Four observations about AI in 2026: a June check-in
Hands-on AI observations from personal and enterprise scale work.
Time flies, and we’re already completing the first half of the year. Here’s my big four takeaways from personal practice and enterprise scale work.
The most efficient and effective use of AI is use-case driven #
Right now, too many large enterprises are on an immature phase of ‘AI everything’. Big AI contracts, big AI platforms, big AI visions.
It’a a top-down approach with limited success. Wanting to ‘jump in the AI wagon’ is an understandable ambition and honestly, a bit too much financially rewarded. But you can’t make that ambition your full technology strategy, because the result is often:
- old problems being half-solved
- new problems created which you couldn’t predict, often complex and unsolvable
- blind spots that are hard to catch
- too much waste all around (budget, time, employee and client patience, brand reputation…)
A more mature strategy is bottom to top: focus ruthlessly on the problem, then get a targeted and efficient solution to it. Whether if it’s AI-enabled or not is irrelevant.
The most useful AI tool in enterprise is the boring tool that accelerated work for specific tasks or it removed frictions (time wasted, or kafkaeske processes everybody hated).
As obvious as this sounds, it has to be said aloud so we don’t lose ourselves too much on promises that might never be delivered.
Benchmarks measure capability. Users perceive calibration. #
The LLM community is obsessed with numbers. Parameter count, context window, tokens per second. SWE-bench, MMLU, GPQA. Leaderboards.
The numbers track something real, but they’re tracking 2024’s bottleneck. In 2024, capability was the limit. Models couldn’t reason, couldn’t follow instructions, hallucinated constantly. More parameters fixed real problems.
In 2026 the bottleneck shifted. Frontier models are extraordinary at reasoning. They hold a million tokens of context. They use tools. The remaining bottleneck is calibration: weighting the current input correctly given everything else. Calibration does not scale with parameter count. Sometimes it inversely correlates.
I’ve watched a 27B model running locally read a casual conversation better than a flagship with full memory and thinking enabled. Same prompt. The flagship had so much context it matched on the wrong feature and fired a safety lecture. The small model just answered the message.
For most real-world tasks, a well-configured small model with a thoughtful prompt beats a flagship with bad setup. The benchmark for what makes AI useful is the felt sense of “this answer matched what I needed.” That benchmark belongs to the user, not the lab.
The full argument, with the Sunday afternoon test that triggered it: From measuring performance to measuring user experience quality.
When AI looks broken, the model is rarely the bug #
The LLM inference chain is more fragile and complex than the marketing suggests. Pipeline, prompt, retrieval, model, sampling, post-processing. Failures look like model failures. They’re usually further upstream.
Earlier this year I caught a thing. A model output that looked miraculous in a long-form text evaluation. Tried to reproduce it. Failed. Designed a controlled experiment that held retrieval constant and varied only the model. Turns out the “extraordinary output” we’d been chasing was temperature stochasticity, not quant or model superiority. RAG was poisoning the context, biasing toward a specific direction, making a retrieval artefact look like a model breakthrough.
When a model fails, it’s tempting to blame size, quantization, or sampling. The actual cause is often one or two layers up. Cached context. Misconfigured retrieval. Prompt template that doesn’t match what the model was trained on. Diagnosis matters more than blame.
Voice will reshape how we use AI. Flagships are not ready. #
Typing is friction. On mobile it’s painful. On desktop it’s a bottleneck for anyone doing parallel tasks. Voice is the obvious next interface, especially for the use cases AI is rapidly being adopted for: coaching, planning, exploring ideas, conversational thinking.
Flagship voice modes are not delivering this yet. Claude’s voice mode interrupts itself in ways that make it unusable on a wide range of devices. Reddit threads full of users begging for fixes. GPT’s voice mode is currently SOTA among flagships, and even there the gap between “good demo” and “reliable daily driver” is wide.
A self-hosted setup can outperform flagship voice on the axes that matter day to day. My own implementation doesn’t interrupt itself. It also doesn’t moralize, doesn’t apologize unnecessarily. None of those are technical breakthroughs. They’re configuration. Which is the point: if a self-hosted setup can match or beat flagship voice on the things users actually feel, the flagships have a calibration problem, not a capability problem.
The interface shift is coming faster than the leaderboards reflect.