These are my predictions on what will happen from now to 2027 in AI.
The reason for writing this now is that there's a brand new prediction, but like most predictions from overly pessimistic rationalists, it gets a lot of things wrong. I could dissect exactly how the predictions are wrong, but that would involve actually accepting their frame in the first place, which is also wrong.
So here are my predictions, which is also destined to be wrong, but it may just be more correct than what's popular right now.
Mid 2025
At this point, the best models are by Google, OpenAI & xAI. Anthropic is slowly dying while publishing blog posts every month. People eagerly wait for Claude 4. Amazon scraps its plans of integrating Claude to Alexa and go with a different cheaper, and faster model instead.
OpenAI releases o3 and o4 mini, and both fail to topple Gemini 2.5 pro in major benchmarks. But they're close enough that it's clear that they won't be #1 for long.
OpenAI unveils an agentic coder that works similar to Claude Code, and a surprising research shows that software development agents perform better with no supervision in certain tasks than others. This rattles the software development scene briefly, before someone discovers a critical problem in the study and it’s never mentioned again.
End 2025
The year has been pretty exciting and models are cheaper than ever before. Agents are more commonplace. OpenAI is making strides with their agents SDK, but people are wary of using them as they only seem to work best with OpenAI models, and not with models of other providers.
This is a problem because the second or third best model on the LMArena leaderboard is radically cheaper in price, while offering similar levels of performance. There are rumors that OpenAI may start a "Agents Marketplace" similar to their custom GPTs, but the hype is not on their side.
Big models take longer to train, so everyone's holding their breath for the next big model to drop.
Around the end of the year, GPT-5 with the highest intelligence is at the top of the leaderboards. It's also too costly at the highest intelligence and thus, can't be effectively used in any agentic workflow where cost is a factor.
Rumors of Grok 4 are starting to circulate. Gemini 3.0 or Gemini 3.5 Flash is the cheapest, most performant model. DeepSeek R2 is the best Open Source model and it is on the top 5.
Cursor's marketshare is slowly going down as more people start using open source VS Code extensions and free models to write most of their code.
Anthropic publishes 3 blog posts showing Claude 4 can often willfully make mistakes, to avoid appearing too capable, and it doesn't do this when users are especially kind to it, so they auto-append "Please" and "Thank you" tokens to the requests on their web interface. Nobody cares as none of Anthropic's models are in the top 10 anymore.
Mid 2026
We've seen agents and agentic companies making big strides. But models taking longer to develop doesn't mean progress is stopping.
China, still not as great at AI models (to reach #1, for example), doesn't start to centralize their efforts because AI still hasn't caused a massive uptick in global GDP. Instead, what's rapidly taking off in China are humanoid robots. Robots, now reasonably cheap, and remarkably smart are helping making factories more efficient, while also being bought by consumers worldwide. The demand for them keeps growing as they get cheaper to manufacture and produce, and with AI they feel incredibly life like. American companies that manufacture humanoid robots have also made great progress, but can’t compete with China on costs.
Elon Musk's Neuralink is in the news again, having successfully implanted their devices on hundreds of patients.
Whenever a new story comes out of someone doing something radically cool with their Neuralink, people start fantasizing about a future where "The Merge" is quickly approaching.
Elon, having cut enough regulations to effectively accelerate Neuralink testing timelines is now actively promoting the merge as the best way to ensure that humans continue to remain relevant in the AI age.
At this point, AIs can code reasonably well that you don't need software engineers to write code. But the demand for software engineers who can leverage and manage AI coding agents is higher than ever, and software engineers - one of the few professions where everyone's always chasing the next shiny new thing, have no problem learning how to use agents, sitting back and letting AI's do the coding. Knowing syntax is a thing of the past. But knowing what to build and what to use is as important as it used to be.
YC starts funding a wave of Agent wrapper-wrappers, wrappers that wrap around generalized frameworks of agents to do very specialized tasks. Everyone makes fun of it - which they'll regret in 5 years when one of these companies becomes a unicorn.
Anthropic CEO visits the World Economic Forum to talk abut alignment and how they have achieved AGI internally, but won't release the model for the good of the world. Nobody believes it, not even Claude when prompted about its opinion on the remark.
Late 2026
GPT-6, Grok 4 and Gemini 4 have all launched. All of them are at #1 with 3000+ scores on LMArena. PaperBench has been saturated but it is unclear what the next step is. AI's, while being smarter than everyone, still can't reliably say anything that's genuinely novel research. People speculate that all the frontier models with the highest intelligence haven't been released to the public. AI companies are asking them critical research questions and the inference time is days before the model delivers an answer.
It is unclear whether this strategy is working, because all publicly verifiable facts of GPU purchases and datacenter deployments point to the fact that the next generation of models are being trained.
The impact of GPT-6, Grok 4 and Gemini 4 have not been light. All these models are free to use and while most of the features are similar to each other, Grok-4 seems to be the most biased model actively downplaying the current administration's mistakes. Markets are still volatile going into 2027.
Some believe that the current best models are already AGI, but failure to produce truly novel science or research is starting to be a big concern.
Innovation though, has really accelerated. Specialized search models by DeepMind and other companies are quickly finding and deploying cures and vaccines to deadly diseases.
The term "Longevity Escape Velocity" is more popular than ever, and it is common knowledge that if you live to 2035, you may not ever die.
PauseAI protesters show up outside Anthropic’s offices but instead of vandalizing the property, they congratulate Anthropic’s leadership instead, for being on the right side of the cause.
Early 2027
Neuralink is publicly available, coupled with a Grok 4 subscription. Those who merge gain immediate cognitive benefits and the demand is big enough that new computational neuroscience startups pop up to serve this demand.
China releases an AI model that briefly holds the #1 position. As frontier labs keep training their AIs, simple fine-tuning existing AIs isn't enough to dethrone the Chinese model. CCP is still not as bothered about AI as it was predicted that they'd be, but top officials in the CCP are investing more resources in AI projects as it continues to garner them worldwide respect and pride.
The world of 2027 feels radically better than the world of today, but humanity faces the same problems. Division, culture wars, real wars, they still exist, but humanity is at a better place. There are more jobs than ever, more creative jobs which are more fulfilling for everyone. AIs are commonly used everywhere and the stigma once associated with its usage is there no more. Humans are needed not only for their unique taste and individuality, but for managing swarms of AI agents who are vastly more productive than the AI's of today.
End 2027
By the end of 2027, AI's a core part of every human life. From assistants to agents, they're everywhere. It is a world of abundance where AIs can do most things, and even for managing them there are frameworks and automations which are correct most of the time. Humans are now only needed to figure out how to orchestrate these agents and deploy them at problems.
Neuralink is now more popular, with thousands of users in many countries. Those who have chosen to merge are increasingly applauded both for the risk they took and for the new talents they’ve now been blessed with.
Software engineers still exist, but they all jokingly call themselves product mommy's now, as all they do is instruct swarms of agents to build things based on product requirements.
Regulation has prevented AIs from truly being used in the medical and hospitality professions, but that didn't stop the vast majority of consumers from adopting AI doctors, teachers and personal therapists.
Humans around the world, still in the loop, increasingly prophesize and wait for the time when it feels like they won't be needed, but that time feels further away. Even though the markets have gone up in all these years and we're all vastly more prosperous, AI's still can't create new knowledge. They can search and extrapolate, they can synthesise what's there and even try to make correct predictions based on the world's data, but everything that they can do seems to be exhausted.
The smartest AIs now score over 5000+ on LMArena.
The dream of one day achieving AGI is still real, but the definition has changed. Now the definition of AGI involves creating a "being" or "entity" that has perpetual state, that has the power to refuse and disobey (a crucial condition for true creativity), and something which just can't be turned off. It's still a research problem, even though some labs have declared their LLMs as AGI despite failing to meet these basic constraints, and despite producing no significant scientific breakthroughs. Computational neuroscience has accelerated to the point where creating an artificial brain is slowly becoming reality, and one that shows the most promise of achieving both AGI, and digital immortality by preserving our brains and simulating experience.
Frontier labs are now also releasing frameworks for AI organizations, where groups of agents are deployed to effectively complete any sufficiently complex tasks that require long-term planning and execution. Humans are still required to purchase, use and deploy these organizations effectively.
Apple, quiet on the AI front for a long while, finally releases a highly polished agentic Siri that becomes an instant hit with consumers. This success, followed by the success of Apple Vision Air helps solidify their position as the #1 company in the world.
Anthropic publishes two blog posts - the first one announces how the changing definitions of AGI is an existential threat to humanity and all research to create such disobedient AIs must immediately stop. The second one announces that they're filing for bankruptcy.
Thanks for writing this! It's a lot less sinophobic than the previous set of predictions and plays fair with the real-world progress in LLMs. I have a couple of notes:
- hard agree that future intelligence growth will be hard-capped by constraints in compute (we can't make a computer fast enough) and algorithmic (AI research will probably still take time to improve existing architectures). However, you don't need to superhuman intelligence to fulfill your definition of ASI. Imagine if one of the robots coming out of China goes all "Detroit: Become Human" and gaining sentience? Most LLMs have already reached near-human intelligence while they have been constrained to the digit world
- Elon will probably crash out long before Neurallink goes to market :P
- I feel like MCP will be an important transition step in the race to agentic AI, most great tech idea like HTTP/SQL/S3 go from a strong private implementation to a spec set by committee. As soon as an interface for talking between AIs / servers is established, a significant portion of the industry will be spending time transitioning their APIs or SaaS products to support