Thanks to everyone who contributed to GiveDirectly<\/a>! Because my readers gave more than $20,000, my wife and I donated an additional $10,000.<\/p>\n 2025 has been a huge year for AI: a flurry of new models, broad adoption of coding agents, and exploding corporate investment were all major themes. It\u2019s also been a big year for self-driving cars. Waymo tripled weekly rides, began driverless operations in several new cities, and started offering freeway service. Tesla launched robotaxi services in Austin and San Francisco.<\/p>\n What will 2026 bring? We asked eight friends of Understanding AI to contribute predictions, and threw another nine in ourselves. We give a confidence score for each prediction; a prediction with 90% confidence should be right nine times out of ten.<\/p>\n We don\u2019t believe AI is a bubble on the verge of popping, but neither do we think we\u2019re close to a \u201cfast takeoff\u201d driven by the invention of artificial general intelligence. Rather, we expect models to continue improving their capabilities \u2014 but we think it will take a while for the full impact to be felt across the economy.<\/p>\n Timothy B. Lee<\/p>\n In 2024, the five main hyperscalers \u2014 Google, Microsoft, Amazon, Meta, and Oracle \u2014 had $241 billion in capital expenditures. This year, those same companies are on track to spend more than $400 billion.<\/p>\n This rapidly escalating spending is a big reason many people believe that there\u2019s a bubble in the AI industry. As we\u2019ve reported<\/a>, tech companies are now investing more, as a percentage of the economy, than the peak year of spending on the Apollo Project or the Interstate Highway System. Many people believe that this level of spending is simply unsustainable.<\/p>\n But I don\u2019t buy it. Industry leaders like Mark Zuckerberg and Satya Nadella have said<\/a> they aren\u2019t building these data centers to prepare for speculative future demand \u2014 they\u2019re just racing to keep up with orders their customers are placing right now. Corporate America is excited about AI and spending unprecedented sums on new AI services.<\/p>\n I don\u2019t expect Big Tech\u2019s capital spending to grow as much in 2026 as it did in 2025, but I do expect it to grow, ultimately exceeding $500 billion for the year.<\/p>\n Timothy B. Lee<\/p>\n Anthropic and OpenAI have both enjoyed impressive revenue growth in 2025.<\/p>\n OpenAI expects<\/a> to generate more than $13 billion for the calendar year, and to end the year with annual recurring revenue<\/a> around $20 billion. A leaked internal document<\/a> indicated OpenAI is aiming for $30 billion in revenue in 2026 \u2014 slightly more than double the 2025 figure.<\/p>\n Anthropic expects<\/a> to generate around $4.7 billion in revenue in 2025. In October, the company said its annual recurring revenue had risen to \u201calmost $7 billion.\u201d<\/a> The company is aiming for 2026 revenue of $15 billion.<\/p>\n I predict that both companies will hit these targets \u2014 and perhaps exceed them. The capabilities of AI models have improved a lot over the last year, and I expect there is a ton of room for businesses to automate parts of their operations even without new model capabilities.<\/p>\n Kai Williams<\/p>\n LLMs have a \u201ccontext window,\u201d the maximum number of tokens they can process. A larger context window lets an LLM tackle more complex tasks, but it is more expensive to run.<\/p>\n When ChatGPT came out in November 2022, it could only process 8,192<\/a> tokens at once. Over the following year and a half, context windows from the major providers increased dramatically. OpenAI started offering a 128,000 token window with GPT-4 Turbo in November 2023. The same month, Anthropic released Claude 2.1, which offered 200,000 token windows. And Google started offering one million tokens of context with Gemini 1.5 Pro in February 2024 \u2014 which it later expanded to two million tokens.<\/p>\n Since then, progress has slowed. Anthropic has not changed its default context size since Claude 2.1. GPT-5.2 has a 400,000 token context window, but that\u2019s less than GPT-4.1, released last April. And Google\u2019s largest context window has shrunk to one million.<\/p>\n I expect context windows to stay fairly constant in 2026. As Tim explained in November<\/a>, larger context window sizes brush up against limitations in the transformer architecture. For most tasks with current capabilities, smaller context windows are cheaper and just as effective. In 2026, there might be some coding-related LLMs \u2014 where it\u2019s useful for the LLM to be able to read an entire codebase \u2014 that have larger context windows. But I predict the context lengths of general-purpose frontier models will stay about the same over the next year.<\/p>\n Timothy B. Lee<\/p>\n The year 2027 has acquired a totemic status in some corners of the AI world. In 2024, former OpenAI researcher Leopold Aschenbrenner penned a widely-read series of essays<\/a> predicting a \u201cfast takeoff\u201d in 2027. Then in April 2025, an all-star team of researchers published AI 2027<\/a>, a detailed forecast for rapid AI progress. They forecast that by the 2027 holiday season, GDP will be \u201cballooning.\u201d One AI 2027 author suggested<\/a> that this could eventually lead to annual GDP growth rates as high as 50%.<\/p>\n They don\u2019t make a specific prediction about 2026, but if these predictions are close to right, we should start seeing signs of it by the end of 2026. If we\u2019re on the cusp of an AI-powered takeoff, that should translate to above-average GDP growth, right?<\/p>\n So here\u2019s my prediction: inflation-adjusted GDP in the third quarter of 2026 will not be more than 3.5% higher than the third quarter of 2025. Over the last decade, year-over-year GDP growth has only been faster than 3.5% in late 2021 and early 2022, a period when the economy was bouncing back from Covid. Outside of that period, year-over-year growth of real GDP has ranged from 1.4% to 3.4%.<\/p>\n I expect the AI industry to continue growing at a healthy pace, and this should provide a modest boost to the US economy. Indeed, data center construction has been supporting the economy over the last year. But I expect the boost from data center construction to be a fraction of one percent \u2014 not enough to push overall economic growth outside its normal range.<\/p>\n Kai Williams<\/p>\n The AI evaluation organization METR released<\/a> the original version of this chart in March. They found that every seven months, the length of software engineering tasks that leading AI models were capable of completing (with a 50% success rate) was doubling. Note that the y-axis of this chart is on a log scale, so the straight line represents an exponential increase.<\/p>\n By mid-2025, LLM releases seemed to be improving more quickly, doubling successful task lengths in just five months. METR estimates that Claude Opus 4.5, released in November, could complete software tasks (with at least a 50% success rate) that took humans nearly five hours.<\/p>\n I predict that this faster trend will continue in 2026. AI companies will have access to significantly more computational resources in 2026 as the first gigawatt-scale clusters start operating<\/a> early in the year, and LLM coding agents are starting to speed up AI development. Still, there are reasons to be skeptical. Both pre-training (with imitation learning) and post-training (with reinforcement learning) have shown diminishing returns.<\/p>\n Whatever happens, whether METR\u2019s line will continue to hold is a crucial question. If the faster trend line holds, the strongest AI models will be at 50% reliability for 20-hour software tasks \u2014 half of a software engineer\u2019s work week.<\/p>\n James Grimmelmann<\/a>, professor at Cornell Tech and Cornell Law School<\/p>\n So far, AI companies are winning against the lawsuits that pose truly existential threats \u2014 most notably, courts in the US, EU, and UK have all held that it\u2019s not copyright infringement to train a model. But for everything else, the courts have been putting real operational limits on them. Anthropic is paying $1.5 billion<\/a> to settle claims that it trained on downloads from shadow libraries, and multiple courts have held or suggested that they need real guardrails against infringing outputs.<\/p>\n I expect the same thing to happen beyond copyright, too: courts won\u2019t enjoin AI companies out of existence, but they will impose serious high-dollar consequences if the companies don\u2019t take reasonable steps to prevent easily predictable harms. It may still take a head on a pike \u2014 my money is on Perplexity\u2019s \u2014 but I expect AI companies to get the message in 2026.<\/p>\n Steve Newman<\/a>, author of Second Thoughts<\/a><\/p>\n There are credible concerns that AI could eventually enable various disaster scenarios. For instance, an advanced AI might help create a chemical or biological weapon, or carry out a devastating cyberattack. This isn\u2019t entirely hypothetical; Anthropic recently uncovered<\/a> a group using its agentic coding tools to carry out cyberattacks with minimal human supervision. And AIs are starting to exhibit advanced capabilities<\/a> in these domains.<\/p>\n However, I do not believe there will be any major \u201cAI catastrophe\u201d in 2026. More precisely: there will be no unusual physical or economic catastrophe (dramatically larger than past incidents of a similar nature) in which AI plays a crucial enabling role. For instance, no unusually impactful bio, cyber, or chemical attack.<\/p>\n Why? It always takes longer than expected for technology to find practical applications \u2014 even bad applications<\/a>. And AI model providers are taking steps<\/a> to make it harder to misuse their models.<\/p>\n Of course, people may jump to blame AI for things that might have happened anyway, just as some tech CEOs blamed AI for layoffs that were triggered by over-hiring during Covid.<\/p>\n Andrew Lee<\/a>, CEO of Tasklet<\/a> (and Tim\u2019s brother)<\/p>\n The Model Context Protocol<\/a> was designed to give AI assistants a standardized way to interact with external tools and data sources. Since its introduction in late 2024, it has exploded in popularity.<\/p>\n But here\u2019s the thing: modern LLMs are already smart enough to reason about how to use conventional APIs directly, given just a description of that API. And those descriptions that MCP servers provide? They\u2019re already baked into the training data or accessible on public websites.<\/p>\n Agents built to access APIs directly can be simpler and more flexible, and they can connect to any service \u2014 not just the ones that support MCP.<\/p>\n By the end of 2026, I predict MCP will be seen as an unnecessary abstraction that adds complexity without meaningful benefit. Major vendors will stop investing in it.<\/p>\n Daniel Abreu Marques, author of The AV Market Strategist<\/a><\/p>\n Waymo has world-class autonomy, broad regulatory acceptance, and a maturing multi-city playbook. But vehicle availability remains a major bottleneck. Waymo is scheduled to begin using vehicles from the Chinese automaker Zeekr in the coming months, but tariff barriers and geopolitical pressures will limit the size of its Zeekr-based fleet. Waymo has also signed a deal with Hyundai<\/a>, but volume production likely won\u2019t begin until after 2026. So for the next year, fleet growth will remain incremental.<\/p>\n Chinese AV players operate under a different set of constraints. Companies like Pony.ai, Baidu Apollo Go, and WeRide have already demonstrated mass-production capability. For example, when Pony rolled out its Gen-7 platform, it reduced<\/a> its bill of materials cost by 70%. Chinese companies are scaling fleets across China, the Middle East, and Europe simultaneously.<\/p>\n At the moment, Waymo has about 2,500 vehicles in its commercial fleet. The biggest Chinese company is probably Pony.ai, with around 1,000 vehicles. Pony.ai is aiming for 3,000 vehicles by the end of 2026, while Waymo will need 4,000 to 6,000 vehicles to meet its year-end goal of one million weekly rides.<\/p>\n But if Waymo\u2019s supply chain ramps slower than expected due to unforeseen problems or delays \u2014 and Chinese players continue to ramp up production volume \u2014 then at least one of them could surpass Waymo in total global robotaxi fleet size by the end of 2026.<\/p>\n Sophia Tung, content editor of the Ride AI newsletter<\/a><\/p>\n Currently many customer-owned vehicles have advanced driverless systems (known as \u201clevel two\u201d in industry jargon), but none are capable of fully driverless operations (\u201clevel four\u201d). I predict that will change in 2026: you\u2019ll be able to buy a car that\u2019s capable of operating with no one behind the wheel \u2014 at least in some limited areas.<\/p>\n One company that might offer such a vehicle is Tensor<\/a>, formerly AutoX. Tensor is working with younger, more eager automakers that already ship vehicles in the US, like VinFast<\/a>, to manufacture and integrate their vehicles. The manufacturing hurdles, while significant, are not insurmountable.<\/p>\n Many people expect Tesla to ship the first fully driverless customer-owned vehicle, but I think that\u2019s unlikely. Tesla is in a fairly comfortable position. Its driver-assistance system performs well enough most of the time. Users believe it is \u201cpretty much\u201d a fully driverless system. Being years behind Waymo<\/a> in the robotaxi market hasn\u2019t hurt Tesla\u2019s credibility with its fans. So Tesla can probably retain the loyalty of its customers even if a little-known startup like Tensor introduces a customer-owned driverless vehicle before Tesla enables driverless operation for its customers.<\/p>\n Tensor has a vested interest in being first and flashiest in the market. It could launch a vehicle that can operate with no driver within a very limited area and credibly claim a first-to-market win. Tensor runs driverless robotaxi testing programs and therefore understands the risks involved. Tesla, in contrast, probably does not want to assume liability or responsibility for accidents caused by its system. So I expect Tesla to wait, observe how Tensor performs, and then adjust its own strategy accordingly.<\/p>\n Timothy B. Lee<\/p>\n In June, Tesla delivered on Elon Musk\u2019s promise<\/a> to launch a driverless taxi service in Austin. But it did so in a sneaky way. There was no one in the driver\u2019s seat, but every Robotaxi had a safety monitor in the passenger seat. When Tesla began offering Robotaxi rides in the San Francisco Bay Area, those vehicles had safety drivers.<\/p>\n It was the latest example of Elon Musk overpromising and underdelivering on self-driving technology. This has led many Tesla skeptics to dismiss Tesla\u2019s self-driving program entirely, arguing that Tesla\u2019s current approach simply isn\u2019t capable of full autonomy.<\/p>\n I don\u2019t buy it. Elon Musk tends to achieve ambitious technical goals eventually. And Tesla has been making genuine progress on its self-driving technology. Indeed, in mid-December, videos<\/a> started to circulate showing Teslas on public roads with no one inside. I think that suggests that Tesla is nearly ready to debut genuinely driverless vehicles, with no Tesla employees anywhere in the vehicle.<\/p>\n Before Tesla fans get too excited, it\u2019s worth noting that Waymo began its first fully driverless service in 2020. Despite that, Waymo didn\u2019t expand commercial service to a second city \u2014 San Francisco \u2014 until 2023. Waymo\u2019s earliest driverless vehicles were extremely cautious and relied heavily on remote assistance, making rapid expansion impractical. I expect the same will be true for Tesla \u2014 the first truly driverless Robotaxis will arrive in 2026, but technical and logistical challenges will limit how rapidly they expand.<\/p>\n Kai Williams<\/p>\n Current LLMs are autoregressive, which means they generate tokens one at a time. But this isn\u2019t the only way that AI models can produce outputs. Another type of generation is diffusion. The basic idea is to train the model to progressively remove noise from an input. When paired with a prompt, a diffusion model can turn random noise into solid outputs.<\/p>\n For a while, diffusion models were the standard way to make image models, but it wasn\u2019t as clear how to adapt that to text models. In 2025, this changed. In February, the startup Inception Labs released Mercury<\/a>, a text diffusion model aimed at coding. In May, Google announced Gemini Diffusion<\/a> as a beta release.<\/p>\n Diffusion models have several key advantages over standard models. For one, they\u2019re much faster because they generate many tokens at once. They also might learn from data more efficiently, at least according to a July study<\/a> by Carnegie Mellon researchers.<\/p>\n While I don\u2019t expect diffusion models to supplant autoregressive models, I think there will be more interest in this space, with at least one established lab (Chinese or American) releasing a diffusion-based LLM for mainstream use.<\/p>\n Charlie Guo, author of Artificial Ignorance<\/a><\/p>\n AI has become a vessel for a number of different anxieties: misinformation, surveillance, psychosis, water usage, and \u201cBig Tech\u201d power in general. As a result, opposition to AI is quickly becoming a bipartisan issue. One example: back in June, Ted Cruz attempted to add an AI regulation moratorium to the budget reconciliation bill (not unlike President Trump\u2019s recent executive order), but it failed 99-1<\/a>.<\/p>\n Interestingly, there are at least two well-funded pro-AI super PACs:<\/p>\n Leading The Future, with over $100 million from prominent Silicon Valley investors, and<\/p>\n Meta California, with tens of millions from Facebook\u2019s parent company.<\/p>\n Meanwhile, there\u2019s no equally organized counterweight on the anti-AI side. This feels like an unstable equilibrium, and I expect to see a group solely dedicated to lobbying against AI-friendly policies by the end of 2026.<\/p>\n Abi Olvera, author of Positive Sum<\/a><\/p>\n We\u2019ve already seen extensive media coverage of cases like the Character.AI lawsuit<\/a>, where a teen\u2019s death became national news. I expect suicides involving LLMs to generate even more media attention in 2026. Specifically, I predict that news mentions of \u201cAI\u201d and \u201csuicide\u201d in media databases will be at least three times higher in 2026 than in 2025.<\/p>\n But increased coverage doesn\u2019t mean increased deaths. The US suicide rate will likely continue on its baseline trends.<\/p>\n The US suicide rate is currently near a historic peak after a mostly steady rise since 2000. While the rate remained high through 2023, recent data shows a meaningful decrease in 2024. I expect suicide rates to stay stable or lower, reverting back toward average away from the 2018 and 2022 peaks.<\/p>\n Florian Brand, editor at the Interconnects<\/a> newsletter<\/p>\n In late 2024, Qwen 2.5, made by the Chinese firm Alibaba, surpassed the best American open model Llama 3. In 2025, we got a lot of insanely good Chinese models<\/a> \u2014 DeepSeek R1, Qwen3, Kimi K2 \u2014 and American open models fell behind. Meta\u2019s Llama 4, Google\u2019s Gemma 3, and other releases were good models for their size, but didn\u2019t reach the frontier. American investment in open weights started to flag; there have been rumors since the summer that Meta is switching to closed models.<\/p>\n But things could change next year. Through advocacy like the ATOM Project<\/a> (led by Nathan Lambert, the founder of Interconnects), more Western companies have indicated interest in building open-weight models. In late 2025, there has been an uptick in solid American\/Western open model releases like Mistral 3, Olmo 3, Rnj, and Trinity. Right now, those models are behind in raw performance, but I predict that this will change in 2026 as Western labs keep up their current momentum. American companies still have substantial resources, and organizations like Nvidia \u2014 which announced<\/a> in December it would release a 500 billion parameter model \u2014 seem ready to invest.<\/p>\n Kai Williams<\/p>\n This fall, OpenAI and Meta both released platforms for short-form AI-generated video. Initially, Sora caught<\/a> all of the positive attention: the app came with a new video generation model and a clever mechanic around making deepfakes of your friends. Meta\u2019s Vibes initially fell flat. Sora quickly became the number one app in Apple\u2019s App Store, while the Meta AI app, which includes Vibes, languished around position 75.<\/p>\n Today, however, the momentum has seemed to shift. Sora\u2019s initial excitement has seemed to wear off<\/a> as the novelty of AI videos faded. Meanwhile, Vibes has been growing, albeit slowly, hitting two million daily active users in mid-November, according to Business Insider<\/a>. Today, the Meta AI app ranks higher on the App Store than Sora.<\/p>\n I think this reversal will continue. From personal experience, Sora\u2019s recommendation algorithm seems very clunky, and Meta is very skilled at building compelling products that grow its user base. I wouldn\u2019t count out Mark Zuckerberg when it comes to growing a social media app.<\/p>\n Timothy B. Lee<\/p>\n This is one of the few places where Kai and I disagreed, so I thought it would be fun to air both sides of the argument.<\/p>\n I was initially impressed by Sora\u2019s clever product design, but the app hasn\u2019t held my attention since my October writeup<\/a>. However, toward the end of that writeup I said this:<\/p>\n I expect the jokes to get funnier as the Sora audience grows. Another obvious direction is licensing content from Hollywood. I expect many users would love to put themselves into scenes involving Harry Potter, Star Wars, or other famous fictional worlds. Right now, Sora tersely declines such requests due to copyright concerns. But that could change if OpenAI writes big enough checks to the owners of these franchises.<\/p>\n![\"\"](\"https:\/\/www.newsbeep.com\/il\/wp-content\/uploads\/2026\/01\/https:\/\/substack-post-media.s3.amazonaws.com\/public\/images\/596b0158-7d1f-4af9-a834-431a703a9abb_3900.jpeg\")
<\/a>Wax sculptures of Mark Zuckerberg, Jeff Bezos, and other tech industry leaders were mounted to robot dogs at a recent exibit by artist Mike Winkelmann in Miami. (Photo by CHANDAN KHANNA \/ AFP via Getty Images)<\/p>\n![\"\"](\"https:\/\/www.newsbeep.com\/il\/wp-content\/uploads\/2026\/01\/https:\/\/substack-post-media.s3.amazonaws.com\/public\/images\/ebcc28e5-23d3-4e05-aeef-5d5658c90515_1342.jpeg\")
<\/a><\/p>\n