Positions · Aug 28, 2026 · 6 min readThe Goalposts Moved. The Work Didn't.

In 2019, a system that could write a coherent paragraph was a landmark. In 2021, passing a professional exam. In 2023, writing production code.

Each of those was, at the time, the thing that would count. Each one got cleared, and each one stopped counting the moment it did.

Leopold Aschenbrenner caught this two years ago, in the middle of an essay series arguing that AGI by 2027 needed no exotic assumptions, just trend extrapolation. Almost as an aside, he wrote that people are now "in the game of downward-defining AGI", moving the line to whatever current systems can't quite do yet.

Two years on, the aside has held up better than most of the forecast.

What the argument was

Situational Awareness broke AI progress into three drivers, each running at roughly half an order of magnitude a year: raw compute, algorithmic efficiency, and a third thing Aschenbrenner thought everyone was underrating.

He called it unhobbling. Fixing the obvious, structural ways models get stopped from using capability they already have.

The examples look mundane now. A base model has enormous latent capability and is nearly unusable, RLHF turned it into a product. A model forced to answer instantly does badly on hard problems, chain-of-thought let it work step by step. A model that can't call a tool can't check anything. A model with a 4,000-token context can't hold a codebase in its head.

None of those are capability gains in the sense of a bigger model. They're removals of constraint. And by Aschenbrenner's accounting they contributed on roughly the same scale as the entire compute buildout. Epoch AI's survey of scaffolding and tool-use techniques found effective compute gains of 5-30x on many benchmarks, without retraining anything.

His bet was that the category had a lot of room left, and that picking it clean would take systems from chatbots to something closer to coworkers.

Where it landed

The unhobbling thesis has held up well enough that it's gone invisible. Reasoning models, tool use, long context, computer use, agentic scaffolding, persistent memory. Every one of those is an unhobbling, and every one has since shipped.

The clearest way to see the aggregate is METR's time-horizon work, which measures the length of task, in human-expert time, that a model finishes with 50% reliability. Their long-run trend has the horizon doubling roughly every seven months since 2019. Restrict it to models released since 2023 on their newer task suite and the doubling time tightens to about four months.

A GPT-5 agent sits around two hours and seventeen minutes on that measure. METR now notes on its own dashboard that measurements above sixteen hours aren't reliable with the current task suite. The benchmark is running out of room, which is its own kind of signal.

Underneath the horizon numbers is the thing Aschenbrenner flagged as the biggest untapped overhang: test-time compute. His framing was that each token is a word of internal monologue, so a model limited to hundreds of them is a smart person given a few minutes. Thousands of tokens is half an hour of thought. Hundreds of thousands is a workweek. Millions is months.

The tokens got smarter and there got to be far more of them. That's most of what happened.

And the loop has started closing on itself. In June, Anthropic published internal figures stating that more than 80% of the code merged into its production codebase in May 2026 was authored by Claude, up from low single digits sixteen months earlier, with the typical engineer merging roughly eight times as much code per day as in 2024.

The part we should be honest about

It would be easy to stop there. The honest version needs the caveats, and the caveats are significant.

Anthropic hedged its own headline. The 8x figure counts quantity rather than quality and is almost certainly an overstatement of the real productivity gain. A separate internal poll put the median researcher's self-estimated uplift closer to 4x, and the company noted even that is probably inflated by known reporting biases.

METR's caveats are sharper. A 50% time horizon doesn't mean tasks of that length can be handed off, plenty of work needs reliability far above a coin flip before automation is worth anything. More to the point, METR is explicit that its tasks are self-contained and well-specified with clean success criteria, while real work runs on prior conversations, tacit knowledge, and familiarity with a system you've used for years. Their own guidance is to read a two-hour task as what someone with no prior context could do in two hours. A new hire or a freelance contractor, not an experienced person who already knows the account.

They also found agent performance drops substantially when output gets judged holistically instead of scored algorithmically.

So the trendlines are real, and the gap between benchmark capability and deployed value is also real. Both are true, and the second one is where almost everyone actually works.

Meanwhile, in the workplace

Here's the number that should be jarring next to all of that.

As of May 2026, 15% of US employees use AI daily in their role. Thirty percent use it a few times a week or more. Just over half, 52%, use it a few times a year or more.

Daily use is concentrated almost entirely in three sectors: technology at 42%, finance at 27%, professional services at 22%. Across the other seven industries Gallup tracks, daily use runs between 9% and 15%.

Two more findings from the same data matter more than the headline.

Only 25% of employees say their organization has communicated a clear plan for integrating AI. That's been flat for a year. Three in four people work somewhere with no articulated strategy at all.

Only 14% strongly agree AI has changed how work gets done, and that's measured among employees at organizations that have already implemented it. Sixty-five percent of those same employees say it improved their productivity.

Gallup's own read on that split is the right one. For most workers, AI is a productivity enhancer inside existing workflows, not a reason to rebuild them.

Two-thirds report going faster. One in seven report anything actually changing.

The schlep was always the bottleneck

Buried in Situational Awareness, in a passage that got far less attention than the 2027 date, Aschenbrenner called this. Intermediate models, he wrote, would take enormous amounts of schlep, meaning workflow changes and infrastructure, before anyone could pull economic value out of them. He expected the schlep to take longer than the unhobbling.

It has.

And it isn't a model problem, which is why more capability hasn't fixed it. Look at what separates a METR benchmark task from a real one. The benchmark task is self-contained, well-specified, cleanly scorable. A real task runs on ten years of accumulated context, on conventions nobody wrote down, on knowing which version of the document is the executed one, and on an answer someone can verify before they put their name on it.

The capability to do the work exists. The context needed to do it correctly is stuck inside systems built to store records, not to reason over them, and the frontier isn't moving toward those systems on its own.

That gap doesn't close with a better model. It closes by building the thing that holds the context.

Where we sit

Niska builds operational software with document intelligence and agent orchestration at the core, configured to each organization's own data and conventions.

That last clause is the whole product. Not a general model pointed at a company's files, and not an AI feature bolted onto software designed in 2009 to file things. A rebuilt operational layer, document management, workflow, reporting, where the intelligence is native instead of appended, and where the system learns a specific firm's conventions instead of inventing its own.

What stays familiar is the surface. People keep working the way they already know how to work: folders, queues, reviews, approvals, an audit trail. Nobody has to become a prompt engineer to get value out of it, and no partner has to trust an answer they can't click through to a source document to check.

What changes is everything underneath.

That's the bridge. On one side is software people already know how to use. On the other is capability that, by every available measure, is running well ahead of most organizations' ability to reach it.

The goalposts will keep moving. That was always going to happen. The more interesting question is why, with frontier systems now writing most of the code at the labs that build them, one in seven workers says anything about their job has actually changed.

Niska builds AI-native operational software for construction, regulated enterprise, and government contracting.