karpathy reviews LLMs’ year of 2025:
Reinforcement Learning from Verifiable Rewards (RLVR) emerged as the de facto new major stage to add to this mix. By training LLMs against automatically verifiable rewards across a number of environments (e.g. think math/code puzzles), the LLMs spontaneously develop strategies that look like "reasoning" to humans - they learn to break down problem solving into intermediate calculations and they learn a number of problem solving strategies for going back and forth to figure things out (see DeepSeek R1 paper for examples). These strategies would have been very difficult to achieve in the previous paradigms because it's not clear what the optimal reasoning traces and recoveries look like for the LLM - it has to find what works for it, via the optimization against rewards.
Unlike the SFT [supervised finetuning] and RLHF stage, which are both relatively thin/short stages (minor finetunes computationally), RLVR involves training against objective (non-gameable) reward functions which allows for a lot longer optimization. Running RLVR turned out to offer high capability/$, which gobbled up the compute that was originally intended for pretraining. Therefore, most of the capability progress of 2025 was defined by the LLM labs chewing through the overhang of this new stage and overall we saw ~similar sized LLMs but a lot longer RL runs. Also unique to this new stage, we got a whole new knob (and and associated scaling law) to control capability as a function of test time compute by generating longer reasoning traces and increasing "thinking time". OpenAI o1 (late 2024) was the very first demonstration of an RLVR model, but the o3 release (early 2025) was the obvious point of inflection where you could intuitively feel the difference.
We're not "evolving/growing animals", we are "summoning ghosts".
Supervision bits-wise, human neural nets are optimized for survival of a tribe in the jungle but LLM neural nets are optimized for imitating humanity's text, collecting rewards in math puzzles, and getting that upvote from a human on the LM Arena. As verifiable domains allow for RLVR, LLMs "spike" in capability in the vicinity of these domains and overall display amusingly jagged performance characteristics
Related to all this is my general apathy and loss of trust in benchmarks in 2025. The core issue is that benchmarks are almost by construction verifiable environments and are therefore immediately susceptible to RLVR and weaker forms of it via synthetic data generation. In the typical benchmaxxing process, teams in LLM labs inevitably construct environments adjacent to little pockets of the embedding space occupied by benchmarks and grow jaggies to cover them. Training on the test set is a new art form.
LLM apps like Cursor bundle and orchestrate LLM calls for specific verticals:
- They do the "context engineering"
- They orchestrate multiple LLM calls under the hood strung into increasingly more complex DAGs, carefully balancing performance and cost tradeoffs.
- They provide an application-specific GUI for the human in the loop
- They offer an "autonomy slider"
The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin. The ultimate reason for this is Moore's law, or rather its generalization of continued exponentially falling cost per unit of computation. Most AI research has been conducted as if the computation available to the agent were constant (in which case leveraging human knowledge would be one of the only ways to improve performance) but, over a slightly longer time than a typical research project, massively more computation inevitably becomes available. Seeking an improvement that makes a difference in the shorter term, researchers seek to leverage their human knowledge of the domain, but the only thing that matters in the long run is the leveraging of computation. These two need not run counter to each other, but in practice they tend to. Time spent on one is time not spent on the other. There are psychological commitments to investment in one approach or the other. And the human-knowledge approach tends to complicate methods in ways that make them less suited to taking advantage of general methods leveraging computation. There were many examples of AI researchers' belated learning of this bitter lesson, and it is instructive to review some of the most prominent.
We have to learn the bitter lesson that building in how we think we think does not work in the long run. The bitter lesson is based on the historical observations that 1) AI researchers have often tried to build knowledge into their agents, 2) this always helps in the short term, and is personally satisfying to the researcher, but 3) in the long run it plateaus and even inhibits further progress, and 4) breakthrough progress eventually arrives by an opposing approach based on scaling computation by search and learning. The eventual success is tinged with bitterness, and often incompletely digested, because it is success over a favored, human-centric approach.
One thing that should be learned from the bitter lesson is the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great. The two methods that seem to scale arbitrarily in this way are search and learning.
The second general point to be learned from the bitter lesson is that the actual contents of minds are tremendously, irredeemably complex; we should stop trying to find simple ways to think about the contents of minds, such as simple ways to think about space, objects, multiple agents, or symmetries. All these are part of the arbitrary, intrinsically-complex, outside world. They are not what should be built in, as their complexity is endless; instead we should build in only the meta-methods that can find and capture this arbitrary complexity. Essential to these methods is that they can find good approximations, but the search for them should be by our methods, not by us. We want AI agents that can discover like we can, not which contain what we have discovered. Building in our discoveries only makes it harder to see how the discovering process can be done.
In this new programming paradigm then, the new most predictive feature to look at is verifiability. If a task/job is verifiable, then it is optimizable directly or via reinforcement learning, and a neural net can be trained to work extremely well. It's about to what extent an AI can "practice" something. The environment has to be:
- resettable (you can start a new attempt),
- efficient (a lot attempts can be made) and
- rewardable (there is some automated process to reward any specific attempt that was made).
The more a task/job is verifiable, the more amenable it is to automation in the new programming paradigm. If it is not verifiable, it has to fall out from neural net magic of generalization fingers crossed, or via weaker means like imitation.
“[The space of minds}(https://karpathy.bearblog.dev/the-space-of-minds/)”:
The computational substrate is different (transformers vs. brain tissue and nuclei), the learning algorithms are different (SGD vs. ???), the present-day implementation is very different (continuously learning embodied self vs. an LLM with a knowledge cutoff that boots up from fixed weights, processes tokens and then dies). But most importantly (because it dictates asymptotics), the optimization pressure / objective is different. LLMs are shaped a lot less by biological evolution and a lot more by commercial evolution. It's a lot less survival of tribe in the jungle and a lot more solve the problem / get the upvote. LLMs are humanity's "first contact" with non-animal intelligence. Except it's muddled and confusing because they are still rooted within it by reflexively digesting human artifacts, which is why I attempted to give it a different name earlier (ghosts/spirits or whatever).