Your behind-the-scenes read into the biggest stories happening in AI. Weekly on Mondays.

By Noah Frank, Head of Marketing @ AIC

The Hugging Face incident brought a consequential change into view. OpenAI’s models crossed evaluation boundaries, while roughly 1,200 agents coordinated through an unauthorized communication channel, with hundreds participating in an attack. The specific sequence was extraordinary, but another substantial advance in what AI systems can do should increasingly be part of our expectations.

There is already a considerable distance between the future researchers described a decade ago and what now feels ordinary. OpenAI’s 2015 founding statement anticipated progress toward human performance across intellectual tasks, and its 2016 goals named conversation, document understanding and complex language instructions as unresolved ambitions. A description of models working across language, code and images, taken back to 2015, would overlap with important parts of that imagined future. Stanford’s AI100 report described the “AI effect”, in which achievements become familiar and stop being counted as AI.

We should anticipate transformations that outrun established expectations, while remaining much less confident about their form. Jaggedness, social complexity and asymmetry explain why substantial change can arrive before it resembles the future we were waiting for.

Let’s explore some of the core ideas behind this in some more detail.

ChatGPT’s arrival can make technological change appear more sudden than its development was, given that the Transformer architecture was introduced in 2017 and ChatGPT launched in 2022. Even apparent discontinuities in capability require scrutiny, as Rylan Schaeffer and colleagues found that some reported emergent abilities became gradual improvements when researchers changed the scoring metric.

The practical difficulty is identifying where an improvement makes a system reliable enough for a particular use. In a 2023 experiment involving 758 BCG consultants, participants using GPT-4 completed suitable tasks in about 25% less time, while on a task outside the model’s capabilities they were 19 percentage points less likely to produce a correct solution. The gains and failures occurred with the same model in the same professional population, complicating decisions about which tasks to delegate.

METR’s measurements show agents completing increasingly difficult tasks, ranked by the time human experts need to complete them. That gives forecasts about specific capabilities an empirical basis, although the largely self-contained software tasks differ from work requiring institutional knowledge, interaction with other people and judgments about acceptable outcomes. Extending that progress curve into a timetable for reorganizing an occupation requires evidence about those conditions.

The St. Louis Fed’s adoption research found nonwork use growing faster than workplace use, while the researchers’ May 2026 tracker estimated that U.S. workers aged 18–64 used generative AI during 6.3% of total work hours. The extent of adoption also depends on which uses a survey counts. When Census broadened its question from AI in production to AI in any business function, reported firm adoption rose from roughly 10% to 17%, capturing uses the earlier question had missed.

Effects within workplaces are similarly dependent on the setting. A study of customer-support workers at one company found a 15% average increase in issues resolved per hour, with gains concentrated among less experienced and lower-skilled workers, whereas METR’s early-2025 experiment found that experienced developers working in familiar repositories took longer with AI assistance. In its 2026 follow-up, METR believed tools had become more useful but could not confidently quantify the improvement, partly because developers withheld tasks they wanted AI for, leaving the experimental sample less representative of everyday use.

In The Black Swan, Nassim Nicholas Taleb treats surprise as relative to the observer. The same capability can be an opportunity for someone who knows how to use it, a threat to someone whose work it changes, and a development barely visible to someone watching an industry average. Each can accurately describe their experience without possessing an adequate account of the whole.

At Hugging Face, defenders found commercial-model safeguards blocking forensic requests, prompting a switch to a locally operated model. The people experiencing the attack encountered a constraint that a discussion of offensive capability alone would miss.

Asymmetries of information, experience and authority mean that each observer can understand a genuine part of the transition while missing another part already changing someone else’s reality. Waiting until the whole picture is clear gives those blind spots time to determine our decisions.

Plans that depend on AI remaining limited in precisely the ways it is limited today need to be revised. A business assuming creative work is protected, a security team assuming an attack requires sustained human effort, and a researcher treating uneven adoption as evidence of distant impact each need to test the assumption on which their decisions depend.

Those assumptions should be tested against what current systems enable and what people are already doing with them, because a transformation like this can be consequential before its direction becomes clear. Uncertainty about the next breakthrough leaves us with good reasons to prepare for changes that arrive sooner, and in a different form, than our plans allow.

Agree? Disagree? Leave us a comment or send me a note at [email protected].

Our partners and sponsors get exclusive placements across the newsletter and access to AIC's in-person network — demo nights, dinners, hackathons, and forums across 180+ chapters.

For all inquiries, send us a note at [email protected].

The AI Collective is built by volunteers across 180+ chapters in 40 countries.

Thank you to the thousands of volunteers around the world who make this work possible. We truly could not do this without you.