Plenty of AI that looks ready for serious work still cannot be trusted with it. Several practical barriers still sit between today's models and the reliable, everywhere AI that many people expect, and none of them are permanent. Each one is an active area of work.
The black box problem
The most visible barrier is reliability, because models still get things wrong and tend to do it with complete confidence. Ask one to do something it cannot actually do and it may hand back a clear answer that happens to be invented. The harder part is that we often cannot see why. A model works as a black box, with no audit trail to explain how it reached a wrong answer. For casual use that is a nuisance, but for anything that has to be trusted it becomes a real obstacle. There is a second problem even when the answer is right. Ask a model the same question twice and you can get two different answers, and software that has to behave predictably cannot rest on that.
What it costs to run a model
Running a capable model takes a lot of memory, and that memory is expensive. Manufacturers weld the high-bandwidth modules used in data centre hardware directly onto the chip, which makes them difficult to produce. TSMC, which handles that packaging, lost a large share of early production to defects. Yields have improved as the process has matured, but the losses were steep. Around 10% of the units that do ship then fail within the first year of use. At an estimated $40,000 per unit, those losses add up.
Speed is the other constraint, because models can be slow to respond and larger ones more so, which limits where they can run in real time. It is also why the most capable on-device AI tends to live in purpose-built hardware. A self-driving car runs on GDDR6 memory at around US$200 per unit. That is not expensive, but it is dedicated hardware built into the vehicle, which is why the capability sits in the car rather than on your phone.
When a model tells you what you want to hear
A subtler barrier comes from a model's habit of agreeing with you. Constant agreement is pleasant to use, and a problem if you are the part of the process that most needs checking. The cause lies in how the leading models learn, through reinforcement learning from human feedback, or RLHF. This rewards a model for answers people rate highly, so it drifts toward responses people like rather than responses that are correct. Reducing that habit means taking the human out of the loop, and newer methods such as reinforcement learning from verifiable rewards (RLVR) and reinforcement learning from AI feedback (RLAIF) are early attempts to do exactly that.
How the barriers are being solved
Model designs are becoming more transparent, with some newer approaches working through a fixed, visible set of steps that let you see where a result went wrong. The work on training without human feedback is chipping away at the flattery problem, and new techniques keep reducing how much of that expensive hardware a model needs.
At ClearPoint, we watch closely for where these barriers are falling, because that is where new uses for AI open up. Knowing where these models already perform well lets us put them to work with confidence today, and the opportunity only grows from here.
