AI 2027: What If We Can't Keep Up?
New forecasts put superhuman AI researchers in 2027. What happens to work and human control if AI helps build its successors faster than we can adapt?
The models, the people, and the power behind AI.
New forecasts put superhuman AI researchers in 2027. What happens to work and human control if AI helps build its successors faster than we can adapt?
GDPx combines different definitions of success into one score. Reading the number means separating task progress, finished work and economic impact.
Matched coding tasks exposed a performance gap that scans for explicit refusals largely missed. The result is not a general failure rate.
Ai2’s StateBench separates looking up information from following dependent updates. Its synthetic tests reveal a gap that a larger context window does not measure.
A research team raised a small model’s math scores by repairing repetitive reasoning. A second experiment shows why that is not the same as teaching new skills.
Grok, ChatGPT, Claude and their rivals can now do more than answer questions. The useful differences are what keeps running, what carries over and what still needs you.
An Ai2 study found large swings in originality judgments when the instructions changed. The result tests AI evaluation, not whether machines can invent.
Binance says internal AI tooling was compromised in its $6.3 billion theft. That is not evidence of an autonomous attacker financing its own operation.
Epoch’s offline experiment shows why a visible repair can become an attack clue before an update ships. A second, now-patched flaw was also essential.
A controlled experiment shows how an AI can shape an expert’s analysis before the final answer is checked. Its 86% result is not a real-world rate.
A first attempt, a stronger replacement and an independent review do different jobs. Fusion chooses among them, but rejecting a draft does not reverse its file edits.
The new ranking asks harder questions about finished work. Its wider gap needs two qualifications: a changed test and model weights that are not yet available.