What are AI's open problems today?

As of July 2026, AI’s open problems cluster into five areas: reliability, autonomy over hours, learning after training, cost, and understanding what happens inside. None of them is new, and none of them is the same as three years ago. The list shrinks more slowly than the announcements suggest and faster than the scepticism predicts.

This article carries a date and ages by design. It gets rewritten rather than patched, and it exists to give the map: what is stuck, what is moving, and how to tell them apart. Its most useful part is the third figure, listing problems that were here and left — what changed in the last twelve months is the fine-grained version of that.

Two symmetrical traps surround any list like this. One is reading a current limitation as a permanent property of the technology. The other is assuming everything falls by extrapolation. Telling real progress from hype is the exercise that avoids both, and hallucination is the best case study, because both kinds of argument have real evidence behind them.

Where the problems sit

Severity and velocity are independent axes, and conflating them is the most common mistake in public commentary. Cost per token is a real problem that is falling on its own. Interpretability barely affects the daily work of anyone shipping a product and is close to stationary relative to the size of the target. Treating the two with the same urgency makes no sense.

Reliability: the model does not know that it does not know

Hallucination is the most discussed problem and the only one with two strong results pointing in opposite directions.

The pessimistic argument is formal. A 2024 paper defined a formal world where hallucination is the inconsistency between a computable LLM and a computable ground-truth function, and used learning theory to show an LLM cannot learn all computable functions. It follows that a model used as a general-purpose solver will inevitably hallucinate¹. That is a result about the limit, not about the rate.

The optimistic argument is empirical and more actionable. A 2025 paper argues models hallucinate because both training and evaluation reward guessing over acknowledging uncertainty: a model optimised to do well on exams learns that guessing pays better than leaving a question blank. Their proposal is socio-technical, changing how the dominant benchmarks score abstention rather than adding yet another hallucination-specific evaluation².

Both can be true at once, and probably are. The rate has plenty of room to fall; zero is not on the menu. In practice that means systems depending on factual correctness need external verification by design, not a better model by waiting.

Autonomy: tasks that run for hours

This is the problem that has moved the most and remains the bottleneck for agentic products.

The most useful measurement appeared in 2025. Instead of asking what percentage of a benchmark the model gets right, the group measured how long a human expert takes on the tasks the model completes with 50% success. Frontier models of that period landed around 50 minutes, and the horizon had been doubling roughly every seven months since 2019, with signs of acceleration in 2024. The authors attribute the gain mainly to greater reliability and better error recovery, not to more knowledge³.

The failure mode underneath is arithmetic. A twenty-step task at 95% per-step accuracy finishes intact 36% of the time. Improving from 95% to 99% per step takes the end-to-end rate from 36% to 82%, which is why progress on reliability pays more than progress on raw capability in that range.

A 2024 benchmark made this visible another way. It simulates conversations between a user and an agent equipped with tools and domain rules, then compares the final database state against the expected one. State-of-the-art function calling agents solved under 50% of the tasks, and the consistency metric across repeated attempts fell below 25% over eight runs in the retail domain⁴. Doing it once and doing it every time are different capabilities, and production wants the second.

There is also a ceiling that arrives earlier than expected. A 2025 study used puzzles of controllable complexity and found complete accuracy collapse past a certain point, alongside a counterintuitive behaviour: reasoning effort grows with complexity up to a limit and then declines, even with token budget left over⁵. The work drew legitimate methodological debate about how much of the collapse comes from the task format, and it is worth reading with that caveat. Even so, the pattern of a model giving up before spending its budget reproduces and is uncomfortable.

Learning: knowledge freezes when training ends

A deployed model is a frozen file. Nothing you write changes the weights, and yesterday’s conversation leaves no trace. What looks like memory is the surrounding system resending text on every call.

That has two consequences with no clean fix as of July 2026. The first is the knowledge cut-off: what was not in the corpus is not there, and updating requires continued pretraining, with proportional cost and a real risk of damaging what already worked. The second is forgetting. A 2023 study measured continual instruction tuning in models from 1 to 7 billion parameters and found broad degradation in domain knowledge, reasoning and reading comprehension, with severity rising with size inside that range⁶.

The workaround the industry settled on is not to solve it: new knowledge goes into the context, through retrieval, and the model stays ignorant between training runs. It works, and it is not the same thing as learning.

Cost: the bill moved somewhere else

Cost per token had been falling steadily, and most 2023 analyses assumed it would stop being a problem. A 2024 result scrambled that forecast. It showed that spending compute at answer time can pay more than growing the model: with adaptive per-prompt allocation, test-time scaling efficiency improved by more than 4 times over a simple baseline, and on problems where a small model already had some success rate, spending compute at inference beat a model 14 times larger at the same total operations⁷.

The effect is that compute migrated from training to inference. A reasoning model can spend thousands of tokens thinking before writing the first word of an answer, and much of that is billed without ever being shown. As of July 2026 the operational question stopped being “what does a token cost” and became “how many invisible tokens does this task consume”.

Two practical consequences follow from that migration. The first is that latency became part of the cost in a new way: an answer that needs thirty seconds of reasoning does not fit a synchronous interface, and the product has to be redesigned around it, not just the invoice. The second is that routing became the highest-return optimisation available. Sending every question to the most expensive model is the default configuration of nearly every hastily built system, and classifying difficulty before choosing the model routinely halves the bill with no change in perceived quality. It is the same adaptive allocation from the 2024 work, applied one level up.

Understanding: nobody can explain an answer

This is the oldest problem on the list and the one that least affects daily work, which explains why it gets less attention than it deserves.

The core difficulty has a name: a neuron in a language model tends to activate across semantically distinct contexts, because the network represents more features than it has neurons. A 2023 paper attacked that with sparse autoencoders that reconstruct internal activations, producing feature sets more interpretable than the directions found by alternative approaches, and managed to pinpoint which features are causally responsible for a specific behaviour at finer granularity than earlier decompositions⁸.

That is real progress and it is still a research foundation. Between it and being able to say, with any guarantee, why a frontier model produced a particular sentence in a particular context, the distance remains large.

Generalisation: the frontier that does not move with the rest

The problem that resists scaling best deserves its own place. In 2025, the second version of a benchmark designed to measure fluid intelligence kept its predecessor’s input-output pair format, with newly curated tasks built to demand abstraction rather than accumulated knowledge, and reported human testing that establishes a baseline accessible to ordinary people and hard for current systems⁹.

The pattern this exposes is the most interesting on the whole list: capabilities that require recombining a few new rules in an unfamiliar setting improve far more slowly than capabilities that require knowledge and execution. It is the difference between a content exam and a riddle.

The map of the subject

Knowing which area a complaint falls into already eliminates half the wrong answers to it. A reliability problem is attacked with external verification. An autonomy problem is attacked with smaller steps and checkpoints. A learning problem is attacked with retrieval. A cost problem is attacked with routing. An understanding problem is not attacked: you either accept it or avoid the use case.

What has fallen in recent years

This figure is what keeps the rest of the text honest, and every line in it has a checkable origin.

GPT-3 worked with a 2,048-token window in 2020¹⁰; as of July 2026 advertised production windows sit on the order of a million. Reasoning on verifiable tasks depended on human-written demonstrations until, in 2025, a group showed pure reinforcement learning incentivises the same capabilities with no labelled trajectories, with self-verification and strategy changes emerging during training¹¹. MMLU, published in 2020 as a test where the largest GPT-3 beat chance by about 20 points, saturated above 90% and stopped informing. And the task horizon has been doubling roughly every seven months since 2019³.

The caveat that keeps the figure honest: none of these fell by itself. Each one fell because somebody worked on it for years, and some fell only halfway. A million-token window does not mean a million usable tokens — a 2024 benchmark evaluated 17 long-context models and found large drops as length grows, with only half of them holding satisfactory performance at 32 thousand tokens despite all of them advertising 32 thousand or more¹².

Where this list fails

It is written by the side that publishes. Nearly all the evidence here comes from labs and universities with a stake in the subject. That is not a reason to discard it, it is a reason to prefer results that measure a drop over results that promise one.

It confuses what is hard with what is unpopular. Some problems stall because they are hard; others because nobody is paying. Continual learning has more of the second than the discourse suggests.

It ages unevenly. Cost moves in months, interpretability in years. Rereading this page twelve months from now will find two sections obsolete and three unchanged, and there is no telling in advance which.

It measures what we know how to measure. The entire autonomy discussion rests on software-task benchmarks, because that is where automatic verifiers exist. That leaves out most human work.

What to do with this

1. When you hear a limitation, ask how old the data is. Half the claims about what AI cannot do describe the state of two years ago.
2. Treat reliability as a systems problem, not a model problem. External verification, small steps and checkpoints pay more than waiting for the next version.
3. Measure your per-step rate before designing a long agent. The composition arithmetic decides the maximum viable length, and it is unforgiving.
4. Do not build on the promise that an item on this list will fall. Some do; which one and when is exactly what nobody gets right.
5. Keep your own list, dated. Write down what blocked your project this quarter and reread it in six months. It is the cheapest instrument there is for calibrating expectations.

Footnotes

1. Xu et al. (2024) formalised hallucination as inconsistency between a computable LLM and a computable ground-truth function and showed, through learning theory, that it is inevitable for a general-purpose solver. ↩

2. Kalai et al. (2025) argued that models hallucinate because training and evaluation reward guessing over acknowledging uncertainty, and proposed changing how dominant benchmarks score abstention rather than adding more hallucination evaluations. ↩

3. Kwa et al. (2025) defined a 50%-task-completion time horizon in terms of how long human experts take, measured around 50 minutes for the frontier models of the time, and found a doubling roughly every seven months since 2019. ↩ ↩²

4. Yao et al. (2024) simulated user-agent conversations with tools and domain rules, comparing the final database state against the expected one: state-of-the-art agents solved under 50% of tasks, with consistency below 25% over eight attempts in retail. ↩

5. Shojaee et al. (2025) used puzzles of controllable complexity and found accuracy collapse past a certain point, with reasoning effort declining despite remaining token budget. ↩

6. Luo et al. (2023) observed catastrophic forgetting during continual instruction tuning in models from 1 to 7 billion parameters, with severity rising with size in that range. ↩

7. Snell et al. (2024) measured test-time compute scaling and found more than a 4-times efficiency gain with adaptive per-prompt allocation, beating a model 14 times larger at the same operation budget on problems with partial success. ↩

8. Cunningham et al. (2023) used sparse autoencoders to reconstruct internal activations and obtained features more interpretable than the directions from alternative methods, pinpointing those responsible for a specific behaviour at finer granularity. ↩

9. Chollet et al. (2025) published the second version of a fluid intelligence benchmark with newly curated tasks and human testing establishing a baseline accessible to people and difficult for current systems. ↩

10. Brown et al. (2020) trained GPT-3 with a 2,048-token context window. ↩

11. DeepSeek (2025) showed reasoning ability can be incentivised through pure reinforcement learning, with no human-labelled trajectories. ↩

12. Hsieh et al. (2024) evaluated 17 long-context models on tasks beyond simple retrieval and found large drops as length increases: half of them failed to hold satisfactory performance at 32 thousand tokens despite advertising 32 thousand or more. ↩