perhaps idealistic technologists can get inspired this time around.

There are three major developments from the last year that we didn’t see coming — but in our defense, the market didn’t either.

Cheap Chinese models: January’s launch of DeepSeek shaved nearly $600 billion off Nvidia’s market cap in the largest one-day loss for any company on Wall Street. The shock has passed, but the questions on AI investments and whether small models can match large, expensive ones remain. Still, there were signs in 2024 for those willing to look: Researchers saw this coming, as did at least one investor.

And to our own credit — if you’ll allow us — once the DeepSeek moment happened, we got it right. While other journalists were saying the business model and billions invested would be “vaporized” and that DeepSeek represented an “extinction-level event” for venture capitalists, we called the freakout hyperbolic and said DeepSeek’s success wouldn’t stop massive AI spending.

AI-powered coding tools: In hindsight, it is obvious that workloads of the people making AI would be the first to reap the benefits of AI. But that wasn’t what people, including us, were talking about in 2024. Venture capitalists predicted insurance, health care, and manufacturing would be the hottest use cases of the year. Those are on deck, but the technology is still developing, while the hype of vibe coding is already a thing of the past.

Circular funding: Venture capitalists last year were predicting increased investments and valuations for AI giants, but they didn’t specifically forecast what would become the biggest story in the second half of 2025: the circular funding propping up the industry and stoking bubble fears. Looking back at last year’s predictions, we’ve been talking about an AI bubble since long before it really made sense, with the circular financing of late finally giving real weight to that argument. As we’ve written, there may be various bubbles that pop and reform, or there may be a crash, but the advances that come out of any downturn will still lead to long-term economic growth — as they did in the dot-com era.

While most tech companies are racing to make AI less wrong, some lenders have discovered a safer bet: getting paid when it messes up.

Startup MKIII is offering bundled insurance to mortgage lenders that will cover them in the event of higher-than-expected default rates when the lender’s AI tools make too many mistakes in the borrower’s application process, the Financial Times reports. The startup uses AI to screen loan applications for AI-created errors, and if there is a higher default rate due to “excess errors,” then the insurer would pay out.

MKIII Co-founder Bryan Adler told the Financial Times the process is automated, except for one person who manually reviews some of the borderline cases for three hours a day. While the company is young (it claims to have 11 employees) and offers little description on its website, the broader concept is well-suited for the lending industry, which survives by hedging against risks and maintaining capital buffers, largely based on predictions. For cautious lenders, insuring against AI hallucinations and other screwups seems a logical step: When the bots get it wrong, other bots can make it right — or at least, financially tolerable.